From 47769981c6fb596f892f0b2ab64c24c8e22b38e9 Mon Sep 17 00:00:00 2001
From: Valery Kharseko <vharseko@3a-systems.ru>
Date: Wed, 23 Sep 2026 07:38:47 +0000
Subject: [PATCH] [#915] Bound the wait of a write transaction for a row lock another session holds (#1010)

---
 opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCRowLockBoundTestCase.java |  803 +++++++++++++++++++
 opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/TestCase.java                 |  101 ++
 opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCDdlLockBoundTestCase.java |   66 +
 opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/CachedConnection.java         |   30 
 opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCStorageRetryTest.java     |  374 ++++++++
 opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/JDBCStorage.java              | 1114 ++++++++++++++++++++++----
 6 files changed, 2,281 insertions(+), 207 deletions(-)

diff --git a/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/CachedConnection.java b/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/CachedConnection.java
index 516960e..d629d0f 100644
--- a/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/CachedConnection.java
+++ b/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/CachedConnection.java
@@ -1325,6 +1325,14 @@
     private volatile boolean poolable;
 
     /**
+     * The session value the row lock bound of this backend displaces on this connection, read once and
+     * kept for its life: see {@link #rowLockBoundDisplaces()}. Volatile for the reason
+     * {@link #poolable} is - it is written by one borrower and read by the next, which is another
+     * thread, and the value is a plain field of an object the pool hands between them.
+     */
+    private volatile Long rowLockBoundDisplaces;
+
+    /**
      * When this connection last answered the database, as a {@link System#nanoTime()} reading:
      * established - the login and the two round trips that set it up have just answered - or
      * validated. It is never stamped on the way back into the pool, although that is where a
@@ -1369,7 +1377,7 @@
     /**
      * Keeps this connection out of the pool: it serves the borrower holding it and is closed rather
      * than pooled when that borrow ends. For a borrower that left something of its own on the session
-     * and could not take it off again - {@code JDBCStorage.restoreDdlLockBound()} is the one that
+     * and could not take it off again - {@code JDBCStorage.restoreLockBound()} is the one that
      * does (#885) - where the blast radius is then this one connection instead of every borrow it
      * would have served after this one.
      */
@@ -1377,6 +1385,26 @@
         poolable = false;
     }
 
+    /**
+     * What this session carried before the row lock bound of {@code JDBCStorage.write()} first
+     * displaced it, or null while that has not been read yet (#915).
+     * <p>
+     * Remembered for the life of the connection because that bound is on the hot path - one write of
+     * the server, one arming - while the readback it saves is a round trip. Only this backend writes
+     * that setting on a connection of this pool, every write puts the value back before the
+     * connection is released, and a connection whose restore failed is kept out of the pool by
+     * {@link #keepOutOfThePool} rather than handed on: so a value remembered here cannot outlive the
+     * session that answered it. Not reset on borrow for the same reason - it describes the session,
+     * which outlives every borrow of it.
+     */
+    Long rowLockBoundDisplaces() {
+        return rowLockBoundDisplaces;
+    }
+
+    void rowLockBoundDisplaces(Long previous) {
+        rowLockBoundDisplaces = previous;
+    }
+
     /** Gives back the right to hold this connection, once and only if it was taken. */
     void releasePermit() {
         if (metered && permitReleased.compareAndSet(false, true)) {
diff --git a/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/JDBCStorage.java b/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/JDBCStorage.java
index 70e45f8..e38d3cf 100644
--- a/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/JDBCStorage.java
+++ b/opendj-server-legacy/src/main/java/org/opends/server/backends/jdbc/JDBCStorage.java
@@ -49,7 +49,9 @@
 import java.util.concurrent.atomic.AtomicBoolean;
 import java.util.concurrent.atomic.AtomicInteger;
 import java.util.concurrent.atomic.AtomicLong;
+import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.Predicate;
+import java.util.function.Supplier;
 
 import static org.opends.server.backends.pluggable.spi.StorageUtils.addErrorMessage;
 import static org.opends.server.util.StaticUtils.stackTraceToSingleLineString;
@@ -63,24 +65,29 @@
 
 	/**
 	 * Wall-clock budget the replays of a {@link #write} may spend, in nanoseconds, measured from the start of the
-	 * first attempt. It is checked between attempts, so an attempt already running is never interrupted, and it
-	 * applies from the first check, with the single exception {@link #grantedPastTheWindow} describes: a conflict
-	 * its engine reports promptly is granted one replay whatever the clock says, because the lock wait that
-	 * precedes such a conflict is charged to the attempt and is unbounded on three of the four engines here, so no
-	 * window survives it. It bounds what {@link #MAX_RETRIES} alone does not - MySQL reports a lock wait timeout
-	 * only after innodb_lock_wait_timeout, 50 s by default and not overridden here, so ten attempts would park a
-	 * worker thread for eight minutes where one releases it after 50 s.
+	 * first attempt. It is checked between attempts, so an attempt already running is never interrupted: the loop
+	 * returns after at most this window plus one attempt. It bounds the conflicts that are slow to report, which
+	 * {@link #MAX_RETRIES} alone does not - MySQL reports a lock wait timeout only after innodb_lock_wait_timeout,
+	 * 50 s by default, so ten attempts would park a worker thread for eight minutes where a single one released it
+	 * after 50 s.
 	 * <p>
-	 * What that costs, stated rather than left to be read off a test row: at the stock innodb_lock_wait_timeout a
-	 * MySQL lock wait timeout is reported at ~50 s, which is past this window on the first check, so such a write
-	 * is never replayed at all - the one conflict class of the set that a MySQL deployment sees most, and the one
-	 * whose replay would most reliably succeed. It is the deliberate half of the trade the other half of which is
-	 * #903: one bounded wait beats two, and a deployment that tunes innodb_lock_wait_timeout below this window
-	 * gets its replays back. The trade only exists because nothing here bounds the attempt: with a session lock
-	 * timeout on the transaction connection (#915) every wait would be shorter than this window, the tuned-down
-	 * case would become the normal one, and this window would govern both classes with no grant needed at all.
+	 * A window is only a bound on the replays while an attempt is shorter than it, and what makes an attempt long
+	 * is the lock wait ahead of the conflict, charged to the attempt that hit it. That wait is bounded inside the
+	 * attempt by {@link #ROW_LOCK_TIMEOUT_PROPERTY} - including the innodb_lock_wait_timeout above, which is taken
+	 * down to it for the length of the write - so a conflicted write is replayed within this window rather than
+	 * spending the whole of it on one attempt and being refused a replay it could have made (#903, #915). The
+	 * replays a write gets are therefore this window divided by that bound: three at both defaults.
+	 * <p>
+	 * Two cases still spend it in one attempt, and they are the two nothing bounds: an oracle session waiting in a
+	 * row-lock enqueue, which has no setting to bound it, and a deployment that sets that property to 0. There the
+	 * loop behaves as it did before #915 - the window applies from the first check, with the single exception
+	 * {@link #grantedPastTheWindow} describes: a conflict its engine reports promptly is granted one replay
+	 * whatever the clock says, since no window survives a wait nothing bounds, and measuring one against it would
+	 * only leave the operation with no replay at all (#903).
 	 */
-	private static final long RETRY_WINDOW_NANOS = TimeUnit.SECONDS.toNanos(10);
+	// package-private so that the suite can assert the one invariant tying it to ROW_LOCK_TIMEOUT_PROPERTY:
+	// a bound at or past this window leaves the replays exactly where #903 found them
+	static final long RETRY_WINDOW_NANOS = TimeUnit.SECONDS.toNanos(10);
 
 	/** Upper bound of the random delay before the second attempt, in milliseconds; it doubles with every attempt. */
 	private static final double BASE_SLEEP_ON_RETRY_MS = 50.0;
@@ -624,23 +631,41 @@
 	private final AtomicBoolean backstopFailedWarned = new AtomicBoolean();
 	private final AtomicBoolean queryTimeoutWarned = new AtomicBoolean();
 	private final AtomicBoolean standingReadBoundWarned = new AtomicBoolean();
-	// The same, for the two ways the lock bound of a DDL degrades: a session that would not take the
-	// setting - or would not say what it carried before it - and one it could not be taken off again.
-	// The second one is a moment rather than a latch, the way CachedConnection throttles the read bound
-	// it could not lift: whatever makes a restore fail - a transaction the server has doomed,
-	// middleware that rejects a SET - recurs on every DDL, and each occurrence now costs the pool the
-	// connection it happened on, so said once for the life of the storage an operator could not tell
-	// one stranded bound from a pool full of them.
-	private final AtomicBoolean ddlLockBoundNotSetWarned = new AtomicBoolean();
-	private final AtomicLong ddlLockBoundLeftBehindWarned = new AtomicLong();
-	private static final long DDL_LOCK_BOUND_WARNING_INTERVAL_MS = 10000;
+	// The same, for the two ways a lock bound of this backend degrades: a session that would not take
+	// the setting - or would not say what it carried before it - and one it could not be taken off
+	// again. The second one is a moment rather than a latch, the way CachedConnection throttles the
+	// read bound it could not lift: whatever makes a restore fail - a transaction the server has
+	// doomed, middleware that rejects a SET - recurs on every statement it is put around, and each
+	// occurrence now costs the pool the connection it happened on, so said once for the life of the
+	// storage an operator could not tell one stranded bound from a pool full of them.
+	// Per LockBound rather than per storage: the two bounds are set by different code on different
+	// paths, and a DDL of the open that could not take its bound says nothing about the writes that
+	// follow it - reported through one latch, the first open would silence every write of that
+	// backend.
+	// Not private, so that a case can read which bound has already said something without reading a log:
+	// what these are is one latch per bound, and a case saying so is what keeps them from becoming one.
+	final EnumMap<LockBound,AtomicBoolean> lockBoundNotSetWarned = perBound(AtomicBoolean::new);
+	final EnumMap<LockBound,AtomicLong> lockBoundLeftBehindWarned = perBound(AtomicLong::new);
+	private static final long LOCK_BOUND_WARNING_INTERVAL_MS = 10000;
 	// And a third way, which is no failure of anything: an engine this backend knows no lock setting
-	// for is left unbounded deliberately, and says so once - see reportTheEngineIsNotKnown(). Not
+	// for is left unbounded deliberately, and says so once - see reportTheEngineIsNotKnown(). Per
+	// LockBound like the two above and for the same reason: the engine is the same for both bounds,
+	// but what an operator is told to act on is the property of the wait that was left unbounded, and
+	// a backend opening on such a driver would otherwise say nothing about the writes behind it. Not
 	// private, so that a case can read what a storage has already said without reading a log.
-	final AtomicBoolean ddlLockBoundEngineUnknownWarned = new AtomicBoolean();
+	final EnumMap<LockBound,AtomicBoolean> lockBoundEngineUnknownWarned = perBound(AtomicBoolean::new);
 	// What was actually said, set beside the flag above and read the same way: the flag alone tells a
 	// case that something was logged, not what it named.
-	volatile String ddlLockBoundEngineUnknownSaid;
+	final EnumMap<LockBound,AtomicReference<String>> lockBoundEngineUnknownSaid = perBound(AtomicReference::new);
+
+	/** One of these per bound, made where this class is: the bounds are an enum, so this is an EnumMap. */
+	private static <T> EnumMap<LockBound,T> perBound(Supplier<T> value) {
+		final EnumMap<LockBound,T> perBound=new EnumMap<>(LockBound.class);
+		for (final LockBound bound : LockBound.values()) {
+			perBound.put(bound, value.get());
+		}
+		return perBound;
+	}
 
 	/**
 	 * The socket read timeout of one connection, and the statements running on it. This second
@@ -1326,7 +1351,149 @@
 
 	/** That bound in seconds, as configured, read by the reader every bound of this backend shares. */
 	static int ddlLockBoundSeconds() {
-		return boundSeconds(DDL_LOCK_TIMEOUT_PROPERTY, COMMENT_LOCK_TIMEOUT_SECONDS);
+		return LockBound.DDL.seconds();
+	}
+
+	/**
+	 * The bound on the wait of a transaction of {@link #write} for a row lock another session holds, in
+	 * seconds. A value of {@code 0}, or a negative one, leaves that wait to the engine, which is what
+	 * this backend did before this bound existed (#915).
+	 * <p>
+	 * The replay of {@link #write} is bounded by a wall clock ({@link #RETRY_WINDOW_NANOS}), and a
+	 * clock cannot bound a wait that nothing else bounds: the lock wait ahead of a conflict is charged
+	 * to the attempt that hit it, so an attempt alone outlasting the window left the operation with no
+	 * replay at all, however transient its conflict (#903). Three engines out of four wait for a row
+	 * lock with no bound of their own ({@code LOCK_TIMEOUT} is -1 on sql server, {@code lock_timeout} is
+	 * 0 on postgres, and oracle's enqueue is unlimited), and the fourth bounds it at 50 s - five times
+	 * the window, which is the same thing as far as the window is concerned. Bounded here, an attempt
+	 * ends inside the window and the window governs the replay again, as it was designed to.
+	 * <p>
+	 * The default is deliberately a small fraction of that window: an attempt that ends at this bound
+	 * spends this much of it, so the replays a conflicted write gets are {@code RETRY_WINDOW_NANOS}
+	 * divided by this value. At the shipped 3 s a write that keeps meeting a lock gets three attempts
+	 * inside the window; raised past the window it would leave exactly one, which is the shape #903
+	 * describes.
+	 * <p>
+	 * What it changes for a deployment is that a write blocked behind a long writer of another session
+	 * now fails once the window is spent, where it used to wait as long as it took and then succeed.
+	 * That is the trade this bound is: an operation held for the length of somebody else's transaction
+	 * cannot be told from one that is never coming back, and the caller of a write is an LDAP client
+	 * with a timeout of its own. A deployment that would rather wait sets this to 0 and keeps the old
+	 * behaviour: the wait is the engine's, and a conflict behind it gets the one replay past the window
+	 * that {@link #grantedPastTheWindow} grants a wait nothing bounds (#904), and no more.
+	 * <p>
+	 * Oracle is not one of the engines this is put on: it has no session-level bound for the enqueue a
+	 * row lock waits in under plain DML - {@code ddl_lock_timeout} is the DDL lock and
+	 * {@code distributed_lock_timeout} the distributed transaction, neither of which is this wait, and
+	 * {@code select ... for update wait} is not a statement this backend issues. An attempt there stays
+	 * bounded only by {@link StatementBound#OPERATION}, whose cancel a session blocked in that enqueue
+	 * does not act on - see {@link #bounded(PreparedStatement, StatementBound, Execution)} - so on oracle
+	 * what answers #903 is not this property but the grant of #904, whatever this property says.
+	 * <p>
+	 * What it costs is the round trips around each write transaction - two on mysql and sql server (the
+	 * setting and the value given back; the readback is a third, and is paid once per pooled connection
+	 * rather than once per write, see {@link #displacedValue}), three on postgres (a {@code set local}
+	 * and the savepoint taken in front of it, which is let go of again once the setting is on), none on
+	 * oracle. Unlike the DDL bound this is the hot path, which is what the readback is cached for.
+	 */
+	static final String ROW_LOCK_TIMEOUT_PROPERTY="org.openidentityplatform.opendj.jdbc.row.lock.timeout";
+
+	/**
+	 * The default of {@link #ROW_LOCK_TIMEOUT_PROPERTY}: a third of {@link #RETRY_WINDOW_NANOS}, so
+	 * that a conflicted write is replayed rather than spending its whole window on one attempt.
+	 */
+	static final int ROW_LOCK_TIMEOUT_SECONDS=3;
+
+	/** That bound in seconds, as configured. */
+	static int rowLockBoundSeconds() {
+		return LockBound.ROW.seconds();
+	}
+
+	/**
+	 * A wait this backend bounds with a session setting, and the property that configures it. Two of
+	 * them: the lock a DDL of {@code openTree()} queues for, and the row lock a transaction of
+	 * {@link #write} queues for. One mechanism sets both ({@link #armLockBound}), and what differs
+	 * between them - which statement each engine takes, what it is called in a message, and where the
+	 * value comes from - is answered here rather than at the call sites.
+	 * <p>
+	 * The statements are asked of the {@link Dialect}, per constant, so that an engine added later
+	 * cannot compile without saying what it sets for both.
+	 */
+	enum LockBound {
+		/** The metadata, schema or DDL lock a {@code create table}, {@code create index} or {@code drop} waits for. */
+		DDL(DDL_LOCK_TIMEOUT_PROPERTY, COMMENT_LOCK_TIMEOUT_SECONDS, "a DDL",
+			"a DDL - the create index of an open whose table is already there, the drop table of a clear -") {
+			@Override
+			String boundSql(Dialect dialect, int seconds, Long previous) {
+				return dialect.ddlLockBoundSql(seconds, previous);
+			}
+
+			@Override
+			String query(Dialect dialect) {
+				return dialect.ddlLockBoundQuery();
+			}
+
+			@Override
+			String restoreSql(Dialect dialect, long previous) {
+				return dialect.ddlLockRestoreSql(previous);
+			}
+		},
+		/** The row lock a statement of a write transaction waits for. */
+		ROW(ROW_LOCK_TIMEOUT_PROPERTY, ROW_LOCK_TIMEOUT_SECONDS, "a write transaction",
+			"a write transaction - the upsert of an add, the delete of a remove -") {
+			@Override
+			String boundSql(Dialect dialect, int seconds, Long previous) {
+				return dialect.rowLockBoundSql(seconds, previous);
+			}
+
+			@Override
+			String query(Dialect dialect) {
+				return dialect.rowLockBoundQuery();
+			}
+
+			@Override
+			String restoreSql(Dialect dialect, long previous) {
+				return dialect.rowLockRestoreSql(previous);
+			}
+		};
+
+		final String property;
+		final int defaultSeconds;
+		/** What the message reporting this bound calls the thing that was waiting. */
+		final String waiter;
+		/**
+		 * The same, spelled out with what this backend issues under it: the one line an operator has to
+		 * act on says which statements are the ones left waiting, since the property alone does not say
+		 * what a deployment would see go slow.
+		 */
+		final String waiterInDetail;
+
+		LockBound(String property, int defaultSeconds, String waiter, String waiterInDetail) {
+			this.property=property;
+			this.defaultSeconds=defaultSeconds;
+			this.waiter=waiter;
+			this.waiterInDetail=waiterInDetail;
+		}
+
+		/** This bound in seconds, as configured, read by the reader every bound of this backend shares. */
+		int seconds() {
+			return boundSeconds(property, defaultSeconds);
+		}
+
+		/**
+		 * The setting bounding this wait on this engine, or null where there is none to put on: an engine
+		 * left to a bound of its own, and a session that already gives up sooner than this one would.
+		 *
+		 * @param previous what the session carries now, as {@link #query(Dialect)} read it, in the unit
+		 *                 that query answers in - or null where nothing was read back
+		 */
+		abstract String boundSql(Dialect dialect, int seconds, Long previous);
+
+		/** What the session carries now, or null where this bound undoes itself. */
+		abstract String query(Dialect dialect);
+
+		/** The setting giving the session back what {@link #query(Dialect)} read off it. */
+		abstract String restoreSql(Dialect dialect, long previous);
 	}
 
 	// The comment statement runs on a connection of its own (newStampConnection() below), and a
@@ -1393,10 +1560,47 @@
 				return null;
 			}
 
+			// the same setting and the same set local, at this bound's own value: lock_timeout bounds the
+			// wait for any lock of the transaction, the row locks of an upsert included, and the commit or
+			// the rollback ending that transaction is what takes it off again. A write() attempt is exactly
+			// such a transaction, so this engine pays no readback and leaves nothing on a pooled connection.
+			// What ends it early is a DDL of the same attempt, which commits (commitStatement): the rest of
+			// that write runs unbounded here, and is given the same treatment on the two engines whose
+			// setting is the session's - commitStatement takes it off there rather than leaving it on. It is
+			// an attempt that has committed part of its work, so it is out of the replay whatever it waits
+			// for, and a bound on a wait nothing can replay would only fail a write that used to go through,
+			// which is why the bound is not armed again behind it.
+			@Override
+			String rowLockBoundSql(int seconds, Long previous) {
+				return "set local lock_timeout = "+seconds*1000L;
+			}
+
+			@Override
+			String rowLockBoundQuery() {
+				return null; // set local: the transaction of the attempt discards it
+			}
+
+			@Override
+			String rowLockRestoreSql(long previous) {
+				return null;
+			}
+
+			@Override
+			String rowLockLiftSql() {
+				// the set local has no value to give back, so the wait of the rest of this transaction is
+				// put back where a transaction that set nothing would have left it
+				return "set local lock_timeout to default";
+			}
+
 			@Override
 			boolean boundLivesInTheTransaction() {
 				return true;
 			}
+
+			@Override
+			boolean oneSettingForBothBounds() {
+				return true; // lock_timeout, for every lock wait of the transaction
+			}
 		},
 		/** mysql: lock_wait_timeout takes seconds; connectTimeout bounds the socket connect and socketTimeout every read after it, both in milliseconds. */
 		MYSQL("set session lock_wait_timeout="+COMMENT_LOCK_TIMEOUT_SECONDS,
@@ -1421,6 +1625,28 @@
 			String ddlLockRestoreSql(long previous) {
 				return "set session lock_wait_timeout="+previous;
 			}
+
+			// innodb_lock_wait_timeout, in seconds, and never the lock_wait_timeout above it: the first is
+			// the row lock a write waits for, the second the metadata lock a DDL waits for, and they are
+			// two settings with two defaults. This is the one engine of the four bounding this wait by
+			// itself - at 50 s, five times the replay window, so a conflict reported at it arrives with the
+			// window already spent and is never replayed. A session giving up sooner keeps what it has: the
+			// variable has no encoding for "wait forever" to be mistaken for a tight value, its range
+			// starting at 1.
+			@Override
+			String rowLockBoundSql(int seconds, Long previous) {
+				return (previous!=null && previous<=seconds) ? null : "set session innodb_lock_wait_timeout="+seconds;
+			}
+
+			@Override
+			String rowLockBoundQuery() {
+				return "select @@session.innodb_lock_wait_timeout";
+			}
+
+			@Override
+			String rowLockRestoreSql(long previous) {
+				return "set session innodb_lock_wait_timeout="+previous;
+			}
 		},
 		/** oracle: ddl_lock_timeout takes seconds and defaults to 0 (give up at once), but it can be raised globally; the connect and read bounds take milliseconds. */
 		ORACLE("alter session set ddl_lock_timeout="+COMMENT_LOCK_TIMEOUT_SECONDS,
@@ -1443,6 +1669,29 @@
 			String ddlLockRestoreSql(long previous) {
 				return null;
 			}
+
+			// There is no session setting bounding the enqueue a row lock waits in under plain DML here:
+			// ddl_lock_timeout is the DDL lock above, distributed_lock_timeout is the distributed
+			// transaction, and the wait can only be named on the statement itself - "select ... for update
+			// wait n", which is not a statement this backend issues. So an attempt of write() on this engine
+			// is bounded by StatementBound.OPERATION alone, and a session blocked in that enqueue does not
+			// act on the cancel behind it: what ends such a wait is the socket read timeout, which takes the
+			// connection with it. Nothing of ours is set here rather than something that would not bound the
+			// wait, so a lock timeout of this engine is nothing this backend claims to have bounded.
+			@Override
+			String rowLockBoundSql(int seconds, Long previous) {
+				return null;
+			}
+
+			@Override
+			String rowLockBoundQuery() {
+				return null; // nothing of ours is set on it
+			}
+
+			@Override
+			String rowLockRestoreSql(long previous) {
+				return null;
+			}
 		},
 		/** ms sql server: lock_timeout takes milliseconds; loginTimeout bounds the socket connect, in seconds, and socketTimeout the prelogin read it leaves open - and every read after it - in milliseconds. */
 		MICROSOFT("set lock_timeout "+(COMMENT_LOCK_TIMEOUT_SECONDS*1000),
@@ -1466,6 +1715,33 @@
 			String ddlLockRestoreSql(long previous) {
 				return "set lock_timeout "+previous;
 			}
+
+			// the same session setting as the DDL bound above - this engine has one LOCK_TIMEOUT for every
+			// lock wait of a session - at this bound's own value, and put back the moment the transaction is
+			// through for the very reason that it is one setting: a value left behind cuts the lock waits of
+			// whoever borrows the connection next, and a read borrowing it has no replay to absorb the
+			// error 1222 it would then see. -1 is "wait forever" rather than a bound tighter than ours and
+			// is replaced; 0 - "do not wait at all" - is tighter, and a session carrying it is left alone.
+			@Override
+			String rowLockBoundSql(int seconds, Long previous) {
+				final long millis=seconds*1000L;
+				return (previous!=null && previous>=0 && previous<=millis) ? null : "set lock_timeout "+millis;
+			}
+
+			@Override
+			String rowLockBoundQuery() {
+				return "select @@lock_timeout";
+			}
+
+			@Override
+			String rowLockRestoreSql(long previous) {
+				return "set lock_timeout "+previous;
+			}
+
+			@Override
+			boolean oneSettingForBothBounds() {
+				return true; // LOCK_TIMEOUT, for every lock wait of the session
+			}
 		};
 
 		final String lockTimeoutSql;
@@ -1520,6 +1796,44 @@
 		abstract String ddlLockRestoreSql(long previous);
 
 		/**
+		 * The session setting bounding the wait of a write transaction for a row lock another session
+		 * holds, or null where there is none to put on: an engine with no such setting - oracle - and a
+		 * session that already gives up sooner than this one would.
+		 * <p>
+		 * A different setting from {@link #ddlLockBoundSql} on the two engines that have two of them, and
+		 * literally the same one on the two that have one: what tells the two bounds apart is the wait
+		 * each is put around, not the statement each issues.
+		 *
+		 * @param previous what the session carries now, as {@link #rowLockBoundQuery()} read it, in the
+		 *                 unit that query answers in - or null where nothing was read back
+		 */
+		abstract String rowLockBoundSql(int seconds, Long previous);
+
+		/**
+		 * What the session carries now, asked before the bound above displaces it - or null where that
+		 * bound undoes itself with the transaction it belongs to.
+		 */
+		abstract String rowLockBoundQuery();
+
+		/** The setting that gives the session back the value {@link #rowLockBoundQuery()} read off it. */
+		abstract String rowLockRestoreSql(long previous);
+
+		/**
+		 * What takes the row lock bound off where there is no displaced value to give back, so that the
+		 * wait it was cutting runs as the deployment left it: a {@code set local} reads nothing back -
+		 * the transaction it belongs to is what discards it - so what puts the setting where it would
+		 * have been is its own default. Null on the engines whose bound is a session setting and whose
+		 * release has a value in hand ({@link #rowLockRestoreSql}), and on those that arm nothing.
+		 * <p>
+		 * Read by the DDL of a write whose own bound is turned off, which is the one place a row bound of
+		 * this backend has to come off before the transaction it belongs to ends - see
+		 * {@code liftTheRowLockBoundForAnUnboundedDdl()}.
+		 */
+		String rowLockLiftSql() {
+			return null;
+		}
+
+		/**
 		 * Whether the bound belongs to the transaction running the DDL rather than to the session. Such a
 		 * setting needs a transaction block to take effect at all, a rollback undoes it, and the commit
 		 * ending the DDL is what takes it off again - so nothing is read back and nothing is put back.
@@ -1527,6 +1841,22 @@
 		boolean boundLivesInTheTransaction() {
 			return false;
 		}
+
+		/**
+		 * Whether one setting of this engine bounds both waits, so that arming either displaces the other.
+		 * True of sql server, whose {@code LOCK_TIMEOUT} bounds every lock wait of a session, and of
+		 * postgres, whose {@code lock_timeout} bounds every lock wait of a transaction; false of mysql,
+		 * which has {@code lock_wait_timeout} for the metadata lock and {@code innodb_lock_wait_timeout}
+		 * for the row lock, and of oracle, which takes neither bound.
+		 * <p>
+		 * What reads it is {@link #armLockBound}, deciding whether a session already gives up sooner than
+		 * the bound being armed: on such an engine the value read back can be this backend's own - the
+		 * row bound of the write a DDL runs inside - and a bound of ours is no reason to leave the other
+		 * one unarmed.
+		 */
+		boolean oneSettingForBothBounds() {
+			return false;
+		}
 	}
 
 	/** Returns the class name of the driver behind the given connection, which names the engine it talks to. */
@@ -2157,120 +2487,310 @@
 
 	/**
 	 * Runs a DDL of this backend under {@link #DDL_LOCK_TIMEOUT_PROPERTY}, and gives the session back
-	 * whatever it carried before.
+	 * whatever it carried before. What is put on the session, and what that costs where it cannot be,
+	 * is {@link #armLockBound}: this is the one statement version of it, where {@link #write} arms the
+	 * same mechanism around a whole transaction.
 	 * <p>
-	 * The dialect is passed in rather than read off the connection here, the way
-	 * {@code commentTable()} takes it: the callers know it, and taking it as a parameter is what makes
-	 * this reachable from a test with no database behind it.
-	 * <p>
-	 * Nothing of ours is set where it could not be taken off again, and a failure of the readback is
-	 * never the failure of the DDL: this runs mostly on a pooled connection, so a setting left behind
-	 * reaches every statement of whoever borrows it next - on sql server that is every lock wait of
-	 * theirs, row locks included, and {@link #isConflict} classifies error 1222 as no replayable
-	 * conflict. A session this backend could not take its bound off again is kept out of the pool for
-	 * that reason ({@link CachedConnection#keepOutOfThePool}), since the validation of the next borrow
-	 * is {@code isValid()} - a liveness check a connection carrying a stale setting passes. The one
-	 * connection here that is not the pool's is the catalog's own, which {@code createCatalogTable()}
-	 * creates its table on: there a setting left behind reaches the rest of that write and goes with
-	 * the connection, which is closed with it.
-	 * <p>
-	 * The DDL runs whatever any of that did, and it runs under the same rewrite either way: a setting
-	 * can reach the server and fail only as the statement carrying it is closed, which no driver tells
-	 * apart from a setting that never arrived, and reporting a DDL that then really did give up at this
-	 * bound as the bare 55P03 it arrives as is the gap this exists to close.
+	 * The DDL runs whatever the arming did, and where a setting of ours was issued it runs under a
+	 * rewrite either way: a setting can reach the server and fail only as the statement carrying it is
+	 * closed, which no driver tells apart from a setting that never arrived, and reporting a DDL that
+	 * then really did give up at this bound as the bare 55P03 it arrives as is the gap that rewrite
+	 * exists to close.
 	 */
 	<T> T withDdlLockBound(Connection con, Dialect dialect, Execution<T> action) throws SQLException {
-		final int seconds=ddlLockBoundSeconds();
-		if (seconds<=0) { // the wait is left exactly as unbounded as it was, and nobody asked otherwise
-			return action.run();
-		}
-		if (dialect==null) {
-			// The one branch where a bound was asked for and none is put on, which is why it is the one
-			// that says so: an engine none of these settings fit is fed none of them - untested SQL is no
-			// thing to send a database on the path a backend opens by - and the silence around that is
-			// what an operator has no way of finding out.
-			reportTheEngineIsNotKnown(con);
-			return action.run();
-		}
-		// Asked with nothing displaced yet, which is what tells an engine this bound is never put on -
-		// oracle - from an engine it is put on. What the session actually carries is read below, and can
-		// take the bound off again all by itself.
-		if (dialect.ddlLockBoundSql(seconds, null)==null) {
-			return action.run();
-		}
-		final String query=dialect.ddlLockBoundQuery();
-		final Long previous=(query==null) ? null : sessionValue(con, dialect, query);
-		if (query!=null && previous==null) { // read it back first: see above
-			return action.run();
-		}
-		// Asked again with it: a session already giving up sooner than this bound is left exactly as it
-		// is, rather than loosened to ours for the length of the DDL. That is the argument leaving oracle
-		// alone, applied where the displaced value is in hand and costs nothing to respect.
-		final String bound=dialect.ddlLockBoundSql(seconds, previous);
-		if (bound==null) {
-			return action.run();
-		}
-		if (dialect.boundLivesInTheTransaction() && !inATransactionBlock(con, dialect, bound)) {
-			return action.run();
-		}
-		final String restore=(previous==null) ? null : dialect.ddlLockRestoreSql(previous);
-		// A statement that fails inside a postgres transaction aborts it, and everything after it - the
-		// DDL included - then fails with 25P02 rather than running "unbounded, as before": a backend that
-		// opened before this bound existed would stop opening because of the bound meant to protect it.
-		// The rollback below goes back to here, which also undoes a set local that did reach the server,
-		// so the DDL really does run as unbounded as the warning says it does.
-		final Savepoint beforeTheBound=savepointBeforeTheBound(con, dialect, bound);
-		try {
-			boundedSessionCall(con, () -> {
-				executeSessionStatement(con, bound);
-				return null;
-			});
-		}catch (SQLException | RuntimeException e) {
-			// The bound is an improvement on a wait, and never a reason to fail a DDL that would have gone
-			// through: a backend that opened before this bound existed has to open still. Whatever the
-			// setting displaced is given back by the finally below, whether it went on or not - giving back
-			// a value the session may never have left costs a round trip and changes nothing.
-			reportTheWaitIsLeftUnbounded(dialect, bound, e);
-			rollbackTheBound(con, dialect, beforeTheBound);
-		}
-		// From here rather than from the top of this method: what the statements above spent is not time
-		// the DDL waited for its lock, and it is the wait that this bound either ended or did not.
-		final long startedAt=nanoTime();
+		final ArmedLockBound armed=armLockBound(con, dialect, LockBound.DDL);
 		try {
 			return action.run();
 		}catch (SQLException e) {
-			throw gaveUpOnTheLock(e, dialect, seconds, startedAt);
+			// only where this bound's own figure is what the session is carrying: a wait this backend put
+			// no bound on was ended by something else, and naming this property for it sends an operator
+			// to a value that changes nothing - the session that already gave up sooner, the engine that
+			// has no such setting, and the deployment whose own value this arming put back on
+			throw armed.ourFigure ? gaveUpOnTheLock(e, dialect, armed.seconds, armed.startedAt) : e;
 		}catch (RuntimeException e) {
 			// Not every statement running under this bound answers with the SQLException it was given:
 			// the lookup deciding each drop of a clear wraps whatever it sees in a
 			// StorageRuntimeException (isExistsTable), and it runs inside the same bound as the drop it
 			// decides. Without this, a lock this bound ended reaches an operator as the bare vendor error
 			// one line away from the drop that would have named the property.
-			throw gaveUpOnTheLock(e, dialect, seconds, startedAt);
+			throw armed.ourFigure ? gaveUpOnTheLock(e, dialect, armed.seconds, armed.startedAt) : e;
 		}finally {
-			if (restore!=null) {
-				restoreDdlLockBound(con, dialect, bound, restore);
-			}
+			releaseLockBound(con, dialect, armed);
 		}
 	}
 
 	/**
+	 * A lock bound of this backend as it stands on one session: what was put on it, what gives it back,
+	 * and whether the wait it was put around is bounded at all. Produced by {@link #armLockBound} and
+	 * handed to {@link #releaseLockBound}, which is what lets {@link #write} arm one around a whole
+	 * transaction where {@link #withDdlLockBound} wraps a single statement.
+	 */
+	static final class ArmedLockBound {
+		final LockBound kind;
+		/** The configured value this was armed at, in seconds - 0 where nothing of ours was put on. */
+		final int seconds;
+		/** The setting issued, or null where none was: an engine or a session that needed none. */
+		final String bound;
+		/** What gives the session its value back, or null where the bound takes itself off. */
+		final String restore;
+		/**
+		 * Whether the wait this was armed around is bounded at no more than {@link LockBound#seconds()}
+		 * - by the setting above, or by a session value that was already tighter and was left alone.
+		 * False wherever this backend could not put a bound on and does not know of one: an engine with
+		 * no such setting, the property turned off, a readback that failed, and a setting the session
+		 * refused. It is what {@link #replayReason} reads: a wait nothing bounds must not be replayed on
+		 * a clock, which is the whole of #903.
+		 */
+		final boolean bounded;
+		/**
+		 * Whether what the session carries under this bound is this bound's own figure, which is what
+		 * lets a failure of the work be reported as this bound's doing
+		 * ({@link #gaveUpOnTheLock(SQLException, Dialect, int, long)}). False
+		 * where a statement of ours was issued all the same: the value put on is the one the deployment
+		 * set, restored over a bound of ours that was tighter than it, and naming
+		 * {@link #DDL_LOCK_TIMEOUT_PROPERTY} for a wait that value ended would send an operator to raise
+		 * a property that governs nothing here.
+		 */
+		final boolean ourFigure;
+		/**
+		 * When the wait itself began, as {@link #nanoTime} reads it: after the round trips of the arming,
+		 * which are not time anything waited for a lock. Zero where {@link #bound} is null, since nothing
+		 * is ever measured against a setting that was not put on.
+		 */
+		final long startedAt;
+
+		private ArmedLockBound(LockBound kind, int seconds, String bound, String restore, boolean bounded,
+				boolean ourFigure, long startedAt) {
+			this.kind=kind;
+			this.seconds=seconds;
+			this.bound=bound;
+			this.restore=restore;
+			this.bounded=bounded;
+			this.ourFigure=ourFigure;
+			this.startedAt=startedAt;
+		}
+
+		/** Nothing was put on and nothing is known to bound this wait: it runs as it did before #885. */
+		static ArmedLockBound none(LockBound kind) {
+			return new ArmedLockBound(kind, 0, null, null, false, false, 0);
+		}
+
+		/**
+		 * The session already gives up sooner than this bound would, so it keeps what it has. Nothing is
+		 * set and nothing has to be put back - and the wait is bounded, which is the whole point of
+		 * leaving that value alone.
+		 */
+		static ArmedLockBound alreadyTighter(LockBound kind, int seconds) {
+			return new ArmedLockBound(kind, seconds, null, null, true, false, 0);
+		}
+	}
+
+	/**
+	 * Puts a lock bound of this backend on a session, and never fails the work it was put around.
+	 * <p>
+	 * The dialect is passed in rather than read off the connection here, the way {@code commentTable()}
+	 * takes it: the callers know it, and taking it as a parameter is what makes this reachable from a
+	 * test with no database behind it.
+	 * <p>
+	 * Nothing of ours is set where it could not be taken off again, and a failure of the readback is
+	 * never the failure of the work: this runs mostly on a pooled connection, so a setting left behind
+	 * reaches every statement of whoever borrows it next - on sql server that is every lock wait of
+	 * theirs, and a read has no replay to absorb the error 1222 it would then see. A session this
+	 * backend could not take its bound off again is kept out of the pool for that reason
+	 * ({@link CachedConnection#keepOutOfThePool}), since the validation of the next borrow is
+	 * {@code isValid()} - a liveness check a connection carrying a stale setting passes. The one
+	 * connection here that is not the pool's is the catalog's own, which {@code createCatalogTable()}
+	 * creates its table on: there a setting left behind reaches the rest of that write and goes with
+	 * the connection, which is closed with it.
+	 */
+	ArmedLockBound armLockBound(Connection con, Dialect dialect, LockBound kind) {
+		final int seconds=kind.seconds();
+		if (seconds<=0) { // the wait is left exactly as unbounded as it was, and nobody asked otherwise
+			return ArmedLockBound.none(kind);
+		}
+		if (dialect==null) {
+			// The one branch where a bound was asked for and none is put on, which is why it is the one
+			// that says so: an engine none of these settings fit is fed none of them - untested SQL is no
+			// thing to send a database on the path a backend opens by - and the silence around that is
+			// what an operator has no way of finding out.
+			reportTheEngineIsNotKnown(con, kind);
+			return ArmedLockBound.none(kind);
+		}
+		// Asked first with nothing displaced yet, which is what tells an engine this bound is never put
+		// on - oracle - from an engine it is put on. What the session actually carries is read below, and
+		// can take the bound off again all by itself.
+		if (kind.boundSql(dialect, seconds, null)==null) {
+			return ArmedLockBound.none(kind);
+		}
+		final String query=kind.query(dialect);
+		final Long previous=(query==null) ? null : displacedValue(con, dialect, kind, query);
+		if (query!=null && previous==null) { // read it back first: see above
+			return ArmedLockBound.none(kind);
+		}
+		// Asked again with it: a session already giving up sooner than this bound is left exactly as it
+		// is, rather than loosened to ours for the length of the work. That is the argument leaving oracle
+		// alone, applied where the displaced value is in hand and costs nothing to respect - and asked of
+		// what the deployment set rather than of what the session happens to carry, since on an engine
+		// with one setting for both waits the value read back a moment ago can be a bound of ours.
+		final Long deployment=carriedByTheDeployment(con, dialect, kind, previous);
+		final String atOurFigure=kind.boundSql(dialect, seconds, deployment);
+		// Where the deployment gives up sooner than this bound would, "it keeps what it has" is the whole
+		// answer only while what it has is what that decision was made against. It is not, on the one
+		// road where the value read back is a bound of ours: then the work would run at neither figure -
+		// see below, which puts the deployment's own value back on for the length of it.
+		final String bound=(atOurFigure!=null) ? atOurFigure : theDeploymentsOwnValue(kind, dialect, deployment,
+			previous);
+		if (bound==null) {
+			return ArmedLockBound.alreadyTighter(kind, seconds);
+		}
+		// whether what goes on the session is this bound's own figure, which is what lets a failure of the
+		// work be named as this bound's doing
+		final boolean ourFigure=(atOurFigure!=null);
+		if (dialect.boundLivesInTheTransaction() && !inATransactionBlock(con, dialect, kind, bound)) {
+			return ArmedLockBound.none(kind);
+		}
+		final String restore=(previous==null) ? null : kind.restoreSql(dialect, previous);
+		// A statement that fails inside a postgres transaction aborts it, and everything after it - the
+		// work this bound was put around included - then fails with 25P02 rather than running "unbounded,
+		// as before": a backend that opened before this bound existed would stop opening because of the
+		// bound meant to protect it. The rollback below goes back to here, which also undoes a set local
+		// that did reach the server, so the work really does run as unbounded as the warning says it does.
+		final Savepoint beforeTheBound=savepointBeforeTheBound(con, dialect, kind, bound);
+		try {
+			boundedSessionCall(con, () -> {
+				executeSessionStatement(con, bound);
+				return null;
+			});
+		}catch (SQLException | RuntimeException e) {
+			// The bound is an improvement on a wait, and never a reason to fail work that would have gone
+			// through: a backend that opened before this bound existed has to open still, and a write that
+			// used to wait for its row lock has to be able to wait for it. Whatever the setting displaced is
+			// given back by the release, whether it went on or not - giving back a value the session may
+			// never have left costs a round trip and changes nothing.
+			reportTheWaitIsLeftUnbounded(dialect, kind, bound, e);
+			rollbackTheBound(con, dialect, beforeTheBound);
+			// The setting is still named as issued, so that the release puts the value back and a failure
+			// that did give up at this bound is still renamed - a setting can reach the server and fail only
+			// as the statement carrying it is closed, and no driver tells that apart from one that never
+			// arrived. What is not claimed is that the wait is bounded: a replay decided on that claim would
+			// be a replay of a wait nothing ends.
+			return new ArmedLockBound(kind, seconds, bound, restore, false, ourFigure, nanoTime());
+		}
+		// The setting is on and there is nothing left to take back, so the subtransaction that net opened
+		// is let go of rather than left to span the work - see there.
+		releaseTheSavepoint(con, dialect, beforeTheBound);
+		// From here rather than from the top of this method: what the round trips above spent is not time
+		// anything waited for a lock, and it is that wait this bound either ends or does not.
+		return new ArmedLockBound(kind, seconds, bound, restore, true, ourFigure, nanoTime());
+	}
+
+	/**
+	 * What the deployment set, to be put on over a bound of this backend that is cutting the wait
+	 * tighter than the deployment asked for - or null where there is nothing to put on.
+	 * <p>
+	 * Asked where a session already gives up sooner than the bound being armed. That is normally the
+	 * end of it: the session keeps its own value, which is the argument that leaves oracle alone. The
+	 * one road where it is not is the DDL of a write on an engine with one setting for both waits: what
+	 * the session carries there is the row bound of that write, armed one statement earlier and tighter
+	 * than what the deployment set, so leaving it alone runs the DDL at a figure nobody asked for and
+	 * hands its failure through bare. The value the row bound displaced goes back on for the length of
+	 * the DDL, and the release puts the row bound back behind it - nothing is loosened past what the
+	 * deployment allows, which is what the decision above already established.
+	 */
+	private String theDeploymentsOwnValue(LockBound kind, Dialect dialect, Long deployment, Long previous) {
+		if (deployment==null || deployment.equals(previous)) {
+			return null; // the session carries what the decision was made against, and keeps it
+		}
+		return kind.restoreSql(dialect, deployment);
+	}
+
+	/** Gives the session back what an armed bound displaced, where it displaced anything. */
+	void releaseLockBound(Connection con, Dialect dialect, ArmedLockBound armed) {
+		if (armed.restore!=null) {
+			restoreLockBound(con, dialect, armed);
+		}
+	}
+
+	/**
+	 * What the session carries before a bound of ours displaces it, read once per pooled connection for
+	 * the bound that is on the hot path.
+	 * <p>
+	 * The readback is a round trip, and the row lock bound takes one around every write of the server
+	 * where the DDL bound takes one around an open. It can be remembered because only this backend
+	 * writes that setting on a connection of this pool: every write puts the value back before the
+	 * connection is released, and a connection whose restore failed is kept out of the pool rather than
+	 * handed on - so a remembered value cannot outlive the session that answered it. The bound of a DDL
+	 * is read every time all the same: it runs inside a write which may be carrying the row bound of
+	 * this backend on the very same setting (sql server has one {@code LOCK_TIMEOUT} for both), and what
+	 * it has to put back is that value rather than the one the connection was borrowed with.
+	 */
+	private Long displacedValue(Connection con, Dialect dialect, LockBound kind, String query) {
+		if (kind!=LockBound.ROW || !(con instanceof CachedConnection)) {
+			return sessionValue(con, dialect, kind, query);
+		}
+		final CachedConnection pooled=(CachedConnection) con;
+		final Long remembered=pooled.rowLockBoundDisplaces();
+		if (remembered!=null) {
+			return remembered;
+		}
+		final Long read=sessionValue(con, dialect, kind, query);
+		if (read!=null) {
+			pooled.rowLockBoundDisplaces(read);
+		}
+		return read;
+	}
+
+	/**
+	 * What the session carried before this backend put anything on it, which is what decides whether it
+	 * already gives up sooner than the bound being armed - and which is not what the readback answers
+	 * where a bound of ours is already on that very setting.
+	 * <p>
+	 * sql server has one {@code LOCK_TIMEOUT} for both waits, and every DDL of this backend but the
+	 * off-write catalog drop is issued from inside a write, which arms the row bound one statement
+	 * earlier. Read live, that DDL sees this backend's own 3 s, answers "already tighter" to its own
+	 * 5 s and runs at the row bound instead: {@link #DDL_LOCK_TIMEOUT_PROPERTY} would govern no DDL of
+	 * a write at all - a deployment raising it for an index build blocked by long readers would still
+	 * get the row bound - and the 1222 such a DDL gives up with would be reported as the bare vendor
+	 * error, the rename of #885 lost with it. What the row bound displaced is remembered on the
+	 * connection for the readback above, so the value the deployment set is in hand here.
+	 * <p>
+	 * Only where one setting carries both waits ({@link Dialect#oneSettingForBothBounds}). On mysql the
+	 * two are different variables - {@code lock_wait_timeout} for the metadata lock,
+	 * {@code innodb_lock_wait_timeout} for the row lock - so the remembered value describes neither the
+	 * other's session nor its default, and reading it there would leave a metadata lock waiting a year
+	 * because a deployment had tightened the row one.
+	 * <p>
+	 * The live readback stays what the release puts back: what a session has to be given back is what it
+	 * carried a statement ago, bound of ours or not. Where the two differ the session is carrying a
+	 * bound of ours rather than the value this decides against, so "already tighter" would leave the
+	 * work at neither figure - see {@link #armLockBound}, which puts the deployment's own value on for
+	 * the length of it instead.
+	 */
+	private Long carriedByTheDeployment(Connection con, Dialect dialect, LockBound kind, Long previous) {
+		if (kind==LockBound.ROW || !dialect.oneSettingForBothBounds() || !(con instanceof CachedConnection)) {
+			return previous;
+		}
+		final Long displacedByTheRowBound=((CachedConnection) con).rowLockBoundDisplaces();
+		return (displacedByTheRowBound!=null) ? displacedByTheRowBound : previous;
+	}
+
+	/**
 	 * Whether a setting that lives in the transaction would take effect on this connection at all.
 	 * Outside a transaction block postgres answers {@code SET LOCAL} with a warning and does nothing
-	 * with it - the driver raises nothing, so the DDL would run with no bound and the log would read
+	 * with it - the driver raises nothing, so the work would run with no bound and the log would read
 	 * exactly like a bounded one. What makes the block is the {@code setAutoCommit(false)} a pooled
 	 * connection is established with and never re-asserts on borrow, so it is asked rather than
 	 * assumed; the call is answered out of the driver's own state, not by a round trip.
 	 */
-	private boolean inATransactionBlock(Connection con, Dialect dialect, String bound) {
+	private boolean inATransactionBlock(Connection con, Dialect dialect, LockBound kind, String bound) {
 		try {
 			if (!con.getAutoCommit()) {
 				return true;
 			}
-			reportTheWaitIsLeftUnbounded(dialect, bound, new SQLException("the connection is in auto-commit,"
+			reportTheWaitIsLeftUnbounded(dialect, kind, bound, new SQLException("the connection is in auto-commit,"
 				+ " where this setting belongs to no transaction and the server answers it with a warning"));
 		}catch (SQLException | RuntimeException e) {
-			reportTheWaitIsLeftUnbounded(dialect, bound, e);
+			reportTheWaitIsLeftUnbounded(dialect, kind, bound, e);
 		}
 		return false;
 	}
@@ -2281,19 +2801,49 @@
 	 * give a savepoint. Nothing of the bound has been set when this is asked, so a failure here only
 	 * leaves the setting below without this net - which is where it was before the net existed.
 	 */
-	private Savepoint savepointBeforeTheBound(Connection con, Dialect dialect, String bound) {
+	private Savepoint savepointBeforeTheBound(Connection con, Dialect dialect, LockBound kind, String bound) {
 		if (!dialect.boundLivesInTheTransaction()) {
 			return null;
 		}
 		try {
 			return boundedSessionCall(con, con::setSavepoint);
 		}catch (SQLException | RuntimeException e) {
-			reportTheWaitIsLeftUnbounded(dialect, bound, e);
+			reportTheWaitIsLeftUnbounded(dialect, kind, bound, e);
 			return null;
 		}
 	}
 
 	/**
+	 * Lets go of that point once the setting is on and there is nothing left to take back.
+	 * <p>
+	 * A savepoint is a subtransaction of the one the work runs in, and one left open spans that work:
+	 * on the hot path that is every write of the server, whose rows then carry the subtransaction's own
+	 * xid rather than the transaction's - which every reader of those rows resolves through
+	 * {@code pg_subtrans} until they are hinted. It costs a round trip to let go of, which is the trade
+	 * this makes: one statement per write against a subtransaction per write. A {@code SET LOCAL}
+	 * survives the release, so the bound this was taken in front of stays on for the rest of the
+	 * transaction - which is the whole of what it was armed around.
+	 * <p>
+	 * Best effort, like every other round trip of the bound: a savepoint that could not be let go of
+	 * costs a subtransaction, never the work.
+	 */
+	private void releaseTheSavepoint(Connection con, Dialect dialect, Savepoint beforeTheBound) {
+		if (beforeTheBound==null) {
+			return;
+		}
+		try {
+			boundedSessionCall(con, () -> {
+				con.releaseSavepoint(beforeTheBound);
+				return null;
+			});
+		}catch (SQLException | RuntimeException e) {
+			logger.trace(LocalizableMessage.raw("jdbc: the point in front of a lock bound could not be let go of on"
+				+ " this %s database, leaving the work it was taken for inside a subtransaction: %s", dialect,
+				stackTraceToSingleLineString(e)));
+		}
+	}
+
+	/**
 	 * Takes the transaction back to the point before the bound was set: it clears the abort a setting
 	 * that failed inside it left behind, and it undoes a {@code set local} that did reach the server
 	 * and failed only as its statement was closed. Best effort - a transaction that cannot be taken
@@ -2346,11 +2896,11 @@
 	/**
 	 * What a session setting of this engine carries right now, or null where it could not be read as a
 	 * number: a server whose session does not have the variable, or one answering with something no
-	 * {@code SET} of it would take back. The DDL then runs as unbounded as it was before this bound
-	 * existed, which is why this is reported rather than thrown - and reported once, since every DDL
-	 * of that backend would say the same thing.
+	 * {@code SET} of it would take back. The work this bound was to be put around then runs as unbounded
+	 * as it ran before the bound existed, which is why this is reported rather than thrown - and reported
+	 * once per bound, since every DDL, or every write, of that backend would say the same thing.
 	 */
-	private Long sessionValue(Connection con, Dialect dialect, String query) {
+	private Long sessionValue(Connection con, Dialect dialect, LockBound kind, String query) {
 		if (logger.isTraceEnabled()) {
 			logger.trace(LocalizableMessage.raw("jdbc: %s",query));
 		}
@@ -2364,70 +2914,79 @@
 				}
 			});
 		}catch (SQLException | RuntimeException e) { // a value that is not a number arrives unchecked
-			reportTheWaitIsLeftUnbounded(dialect, query, e);
+			reportTheWaitIsLeftUnbounded(dialect, kind, query, e);
 			return null;
 		}
 	}
 
 	/**
-	 * Said once per storage, whichever round trip of the bound around a DDL the connection would not
-	 * take: every DDL of that backend would say the same thing, and a backend opening its trees issues
-	 * about 25 of them. The DDL itself is unaffected - it waits as it did before this bound existed.
+	 * Said once per storage and per bound, whichever round trip of the arming the connection would not
+	 * take: every DDL of that backend would say the same thing - a backend opening its trees issues about
+	 * 25 of them - and so would every write. The work itself is unaffected: it waits as it did before this
+	 * bound existed.
 	 * <p>
 	 * "May bound nothing" rather than "bounds nothing": a setting that reached the server and failed
-	 * only as the statement carrying it was closed leaves the DDL bounded after all, and no driver says
+	 * only as the statement carrying it was closed leaves the wait bounded after all, and no driver says
 	 * which of the two happened. Where the bound lives in the transaction the rollback that follows
-	 * this settles it - there the DDL really does run unbounded.
+	 * this settles it - there the work really does run unbounded.
 	 */
-	private void reportTheWaitIsLeftUnbounded(Dialect dialect, String sql, Exception e) {
-		if (ddlLockBoundNotSetWarned.compareAndSet(false, true)) {
-			logger.warn(LocalizableMessage.raw("jdbc: the wait of a DDL for its lock is not bounded as this backend"
+	private void reportTheWaitIsLeftUnbounded(Dialect dialect, LockBound kind, String sql, Exception e) {
+		if (lockBoundNotSetWarned.get(kind).compareAndSet(false, true)) {
+			logger.warn(LocalizableMessage.raw("jdbc: the wait of %s for its lock is not bounded as this backend"
 				+ " means to bound it on this %s database: \"%s\" did not go through, so %s may bound nothing here"
-				+ " and a DDL can wait for a lock another session holds for as long as this engine lets it (%s)",
-				dialect, sql, DDL_LOCK_TIMEOUT_PROPERTY, stackTraceToSingleLineString(e)));
+				+ " and %s can wait for a lock another session holds for as long as this engine lets it (%s)",
+				kind.waiter, dialect, sql, kind.property, kind.waiter, stackTraceToSingleLineString(e)));
 		}
 	}
 
 	/**
-	 * Said once per storage, where the engine behind a connection is not one this backend knows a lock
-	 * setting for. Leaving the bound off such an engine is the conservative reading and stays - untested
-	 * SQL is no thing to send a database on the path a backend opens by - but a deployment that asked
-	 * for the bound has no way of finding out that it got none, which is the silence this ends. It is
-	 * the argument the strict parsing of {@value #DDL_LOCK_TIMEOUT_PROPERTY} is made with, one property
-	 * later.
+	 * Said once per storage and per bound, where the engine behind a connection is not one this backend
+	 * knows a lock setting for. Leaving the bound off such an engine is the conservative reading and
+	 * stays - untested SQL is no thing to send a database on the path a backend opens by - but a
+	 * deployment that asked for the bound has no way of finding out that it got none, which is the
+	 * silence this ends. It is the argument the strict parsing of the property is made with, one
+	 * property later.
+	 * <p>
+	 * Per bound, like every other latch of these bounds: the engine is the same for both, but what an
+	 * operator is told to act on is the property of the wait that was left unbounded, and a backend
+	 * opening its trees on such a driver would otherwise have said the only line there is to say
+	 * before the first write of that backend ever ran.
 	 * <p>
 	 * {@link #dialectOf} reads the engine off the class name of the driver, so this is not the engine
 	 * of an exotic database alone: a mariadb, percona or aurora driver against a live mysql answers
-	 * null here, and that is a session whose {@code lock_wait_timeout} is a year - the very wait this
-	 * bound exists to end. {@link CachedConnection} says the same of a url it knows no connect bound
-	 * for and cannot say it for this one: it keys on the url, which such a driver takes as a mysql one.
+	 * null here, and that is a session whose {@code lock_wait_timeout} is a year and whose
+	 * {@code innodb_lock_wait_timeout} is 50 s - the very waits these bounds exist to end.
+	 * {@link CachedConnection} says the same of a url it knows no connect bound for and cannot say it
+	 * for this one: it keys on the url, which such a driver takes as a mysql one.
 	 * <p>
 	 * The driver is named rather than the url, since the driver is what this reads and what a
 	 * deployment would change - and a url carries the password of the account this backend works as.
 	 * <p>
-	 * The create table this report is issued beside ({@code createCatalogTable()}, and the one
+	 * The create table the DDL report is issued beside ({@code createCatalogTable()}, and the one
 	 * {@code openTree()} issues for a tree of its own) is not itself a wait such an engine leaves
 	 * unbounded: {@code getTableDialect()} answers an unrecognised engine with postgres' column
 	 * types, which a mysql-family server rejects as a syntax error before it ever queues for a lock.
 	 * What this warns of - the wait an operator would otherwise meet with no word on why - is the
-	 * create index of an open whose table is already there, and the drop table of a clear.
+	 * create index of an open whose table is already there, the drop table of a clear, and every row
+	 * lock a write of that backend queues for.
 	 */
-	private void reportTheEngineIsNotKnown(Connection con) {
-		if (ddlLockBoundEngineUnknownWarned.compareAndSet(false, true)) {
-			final String said=String.format("jdbc: the wait of a DDL for a lock is left unbounded on this"
+	private void reportTheEngineIsNotKnown(Connection con, LockBound kind) {
+		if (lockBoundEngineUnknownWarned.get(kind).compareAndSet(false, true)) {
+			final String said=String.format("jdbc: the wait of %s for a lock is left unbounded on this"
 				+ " database: %s is not a driver this backend knows a lock setting of an engine for, so %s"
-				+ " bounds nothing here and a DDL - the create index of an open whose table is already there,"
-				+ " the drop table of a clear - waits for a lock another session holds for as long as this"
-				+ " engine lets it", driverNameOf(con), DDL_LOCK_TIMEOUT_PROPERTY);
-			ddlLockBoundEngineUnknownSaid=said; // read back by a case the way it reads the flag beside it
+				+ " bounds nothing here and %s waits for a lock another session holds for as long as this"
+				+ " engine lets it", kind.waiter, driverNameOf(con), kind.property, kind.waiterInDetail);
+			// read back by a case the way it reads the flag beside it
+			lockBoundEngineUnknownSaid.get(kind).set(said);
 			logger.warn(LocalizableMessage.raw(said));
 		}
 	}
 
 	/**
-	 * Gives the session back the value it carried. Best effort, and never the outcome of the DDL: this
-	 * runs from a {@code finally} while the caller may be being unwound, where a throw would replace
-	 * the failure that brought it there (JLS 14.20.2) - the very one saying what went wrong.
+	 * Gives the session back the value it carried. Best effort, and never the outcome of the work this
+	 * bound was put around: this runs from a {@code finally} while the caller may be being unwound,
+	 * where a throw would replace the failure that brought it there (JLS 14.20.2) - the very one saying
+	 * what went wrong.
 	 * <p>
 	 * A connection this failed on is handed on to nobody. A pooled one is closed rather than given
 	 * back, and the catalog's own - the one connection reaching this that was never in the pool -
@@ -2435,10 +2994,11 @@
 	 * that, and is closed with that write. Leaving it to the next borrow to
 	 * notice does not work: that validation is {@code con.isValid()}, a liveness check which a
 	 * connection whose reset failed for a transient reason passes while still carrying our bound, and
-	 * on sql server it would then cut every lock wait of that borrower at it - row locks included,
-	 * which {@link #isConflict} classifies as no replayable conflict, so {@code write()} does not
-	 * replay them and a client sees a hard failure. That is the hazard this bound is scoped to a DDL to
-	 * avoid, arriving through the back door. Kept out of the pool, its blast radius is this one
+	 * on sql server it would then cut every lock wait of that borrower at it - row locks included. A
+	 * write of that borrower would replay them ({@link #replayReason} reads the bound this attempt was
+	 * armed with, not the one left on the session), but a read replays nothing at all, so a client of
+	 * one sees a hard failure. That is the hazard each of these bounds is scoped to its own piece of
+	 * work to avoid, arriving through the back door. Kept out of the pool, its blast radius is this one
 	 * connection instead of the rest of its life.
 	 * <p>
 	 * The value is named as the statement that set it rather than read back off the property at log
@@ -2446,25 +3006,26 @@
 	 * put on it - which is not the configured figure either, where the session's own value was the
 	 * tighter one.
 	 */
-	private void restoreDdlLockBound(Connection con, Dialect dialect, String bound, String restore) {
+	private void restoreLockBound(Connection con, Dialect dialect, ArmedLockBound armed) {
 		try {
 			boundedSessionCall(con, () -> {
-				executeSessionStatement(con, restore);
+				executeSessionStatement(con, armed.restore);
 				return null;
 			});
 		}catch (SQLException | RuntimeException e) {
 			if (con instanceof CachedConnection) {
 				((CachedConnection) con).keepOutOfThePool();
 			}
+			final AtomicLong warnedAt=lockBoundLeftBehindWarned.get(armed.kind);
 			final long now=System.currentTimeMillis();
-			final long last=ddlLockBoundLeftBehindWarned.get();
-			if (now-last >= DDL_LOCK_BOUND_WARNING_INTERVAL_MS && ddlLockBoundLeftBehindWarned.compareAndSet(last, now)) {
-				logger.warn(LocalizableMessage.raw("jdbc: the lock bound of a DDL could not be taken off a connection"
+			final long last=warnedAt.get();
+			if (now-last >= LOCK_BOUND_WARNING_INTERVAL_MS && warnedAt.compareAndSet(last, now)) {
+				logger.warn(LocalizableMessage.raw("jdbc: the lock bound of %s could not be taken off a connection"
 					+ " of this %s database, which may have been left carrying \"%s\" instead of the value it had:"
 					+ " that connection is not handed on - a pooled one is closed rather than given back, and the"
 					+ " catalog's own carries it for the rest of the write that opened it and is closed with that"
 					+ " write - so no later write gives up on a lock at a bound of %s it never asked for (%s)",
-					dialect, bound, DDL_LOCK_TIMEOUT_PROPERTY,
+					armed.kind.waiter, dialect, armed.bound, armed.kind.property,
 					stackTraceToSingleLineString(e)));
 			}
 		}
@@ -3939,13 +4500,18 @@
 	 * better than never returning. It is bounded twice - by {@link #MAX_RETRIES} attempts and by the
 	 * {@link #RETRY_WINDOW_NANOS} wall-clock window - because an attempt is not guaranteed to be short: a conflict
 	 * an engine reports only after its own lock wait timeout would otherwise multiply that wait by the attempt
-	 * count. The window alone is not enough either, in the other direction: it is shorter than the wait that
-	 * precedes a conflict the engine reports promptly, so measured against such a conflict it does not bound that
-	 * wait but only leaves the operation with no replay at all, which is what master did with the deadlock of
-	 * issue #903. One replay is therefore granted to a prompt conflict whatever the clock says; see
-	 * {@link #grantedPastTheWindow}. The window is checked between attempts, so an attempt already running is
-	 * never interrupted: a conflicted operation holds its caller for the window plus one attempt, and a prompt
-	 * conflict for two attempts when that is longer.
+	 * count. The window is checked between attempts, so an attempt already running is never interrupted: a
+	 * conflicted operation holds its caller for the window plus one attempt.
+	 * <p>
+	 * What makes those two bounds enough is that the attempt itself is bounded: the transaction runs under
+	 * {@link #ROW_LOCK_TIMEOUT_PROPERTY}, armed on the session here and taken off again before the connection is
+	 * released, so the wait ahead of a conflict ends inside the window rather than outlasting it (#915). Where
+	 * nothing bounds the attempt - oracle, which has no such setting, and a deployment that turns the property off
+	 * - the window is shorter than the wait that precedes a conflict the engine reports promptly, so measured
+	 * against such a conflict it does not bound that wait but only leaves the operation with no replay at all,
+	 * which is what master did with the deadlock of issue #903. One replay is therefore granted to a prompt
+	 * conflict whatever the clock says; see {@link #grantedPastTheWindow}. Such a conflict holds its caller for
+	 * two attempts when that is longer than the window plus one.
 	 * <p>
 	 * Only the operation itself is replayed: a failure of {@link #getConnection()} or of the implicit
 	 * {@link Connection#close()} - which returns the connection to the pool after a rollback - leaves the loop, so
@@ -3956,7 +4522,7 @@
 	 * connection handed out unvalidated and found dead costs an attempt rather than the operation, and a write of
 	 * the replication replay - which records a failed operation as applied and advances the server state past it,
 	 * see #889 - never sees it. Only while nothing of the attempt may have been committed yet, though: see
-	 * {@link #replayReason(Conflict, Throwable, boolean, boolean, boolean)}.
+	 * {@link #replayReason}.
 	 */
 	@Override
 	public void write(WriteOperation writeOperation) throws Exception {
@@ -3964,6 +4530,10 @@
 		for (int attempt=1;;attempt++) {
 			Exception failure=null;
 			String driver=null;
+			Dialect dialect=null;
+			//what bounds the wait of this attempt for a row lock, and whether anything does: the window
+			//below is a clock, and a clock cannot bound a wait nothing else bounds (#903, #915)
+			ArmedLockBound rowLock=ArmedLockBound.none(LockBound.ROW);
 			boolean committing=false;
 			boolean dropped=false;
 			boolean partlyCommitted=false;
@@ -3972,6 +4542,7 @@
 			final Connection con=getConnection();
 			try (con) {
 				driver=driverNameOf(con);
+				dialect=dialectOf(con);
 				final WriteableTransactionTransactionImpl txn=new WriteableTransactionTransactionImpl(con);
 				//the connect of the catalog is made inside this attempt and retries the way a borrow does,
 				//up to the pool timeout - six times this window at the defaults. Left to its own deadline
@@ -3982,6 +4553,17 @@
 				//each of them. The grant of #903 is deliberately not passed on - it buys one more replay
 				//of the operation, not one more connect of the catalog inside it
 				txn.catalogSession.boundedAlsoBy(startedAt+RETRY_WINDOW_NANOS);
+				//armed against the try that takes it off again, and against nothing else: the release below
+				//runs from that finally, before this connection goes back to the pool, and a bound left on a
+				//pooled session would cut the lock waits of whoever borrows it next - a read among them, which
+				//has no replay to absorb the failure that is. It arms nothing where it cannot (armLockBound),
+				//so the operation runs whatever the session made of it.
+				//Handed to the transaction as well as kept here: a DDL of this attempt commits, which ends the
+				//transaction this bound was scoped to, and the transaction is what takes it off there. This
+				//copy is what the attempt ran under, which is what the replay is decided on; txn.rowLock is
+				//what is still on the session, which is what the finally below has to give back
+				rowLock=armLockBound(con, dialect, LockBound.ROW);
+				txn.rowLock=rowLock;
 				try {
 					writeOperation.run(txn);
 					committing=true;
@@ -4034,6 +4616,30 @@
 									stackTraceToSingleLineString(e)));
 						}
 					}
+					//after the rollback above rather than beside it: on the two engines whose bound is a session
+					//setting a rollback leaves that setting exactly where it was, so this is what takes it off -
+					//and it runs while this attempt still holds the connection, before the release hands it on.
+					//It reports what it cannot do and throws nothing, the way every other statement of this
+					//finally does: a value that could not be given back costs the pool this connection, not the
+					//write its failure. Asked of the transaction rather than of the local above: a DDL of this
+					//attempt takes the bound off where it commits, and what is left to give back is what the
+					//transaction still carries.
+					//Not on a connection the driver reports closed: there is nothing there to give a value back
+					//to, and the setting would fail on a dead session and be reported as a bound left behind -
+					//so a restart under write load would read as a stream of stranded bounds for connections
+					//that are simply gone. Decided on that and not on `dropped`: this attempt is classified as
+					//a dropped connection for a class 08 anywhere in the chains of its failure, and a refused
+					//catalog connect puts one there with this connection untouched (newCatalogConnection: mysql
+					//answers its connection limit with 08004, sql server with 08S01). That connection goes back
+					//to the pool - the rollback above went through, the pool's validation is isValid(), and a
+					//live session passes it - so on the two engines whose bound is a session setting, skipping
+					//the restore here hands the next borrow this backend's own bound, a read among them, which
+					//has no replay to absorb the error it would then give up with.
+					//A dead session the driver has not flagged yet takes the other road, which is safe: the
+					//restore fails, it is reported, and the connection is kept out of the pool.
+					if (!isClosed(con)) {
+						releaseLockBound(con, dialect, txn.rowLock);
+					}
 				}
 			} catch (Exception e) {
 				//anything the operation did not throw comes from around it - the name of the driver, the
@@ -4069,7 +4675,7 @@
 			//is also the path most likely to carry deeply wrapped chains, since RootContainer.open() commits
 			//DDL and raises the flag for the rest of the write
 			final ConflictVerdict verdict=partlyCommitted ? NOT_CLASSIFIED : conflictVerdict(failure,driver);
-			final String reason=replayReason(verdict.conflict,failure,committing,partlyCommitted,dropped);
+			final String reason=replayReason(verdict.conflict,failure,committing,partlyCommitted,dropped,dialect,rowLock);
 			//nanoTime()-startedAt is the overflow safe form of the elapsed time
 			final long elapsedNanos=nanoTime()-startedAt;
 			if (reason==null || !replayableWithin(attempt, elapsedNanos, verdict.conflict)) {
@@ -4131,20 +4737,60 @@
 	 * OpenDJ issue #896 - which masked the failure that caused the replay and left the indexes of the previous
 	 * attempt behind with their configuration listeners.
 	 *
+	 * A wait for a row lock that this backend bounded is replayable as well, and only where this backend bounded
+	 * it. The engine ends such a wait with a failure of its own - 55P03 on postgres, error 1222 on sql server,
+	 * both of which {@link #isConflict} matches as no conflict, and the class 40 of a mysql lock wait timeout,
+	 * which it already does - and it ends it having rolled nothing back, unlike a deadlock: the transaction is
+	 * still open and the blocker is still holding the lock. What makes it replayable all the same is the
+	 * rollback {@link #write} issues before it asks, and what makes it worth replaying is that the wait it took
+	 * fits inside the replay window rather than outlasting it, which is what {@link ArmedLockBound#bounded}
+	 * says. Where nothing bounded that wait - {@link #ROW_LOCK_TIMEOUT_PROPERTY} at 0, oracle, a session that
+	 * would not take the setting - the failure is left exactly as unreplayable as it was before #915: a bound
+	 * an operator set for themselves is not a licence for this loop to take that wait again, and a wait nothing
+	 * bounds is the one thing the window cannot govern. It is deliberately no class of {@link Conflict}: the
+	 * verdict of such a failure stays {@link Conflict#NONE}, so it is replayed on the window alone and never
+	 * granted past it by {@link #grantedPastTheWindow} - a grant that exists for the waits nothing bounds,
+	 * which this one is not.
+	 * <p>
+	 * The wait a DDL of the same attempt took for its own lock reaches this branch the same way, an engine having
+	 * one way of saying a lock was not available: a {@code create index} that gave up at
+	 * {@link #DDL_LOCK_TIMEOUT_PROPERTY} is replayed rather than failing at once, on the terms of every other
+	 * replay here. Both properties default well under the window, so an attempt of the defaults spends a fraction
+	 * of it either way; a deployment that raises one of them past the window buys the trailing attempt this loop
+	 * has always allowed - what it cannot buy is an attempt nothing ends at all.
+	 *
 	 * @param conflict the class {@link #conflictVerdict} read from the failure, asked of it once by the caller
 	 * @param committing whether the failure was reported by {@code commit()}, which leaves the outcome unknown
 	 * @param partlyCommitted whether the attempt committed part of its work before it failed
 	 * @param connectionClosed whether the driver closed the connection under the failure - evidence no SQLState
 	 * carries on mssql-jdbc, which reports a killed session as S0001 and closes the connection behind it
+	 * @param dialect the engine, which is what names its own way of saying a lock was not available
+	 * @param rowLock the row lock bound this attempt ran under, as {@link #armLockBound} left it
 	 */
 	static String replayReason(Conflict conflict, Throwable failure, boolean committing, boolean partlyCommitted,
-			boolean connectionClosed) {
+			boolean connectionClosed, Dialect dialect, ArmedLockBound rowLock) {
 		if (partlyCommitted) {
 			return null;
 		}
 		if (conflict!=Conflict.NONE) {
 			return "a conflict";
 		}
+		// after the conflict above, which is the class a mysql lock wait timeout arrives in: both are
+		// replayed, and the reason a line names should be the strongest thing that can be said of the
+		// failure. Not while committing, for the reason a drop is not replayed there: the outcome of a
+		// commit that did not answer is unknown, and this loop must not apply a write twice.
+		// Read off every link of the chain, so a lock timeout of the catalog session - a connection of
+		// its own, carrying no bound of ours (enrolInCatalog) and giving up only where a deployment set
+		// a lock_timeout for itself - is replayed under this reason as well. That reads wider than what
+		// the bound of this attempt can say, and stays: the catalog session is made and closed inside
+		// the attempt, its statement was rolled back with the rest, and what such a replay costs is
+		// bounded by the window like any other - a wait long enough to outlast it is not replayed at all.
+		// Which is why the line names the attempt rather than the row lock: a DDL of this attempt that
+		// gave up at DDL_LOCK_TIMEOUT_PROPERTY and a lock timeout of its catalog session are both
+		// replayed here, and "a row lock" is true of neither
+		if (rowLock.bounded && !committing && lockNotAvailable(failure, dialect)) {
+			return "a lock wait of an attempt this backend bounded";
+		}
 		if (!committing && (connectionClosed || isConnectionFailure(failure))) {
 			return "a connection the database dropped";
 		}
@@ -4313,8 +4959,11 @@
 	 * <p>
 	 * Declared in order of how much they restrict the replay, which is the order {@link #conflictVerdict}
 	 * compares them in: the strongest class any link of a failure carries is the class of that failure. Only
-	 * {@link #PROMPT} is granted the replay past the window, so every class this list gains - #915 adds one -
-	 * has to be placed against that grant rather than merely appended.
+	 * {@link #PROMPT} is granted the replay past the window, so every class this list gains has to be placed
+	 * against that grant rather than merely appended. The wait for a row lock that #915 bounds is deliberately
+	 * not one of them: its failure is no conflict at all - see {@link #replayReason}, which replays it on the
+	 * strength of the bound - and a class here would put it in reach of a grant meant for the waits nothing
+	 * bounds.
 	 */
 	enum Conflict {
 		/** Not a conflict: no replay resolves it. */
@@ -4426,7 +5075,7 @@
 	 * The strongest class a conflict raised under the given engine can carry, which is where
 	 * {@link #conflictVerdict} stops walking: nothing further along the chains can outrank it. It is the maximum
 	 * of what {@link #classOf} returns for that dialect and has to be read together with it - a property of the
-	 * engine rather than the last constant of {@link Conflict}, so that the class #915 adds cannot silently move
+	 * engine rather than the last constant of {@link Conflict}, so that a class added later cannot silently move
 	 * the stop condition, and so that the walk of the three engines reporting no lock wait timeout of their own
 	 * ends on the first conflict it meets rather than at the end of every chain.
 	 */
@@ -4455,8 +5104,9 @@
 	 * {@code openTree(createOnDemand)} is guarded by a catalog read - and its writes go down the ANSI branch of
 	 * {@code upsert}, which is an {@code update} and an {@code insert}, not a statement a MySQL-wire engine
 	 * refuses. The cost of the class is one replay of the window's own length for an engine whose conflicts are
-	 * in fact prompt, which is the direction worth being wrong in; #915 removes the trade by bounding the
-	 * attempt itself.
+	 * in fact prompt, which is the direction worth being wrong in. The bound #915 puts on the attempt itself
+	 * does not remove the trade here: it is armed through the {@link Dialect} the driver name resolves to, so
+	 * an unrecognised driver is exactly where it arms nothing.
 	 */
 	private static Conflict classOf(SQLException e, Dialect dialect) {
 		if (!isConflict(e, dialect)) {
@@ -4493,12 +5143,12 @@
 	/**
 	 * Whether this replay is the one {@link #RETRY_WINDOW_NANOS} does not get to deny: the first replay of a
 	 * conflict its engine reports promptly, taken although the window is already spent. The wait an engine spends
-	 * before reporting such a conflict is charged to the attempt that hit it and is unbounded on three of the four
-	 * engines here - SQL Server took some 12 s to pick a victim in CI - so there is no window that some wait does
-	 * not outlast, and measuring one against it only leaves the operation with no replay at all, which is issue
-	 * #903. The grant does not extend to {@link Conflict#AFTER_LOCK_WAIT}, whose wait the engine has already
-	 * bounded for us: replaying that costs the same bounded wait again, which is exactly what the window is here
-	 * to refuse. Nor to {@link Conflict#UNKNOWN_ENGINE}, of which the same cannot be ruled out.
+	 * before reporting such a conflict is charged to the attempt that hit it, and where nothing bounds that wait -
+	 * SQL Server took some 12 s to pick a victim in CI, before #915 bounded it - there is no window that some
+	 * wait does not outlast, and measuring one against it only leaves the operation with no replay at all, which
+	 * is issue #903. The grant does not extend to {@link Conflict#AFTER_LOCK_WAIT}, whose wait the engine has
+	 * already bounded for us: replaying that costs the same bounded wait again, which is exactly what the window
+	 * is here to refuse. Nor to {@link Conflict#UNKNOWN_ENGINE}, of which the same cannot be ruled out.
 	 * <p>
 	 * Asked as a question of its own so that the line reporting the replay can name the bound that was actually
 	 * applied instead of inferring it from the clock: {@code elapsed >= window} coincides with this grant only
@@ -4507,8 +5157,14 @@
 	 * <p>
 	 * {@code attempt==1} is a proxy and not the invariant: the invariant is that no clock can bound a wait
 	 * nothing else bounds, and that holds on every attempt, not only the first. Widening the grant to all of them
-	 * would leave {@link #MAX_RETRIES} as the only real cap, so it is held to one replay until the attempt itself
-	 * carries a lock bound - see #915, which retires this method rather than widening it.
+	 * would leave {@link #MAX_RETRIES} as the only real cap, so it is held to one replay.
+	 * <p>
+	 * What is left for it to do, now that {@link #ROW_LOCK_TIMEOUT_PROPERTY} bounds the attempt, is the attempts
+	 * that bound does not reach: oracle, which has no setting for the wait, the property at 0, and a session that
+	 * would not take the setting - there the wait is as unbounded as it was when this grant was written. Where
+	 * the bound is armed an attempt gives up on its lock long before this window is spent, so the grant is not
+	 * asked whether the bound was on: a prompt conflict that still reaches here past the window came out of a
+	 * long transaction rather than a long wait, and the one replay it is granted runs under the same bound.
 	 */
 	static boolean grantedPastTheWindow(int attempt, long elapsedNanos, Conflict conflict) {
 		return attempt==1 && conflict==Conflict.PROMPT && elapsedNanos>=RETRY_WINDOW_NANOS;
@@ -4761,6 +5417,14 @@
 		 */
 		boolean partlyCommitted;
 
+		/**
+		 * The row lock bound still on the session of this transaction, as {@link JDBCStorage#write} armed
+		 * it and {@link #commitStatement} may have taken it off again. {@code none} while nothing of ours
+		 * is on: a transaction of the importer, one on an engine or a setting that took no bound, and
+		 * every attempt after its own DDL committed.
+		 */
+		ArmedLockBound rowLock=ArmedLockBound.none(LockBound.ROW);
+
 		public WriteableTransactionTransactionImpl(Connection con) {
 			this(con, StatementBound.OPERATION);
 		}
@@ -4792,6 +5456,15 @@
 		 * raised in front of it would take a conflict the engine itself undid out of the replay. On those the
 		 * flag goes up in front of the commit instead, which is the call that leaves the outcome of the
 		 * transaction unknown when it fails.
+		 * <p>
+		 * The row lock bound of the attempt ends here as well, where the statement is a DDL: it was armed
+		 * around a transaction, and this commit is the end of that transaction. On postgres the
+		 * {@code set local} goes with the commit by itself, and on the two engines whose setting is the
+		 * session's this is what takes it off - the attempt behind it has committed part of its work, so
+		 * it is out of the replay whatever it waits for, and a bound on a wait nothing can replay is a
+		 * write that fails where it used to go through. In front of the DDL rather than behind it where
+		 * the DDL bound is turned off, which is the one case the DDL would otherwise run under the row
+		 * bound of this write - see {@link #liftTheRowLockBoundForAnUnboundedDdl}.
 		 *
 		 * @param ddl whether the statement is a DDL one, which two of the four engines commit before
 		 */
@@ -4810,15 +5483,76 @@
 				return null;
 			};
 			if (ddl) {
-				// The DDL of this backend is the part of it that takes locks, and the part that waits for
-				// one with no bound of its own. The delete from of clearTree() waits for row locks, which
-				// write() replays a conflict of and which this bound has no business ending.
-				withDdlLockBound(con, dialectOf(con), issue);
+				try {
+					liftTheRowLockBoundForAnUnboundedDdl();
+					// The DDL of this backend is the part of it that takes locks, and the part that waits for
+					// one with no bound of its own. The delete from of clearTree() waits for row locks, which
+					// write() replays a conflict of and which this bound has no business ending.
+					withDdlLockBound(con, dialectOf(con), issue);
+				}finally {
+					// From a finally, since the DDL that failed ended the transaction just as surely on the two
+					// engines that commit before one - and a bound left on for the rest of a write that is out
+					// of the replay is exactly what this takes off. Best effort, like every other release of
+					// one: it says what it could not do and throws nothing over the failure of the DDL
+					releaseTheRowLockBound();
+				}
 			}else {
 				issue.run();
 			}
 		}
 
+		/**
+		 * Takes the row lock bound of this attempt off in front of a DDL that is to wait as this backend
+		 * waited before {@link JDBCStorage#DDL_LOCK_TIMEOUT_PROPERTY} existed, which is what that
+		 * property at 0 asks for.
+		 * <p>
+		 * At 0 nothing of ours is armed around the DDL ({@link JDBCStorage#armLockBound}), and on the two
+		 * engines with one setting for both waits what it then runs under is not the wait it had before
+		 * either bound existed but the row bound of this very write - 3 s at the defaults, issued one
+		 * statement earlier: {@code set lock_timeout 3000} on the sql server session,
+		 * {@code set local lock_timeout = 3000} on the postgres transaction, which bounds every lock wait
+		 * of it. A {@code create table} of an open blocked by another session then gives up at 3 s, and
+		 * gives up as the bare vendor error - nothing of ours being armed, there is no bound to name it
+		 * by - which {@link JDBCStorage#replayReason} replays as a lock wait of a bounded attempt,
+		 * re-issuing the DDL until the window is spent. So the bound comes off in front of that DDL
+		 * rather than behind it: on postgres the {@code set local} is reset, having no value to give
+		 * back, and on sql server the release puts the deployment's own {@code LOCK_TIMEOUT} back ahead
+		 * of the statement.
+		 * <p>
+		 * Only where this backend armed anything at all: a session already tighter than the bound, an
+		 * engine that takes no such setting, a setting the session refused and mysql - whose metadata
+		 * lock is a variable of its own - leave the DDL exactly where it was. And it throws rather than
+		 * reporting, unlike the release: a {@code set local} that fails has aborted the transaction
+		 * anyway, and the failure saying why is worth more to an operator than the 25P02 of the DDL
+		 * behind it.
+		 */
+		private void liftTheRowLockBoundForAnUnboundedDdl() throws SQLException {
+			final Dialect dialect=dialectOf(con);
+			if (LockBound.DDL.seconds()>0 || rowLock.bound==null || dialect==null
+					|| !dialect.oneSettingForBothBounds()) {
+				return;
+			}
+			final String lift=dialect.rowLockLiftSql();
+			if (lift!=null) {
+				boundedSessionCall(con, () -> {
+					executeSessionStatement(con, lift);
+					return null;
+				});
+			}
+			releaseTheRowLockBound();
+		}
+
+		/**
+		 * Gives the session back what the row lock bound of this attempt displaced, and records that the
+		 * transaction carries none - so the finally of {@link JDBCStorage#write} has nothing left to give
+		 * back and nothing is put back twice.
+		 */
+		private void releaseTheRowLockBound() {
+			final ArmedLockBound armed=rowLock;
+			rowLock=ArmedLockBound.none(LockBound.ROW);
+			releaseLockBound(con, dialectOf(con), armed);
+		}
+
 		/** Whether this engine commits the transaction before a DDL statement whether asked to or not. */
 		private boolean commitsBeforeDdl() {
 			final Dialect dialect=dialectOf(con);
diff --git a/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCDdlLockBoundTestCase.java b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCDdlLockBoundTestCase.java
index d8875c3..3210fd1 100644
--- a/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCDdlLockBoundTestCase.java
+++ b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCDdlLockBoundTestCase.java
@@ -18,6 +18,7 @@
 import org.forgerock.opendj.server.config.server.JDBCBackendCfg;
 import org.opends.server.DirectoryServerTestCase;
 import org.opends.server.backends.jdbc.JDBCStorage.Dialect;
+import org.opends.server.backends.jdbc.JDBCStorage.LockBound;
 import org.opends.server.backends.jdbc.JDBCStorage.StatementBound;
 import org.opends.server.backends.pluggable.spi.AccessMode;
 import org.opends.server.backends.pluggable.spi.Importer;
@@ -59,6 +60,7 @@
 import static org.mockito.Mockito.when;
 import static org.testng.Assert.assertEquals;
 import static org.testng.Assert.assertFalse;
+import static org.testng.Assert.assertNotNull;
 import static org.testng.Assert.assertNull;
 import static org.testng.Assert.assertSame;
 import static org.testng.Assert.assertTrue;
@@ -216,9 +218,9 @@
 
 		storage.withDdlLockBound(con, null, theDdl());
 
-		assertTrue(storage.ddlLockBoundEngineUnknownWarned.get(),
+		assertTrue(storage.lockBoundEngineUnknownWarned.get(LockBound.DDL).get(),
 			"a driver this backend knows no lock bound for left the wait of every DDL unbounded and unsaid");
-		final String said = storage.ddlLockBoundEngineUnknownSaid;
+		final String said = storage.lockBoundEngineUnknownSaid.get(LockBound.DDL).get();
 		assertTrue(said != null && said.contains(JDBCStorage.driverNameOf(con)),
 			"the driver left unbounded was not named: " + said);
 		assertTrue(said != null && said.contains(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY),
@@ -233,12 +235,35 @@
 	public void testAnEngineThisBackendDoesNotKnowIsReportedOnce() throws Exception {
 		final Connection con = recording(mock(Connection.class), "0");
 		storage.withDdlLockBound(con, null, theDdl());
-		assertTrue(storage.ddlLockBoundEngineUnknownWarned.get(), "the first DDL on the engine said nothing");
+		assertTrue(storage.lockBoundEngineUnknownWarned.get(LockBound.DDL).get(),
+			"the first DDL on the engine said nothing");
 
-		storage.ddlLockBoundEngineUnknownSaid = null;
+		storage.lockBoundEngineUnknownSaid.get(LockBound.DDL).set(null);
 		storage.withDdlLockBound(con, null, theDdl());
 
-		assertNull(storage.ddlLockBoundEngineUnknownSaid, "an engine this backend does not know was reported twice");
+		assertNull(storage.lockBoundEngineUnknownSaid.get(LockBound.DDL).get(),
+			"an engine this backend does not know was reported twice");
+	}
+
+	/**
+	 * Once per bound, though: the row lock bound of a write is armed by other code on another path,
+	 * and the line an operator acts on names the property of the wait that was left unbounded. An
+	 * open reporting the DDL bound of this engine says nothing about the writes behind it, and a
+	 * single latch would leave every one of them silent.
+	 */
+	@Test
+	public void testTheRowBoundOfAnEngineThisBackendDoesNotKnowIsReportedOnItsOwn() throws Exception {
+		final Connection con = recording(mock(Connection.class), "0");
+		storage.withDdlLockBound(con, null, theDdl());
+
+		storage.armLockBound(con, null, LockBound.ROW);
+
+		final String said = storage.lockBoundEngineUnknownSaid.get(LockBound.ROW).get();
+		assertNotNull(said, "the open of a backend silenced the row lock bound of every write behind it");
+		assertTrue(said.contains(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY),
+			"the property that would have bounded the write was not named: " + said);
+		assertFalse(said.contains(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY),
+			"the write was sent to the property of the DDL bound: " + said);
 	}
 
 	/** A deployment that turned the bound off asked for none anywhere, and has nothing to act on. */
@@ -248,7 +273,7 @@
 
 		storage.withDdlLockBound(recording(mock(Connection.class), "0"), null, theDdl());
 
-		assertFalse(storage.ddlLockBoundEngineUnknownWarned.get(),
+		assertFalse(storage.lockBoundEngineUnknownWarned.get(LockBound.DDL).get(),
 			"a bound nobody asked for was reported as an engine this backend does not know");
 	}
 
@@ -260,7 +285,7 @@
 	public void testOracleIsNotReportedAsAnEngineThisBackendDoesNotKnow() throws Exception {
 		storage.withDdlLockBound(recording(mock(Connection.class), "0"), Dialect.ORACLE, theDdl());
 
-		assertFalse(storage.ddlLockBoundEngineUnknownWarned.get(),
+		assertFalse(storage.lockBoundEngineUnknownWarned.get(LockBound.DDL).get(),
 			"the engine left alone deliberately was reported as one this backend cannot bound");
 	}
 
@@ -447,8 +472,13 @@
 	}
 
 	/**
-	 * And a sql server lock wait is left as unreplayable as it was: error 1222 is no conflict of
-	 * {@code isConflict()}, and a DDL made to look like one would be replayed into the same wait.
+	 * And the rename leaves a sql server lock wait no conflict of {@code isConflict()}: error 1222 is
+	 * none, and a DDL made to look like one would be replayed on that class - past the window, where the
+	 * class is granted one - rather than on what bounded the wait. No conflict is not the same as never
+	 * replayed: a DDL issued from a write runs inside an attempt carrying a row lock bound of this
+	 * backend, and a lock wait of such an attempt is replayed on the window like any other (see
+	 * {@code JDBCStorageRetryTest.testADdlLockWaitInsideAWriteIsReplayedOnTheBoundTheAttemptRanUnder}) -
+	 * on the window alone, and never past it.
 	 */
 	@Test
 	public void testARewrittenSqlServerLockWaitIsMadeNoMoreReplayable() throws Exception {
@@ -514,10 +544,13 @@
 
 		bounded.write(txn -> txn.deleteTree(TREE));
 
-		// the search path in front of them is the lookup that decides whether there is a table to drop
-		// at all, narrowed to the schemas an unqualified name of this connection resolves in (#888): it
-		// reads a session setting rather than the data, and takes a bound of its own
-		assertEquals(issued, asList("select unnest(current_schemas(true))",
+		// the first setting is the row lock bound of the write this drop runs inside (#915), armed once
+		// per attempt and discarded with the transaction; the search path after it is the lookup that
+		// decides whether there is a table to drop at all, narrowed to the schemas an unqualified name of
+		// this connection resolves in (#888): it reads a session setting rather than the data, and takes a
+		// bound of its own. The DDL bound is the third, and is the one this case is about
+		assertEquals(issued, asList("set local lock_timeout = 3000",
+			"select unnest(current_schemas(true))",
 			"set local lock_timeout = 5000",
 			"drop table " + JDBCStorage.toTableName(TREE)));
 	}
@@ -893,7 +926,11 @@
 		assertEquals(issued, singletonList(THE_DDL));
 	}
 
-	/** And a setting that went through is left standing: it is what bounds the DDL that follows it. */
+	/**
+	 * And a setting that went through is left standing: it is what bounds the DDL that follows it. The
+	 * point it was taken in front of is let go of instead - a {@code SET LOCAL} survives the release,
+	 * and a subtransaction left open would span the work this bound was put around.
+	 */
 	@Test
 	public void testATransactionWhoseSettingWentThroughIsNotTakenBack() throws Exception {
 		final Connection con = recording(mock(Connection.class), "0");
@@ -903,6 +940,7 @@
 		storage.withDdlLockBound(con, Dialect.POSTGRES, theDdl());
 
 		verify(con, never()).rollback(beforeTheBound);
+		verify(con).releaseSavepoint(beforeTheBound);
 		assertEquals(issued, asList("set local lock_timeout = 5000", THE_DDL));
 	}
 
diff --git a/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCRowLockBoundTestCase.java b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCRowLockBoundTestCase.java
new file mode 100644
index 0000000..1be9b8d
--- /dev/null
+++ b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCRowLockBoundTestCase.java
@@ -0,0 +1,803 @@
+/*
+ * The contents of this file are subject to the terms of the Common Development and
+ * Distribution License (the License). You may not use this file except in compliance with the
+ * License.
+ *
+ * You can obtain a copy of the License at legal/CDDLv1.0.txt. See the License for the
+ * specific language governing permission and limitations under the License.
+ *
+ * When distributing Covered Software, include this CDDL Header Notice in each file and include
+ * the License file at legal/CDDLv1.0.txt. If applicable, add the following below the CDDL
+ * Header, with the fields enclosed by brackets [] replaced by your own identifying
+ * information: "Portions copyright [year] [name of copyright owner]".
+ *
+ * Copyright 2026 3A Systems, LLC.
+ */
+package org.opends.server.backends.jdbc;
+
+import org.forgerock.opendj.server.config.server.JDBCBackendCfg;
+import org.opends.server.DirectoryServerTestCase;
+import org.opends.server.backends.jdbc.JDBCStorage.ArmedLockBound;
+import org.opends.server.backends.jdbc.JDBCStorage.Dialect;
+import org.opends.server.backends.jdbc.JDBCStorage.LockBound;
+import org.opends.server.backends.pluggable.spi.StorageRuntimeException;
+import org.testng.annotations.AfterMethod;
+import org.testng.annotations.BeforeMethod;
+import org.testng.annotations.DataProvider;
+import org.testng.annotations.Test;
+
+import java.sql.Connection;
+import java.sql.ResultSet;
+import java.sql.SQLException;
+import java.sql.Savepoint;
+import java.sql.Statement;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicReference;
+
+import static java.util.Arrays.asList;
+import static java.util.Collections.emptyList;
+import static java.util.Collections.singletonList;
+import static org.forgerock.opendj.config.ConfigurationMock.mockCfg;
+import static org.mockito.Mockito.anyString;
+import static org.mockito.Mockito.doAnswer;
+import static org.mockito.Mockito.mock;
+import static org.mockito.Mockito.never;
+import static org.mockito.Mockito.verify;
+import static org.mockito.Mockito.when;
+import static org.testng.Assert.assertEquals;
+import static org.testng.Assert.assertFalse;
+import static org.testng.Assert.assertNull;
+import static org.testng.Assert.assertTrue;
+import static org.testng.Assert.fail;
+
+/**
+ * What a transaction of {@code JDBCStorage.write()} is told to do about a row lock another session
+ * holds (#915), and what the replay makes of the failure that ends such a wait.
+ * <p>
+ * It needs no database: the connection is a mock, so what each engine is told - and what it is told
+ * to put back before the connection goes to the next borrower - is pinned wherever the build runs,
+ * while the container suites cover a write really queued behind another session's row lock. The
+ * replay that this bound exists for is driven end to end in {@code JDBCStorageRetryTest}.
+ */
+@SuppressWarnings("javadoc")
+@Test(groups = { "precommit", "jdbc" }, sequential = true)
+public class JDBCRowLockBoundTestCase extends DirectoryServerTestCase {
+
+	/** A driver name of an engine none of the settings fit, which is what the null dialect stands for. */
+	private static final String UNKNOWN_ENGINE = "com.example.jdbc.Connection";
+	private static final String MYSQL = "com.mysql.cj.jdbc.ConnectionImpl";
+
+	/** What the connection of a case was asked to run, in the order it was asked to run it. */
+	private final List<String> issued = new ArrayList<>();
+
+	private JDBCStorage storage;
+
+	@BeforeMethod
+	public void createStorage() {
+		storage = new JDBCStorage(backendCfg(), null);
+		issued.clear();
+	}
+
+	private static JDBCBackendCfg backendCfg() {
+		final JDBCBackendCfg cfg = mockCfg(JDBCBackendCfg.class);
+		when(cfg.getBackendId()).thenReturn("rowLockBound");
+		return cfg;
+	}
+
+	@AfterMethod
+	public void clearProperties() {
+		System.clearProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY);
+		// the cases arming a DDL bound inside this one raise it: the two are decided against each other
+		System.clearProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY);
+	}
+
+	/**
+	 * The window bounding the replays is a clock, and a clock only bounds them while an attempt is
+	 * shorter than it: the default is a third of that window, so a write meeting a lock gets three
+	 * attempts inside it rather than spending the whole window on one and being refused a replay
+	 * (#903). A default at or past the window would put this back exactly as it was.
+	 */
+	@Test
+	public void testTheDefaultLeavesTheReplayWindowRoomForReplays() {
+		assertEquals(JDBCStorage.rowLockBoundSeconds(), 3);
+		assertTrue(JDBCStorage.rowLockBoundSeconds() * 1000L * 1000L * 1000L * 3 <= JDBCStorage.RETRY_WINDOW_NANOS,
+			"the default bound leaves the replay window room for fewer than three attempts");
+	}
+
+	/** Zero is what a deployment that would rather wait for its row lock sets, and so is a negative value. */
+	@Test
+	public void testTheBoundIsTurnedOffByZero() {
+		System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, "0");
+		assertEquals(JDBCStorage.rowLockBoundSeconds(), 0);
+		System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, "-1");
+		assertEquals(JDBCStorage.rowLockBoundSeconds(), 0);
+	}
+
+	/** A value that is not a number keeps the default, so a typo cannot silently unbound the wait. */
+	@Test
+	public void testAValueThatIsNotANumberKeepsTheDefault() {
+		System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, "three seconds");
+		assertEquals(JDBCStorage.rowLockBoundSeconds(), JDBCStorage.ROW_LOCK_TIMEOUT_SECONDS);
+	}
+
+	/** And a value past what a bound of this backend can hold is taken down to it, not read as no bound. */
+	@Test
+	public void testAValueBeyondTheCeilingIsClamped() {
+		System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, String.valueOf(Integer.MAX_VALUE));
+		assertEquals(JDBCStorage.rowLockBoundSeconds(), JDBCStorage.MAX_BOUND_SECONDS);
+	}
+
+	/**
+	 * What each engine is told around the transaction, in the unit its own setting takes: the row lock
+	 * of mysql is innodb_lock_wait_timeout and never the lock_wait_timeout of a metadata lock, sql
+	 * server has one LOCK_TIMEOUT for every lock wait of a session, postgres takes a set local that the
+	 * transaction discards, and oracle has no setting for this wait at all.
+	 */
+	@DataProvider
+	public Object[][] engines() {
+		return new Object[][] {
+			{ "postgres takes a set local and needs no value back", Dialect.POSTGRES, "0",
+				singletonList("set local lock_timeout = 3000"), true },
+			{ "mysql bounds the row lock, not the metadata lock", Dialect.MYSQL, "50",
+				asList("select @@session.innodb_lock_wait_timeout", "set session innodb_lock_wait_timeout=3",
+					"set session innodb_lock_wait_timeout=50"), true },
+			{ "sql server replaces the -1 that waits forever", Dialect.MICROSOFT, "-1",
+				asList("select @@lock_timeout", "set lock_timeout 3000", "set lock_timeout -1"), true },
+			{ "oracle has no session setting for a row lock", Dialect.ORACLE, "0", emptyList(), false },
+			{ "an engine this backend does not know is told nothing", null, "0", emptyList(), false },
+		};
+	}
+
+	@Test(dataProvider = "engines")
+	public void testWhatEachEngineIsToldAroundTheTransaction(String name, Dialect dialect, String carries,
+			List<String> expected, boolean bounded) throws Exception {
+		final Connection con = recording(mock(Connection.class), carries);
+
+		final ArmedLockBound armed = storage.armLockBound(con, dialect, LockBound.ROW);
+		storage.releaseLockBound(con, dialect, armed);
+
+		assertEquals(issued, expected, name);
+		assertEquals(armed.bounded, bounded, name + ": the wait was reported as bounded when it is not, or the other"
+			+ " way round - which is what decides whether the replay may take that wait again");
+	}
+
+	/**
+	 * A session that gives up sooner than this bound keeps exactly what it has - the bound is never
+	 * loosened to ours - and the wait is bounded all the same, which is the whole reason for leaving
+	 * that value alone. The replay reads that, not whether a statement of ours was issued.
+	 */
+	@DataProvider
+	public Object[][] sessionsAlreadyTighter() {
+		return new Object[][] {
+			{ "mysql giving up after a second", Dialect.MYSQL, "1" },
+			{ "sql server told not to wait at all", Dialect.MICROSOFT, "0" },
+		};
+	}
+
+	@Test(dataProvider = "sessionsAlreadyTighter")
+	public void testASessionAlreadyTighterKeepsWhatItHasAndIsStillBounded(String name, Dialect dialect,
+			String carries) throws Exception {
+		final Connection con = recording(mock(Connection.class), carries);
+
+		final ArmedLockBound armed = storage.armLockBound(con, dialect, LockBound.ROW);
+		storage.releaseLockBound(con, dialect, armed);
+
+		assertEquals(issued, singletonList(dialect == Dialect.MYSQL
+			? "select @@session.innodb_lock_wait_timeout" : "select @@lock_timeout"), name);
+		assertTrue(armed.bounded, name + ": a wait the session itself bounds was reported as unbounded");
+	}
+
+	/** While a session looser than this bound is given ours, and gets its own value back afterwards. */
+	@Test
+	public void testASessionLooserThanTheBoundIsGivenOursAndGetsItBack() throws Exception {
+		final Connection con = recording(mock(Connection.class), "5000");
+
+		final ArmedLockBound armed = storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW);
+		storage.releaseLockBound(con, Dialect.MICROSOFT, armed);
+
+		assertEquals(issued, asList("select @@lock_timeout", "set lock_timeout 3000", "set lock_timeout 5000"));
+	}
+
+	/** Turned off, nothing is asked of the session and nothing is claimed about the wait. */
+	@Test
+	public void testTheBoundTurnedOffAsksTheSessionNothing() throws Exception {
+		System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, "0");
+		final Connection con = recording(mock(Connection.class), "50");
+
+		final ArmedLockBound armed = storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW);
+		storage.releaseLockBound(con, Dialect.MYSQL, armed);
+
+		assertEquals(issued, emptyList());
+		assertFalse(armed.bounded, "a wait nothing bounds was reported as bounded");
+	}
+
+	/**
+	 * A session answering the readback with something no SET of it would take back is left alone
+	 * altogether: nothing of ours is ever set where it could not be taken off again.
+	 */
+	@Test
+	public void testASessionThatWillNotSayWhatItCarriesIsLeftAlone() throws Exception {
+		final Connection con = recording(mock(Connection.class), "unlimited");
+
+		final ArmedLockBound armed = storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW);
+		storage.releaseLockBound(con, Dialect.MYSQL, armed);
+
+		assertEquals(issued, singletonList("select @@session.innodb_lock_wait_timeout"));
+		assertFalse(armed.bounded, "a session that never took the bound was reported as bounded");
+	}
+
+	/**
+	 * A setting the session refuses leaves the write running as it ran before this bound existed - the
+	 * bound is an improvement on a wait and never a reason to fail a write that would have gone
+	 * through - and the wait is not claimed to be bounded, so the replay does not take it again.
+	 */
+	@Test
+	public void testASettingTheSessionRefusesLeavesTheWaitUnbounded() throws Exception {
+		final Connection con = recording(mock(Connection.class), "50");
+		final Statement statement = con.createStatement();
+		doAnswer(invocation -> {
+			issued.add((String) invocation.getArguments()[0]);
+			throw new SQLException("this session takes no such setting", "42000");
+		}).when(statement).execute(anyString());
+
+		final ArmedLockBound armed = storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW);
+		storage.releaseLockBound(con, Dialect.MYSQL, armed);
+
+		assertFalse(armed.bounded, "a setting the session refused was reported as bounding the wait");
+		// the value is given back all the same: a setting can reach the server and fail only as the
+		// statement carrying it is closed, and giving back a value the session may never have left
+		// costs a round trip and changes nothing
+		assertEquals(issued, asList("select @@session.innodb_lock_wait_timeout",
+			"set session innodb_lock_wait_timeout=3", "set session innodb_lock_wait_timeout=50"));
+	}
+
+	/**
+	 * A connection left carrying a bound this backend could not take off again does not go back into
+	 * the pool: on sql server that setting would cut every lock wait of the next borrower, and a read -
+	 * which replays nothing, deliberately - would see the error 1222 it ends with.
+	 */
+	@Test
+	public void testAConnectionWhoseBoundCouldNotBeTakenOffIsKeptOutOfThePool() throws Exception {
+		final AtomicBoolean keptOut = new AtomicBoolean();
+		final Connection parent = refusingToGiveTheValueBack(mock(Connection.class), "-1");
+		try (final CachedConnection con = new CachedConnection("jdbc:mock", parent) {
+			@Override
+			void keepOutOfThePool() {
+				keptOut.set(true);
+				super.keepOutOfThePool();
+			}
+		}) {
+			storage.releaseLockBound(con, Dialect.MICROSOFT,
+				storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW));
+		}
+
+		assertTrue(keptOut.get(), "a connection left carrying our bound was handed back to the pool");
+	}
+
+	/**
+	 * Postgres needs a transaction block for a set local to mean anything: outside one the server
+	 * answers it with a warning no driver raises, so the write would run unbounded while the log read
+	 * exactly like a bounded one.
+	 */
+	@Test
+	public void testPostgresInAutoCommitIsToldNothing() throws Exception {
+		final Connection con = recording(mock(Connection.class), "0");
+		when(con.getAutoCommit()).thenReturn(true);
+
+		final ArmedLockBound armed = storage.armLockBound(con, Dialect.POSTGRES, LockBound.ROW);
+
+		assertEquals(issued, emptyList());
+		assertFalse(armed.bounded, "a set local that reaches no transaction was reported as bounding the wait");
+	}
+
+	/**
+	 * A failed setting takes the transaction back to the savepoint in front of it: a statement that
+	 * fails inside a postgres transaction aborts it, and the write would then fail with 25P02 rather
+	 * than running as unbounded as it ran before this bound existed.
+	 */
+	@Test
+	public void testPostgresTakesASavepointInFrontOfTheSetting() throws Exception {
+		final Connection con = recording(mock(Connection.class), "0");
+
+		storage.armLockBound(con, Dialect.POSTGRES, LockBound.ROW);
+
+		verify(con).setSavepoint();
+	}
+
+	/**
+	 * And lets go of it once the setting is on: a savepoint is a subtransaction of the write, and one
+	 * left open spans the whole attempt - every row that attempt writes would then carry the
+	 * subtransaction's own xid, which readers of those rows resolve through {@code pg_subtrans}. The
+	 * {@code set local} survives the release, so the bound it was taken in front of stays on.
+	 */
+	@Test
+	public void testPostgresLetsGoOfThatSavepointOnceTheSettingIsOn() throws Exception {
+		final Connection con = recording(mock(Connection.class), "0");
+		final Savepoint beforeTheBound = mock(Savepoint.class);
+		when(con.setSavepoint()).thenReturn(beforeTheBound);
+
+		storage.armLockBound(con, Dialect.POSTGRES, LockBound.ROW);
+
+		assertEquals(issued, singletonList("set local lock_timeout = 3000"),
+			"the bound itself was not issued, or was taken back");
+		verify(con).releaseSavepoint(beforeTheBound);
+		verify(con, never()).rollback(beforeTheBound);
+	}
+
+	/**
+	 * A setting that failed is taken back to that point instead, and the point is not let go of in front
+	 * of a rollback that still has to reach it.
+	 */
+	@Test
+	public void testASettingThatFailedIsTakenBackToThatSavepoint() throws Exception {
+		final Connection con = refusingTheSetting(mock(Connection.class), "0");
+		final Savepoint beforeTheBound = mock(Savepoint.class);
+		when(con.setSavepoint()).thenReturn(beforeTheBound);
+
+		final ArmedLockBound armed = storage.armLockBound(con, Dialect.POSTGRES, LockBound.ROW);
+
+		assertFalse(armed.bounded, "a setting the session refused was reported as bounding the wait");
+		verify(con).rollback(beforeTheBound);
+		verify(con, never()).releaseSavepoint(beforeTheBound);
+	}
+
+	/**
+	 * The readback is a round trip, and this bound is armed around every write of the server where the
+	 * DDL bound is armed around an open: it is paid once per pooled connection. Only this backend
+	 * writes that setting on a connection of this pool, every write puts the value back, and a
+	 * connection whose restore failed is kept out of the pool - so what was read cannot go stale.
+	 */
+	@Test
+	public void testTheReadbackIsPaidOncePerPooledConnection() throws Exception {
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				recording(mock(Connection.class), "50"))) {
+			storage.releaseLockBound(con, Dialect.MYSQL, storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW));
+			storage.releaseLockBound(con, Dialect.MYSQL, storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW));
+		}
+
+		assertEquals(issued, asList("select @@session.innodb_lock_wait_timeout",
+			"set session innodb_lock_wait_timeout=3", "set session innodb_lock_wait_timeout=50",
+			"set session innodb_lock_wait_timeout=3", "set session innodb_lock_wait_timeout=50"),
+			"the value a pooled session carries was read back more than once");
+	}
+
+	/**
+	 * A connection that is not one of this pool is asked every time: the memo above is a property of a
+	 * session this backend owns for its life, and the catalog and stamp connections are not that.
+	 */
+	@Test
+	public void testAConnectionOutsideThePoolIsAskedEveryTime() throws Exception {
+		final Connection con = recording(mock(Connection.class), "50");
+
+		storage.releaseLockBound(con, Dialect.MYSQL, storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW));
+		storage.releaseLockBound(con, Dialect.MYSQL, storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW));
+
+		assertEquals(issued.stream().filter("select @@session.innodb_lock_wait_timeout"::equals).count(), 2L,
+			"a connection outside the pool was asked for the value it carries only once");
+	}
+
+	/**
+	 * The failure an engine ends a bounded wait with is replayable, and only where this backend bounded
+	 * that wait: 55P03 and error 1222 are no conflict - the engine rolled nothing back and the blocker
+	 * is still holding the lock - and what makes them worth replaying is that the wait they cost fits
+	 * inside the replay window. Where nothing bounded the wait they stay exactly as unreplayable as
+	 * they were: a bound an operator set for themselves is not a licence for this loop to take that
+	 * wait again, and a wait nothing bounds is the one thing the window cannot govern (#903).
+	 */
+	@DataProvider
+	public Object[][] lockTimeouts() {
+		return new Object[][] {
+			{ "postgres lock_timeout", Dialect.POSTGRES, new SQLException("canceling statement due to lock timeout",
+				"55P03") },
+			{ "sql server LOCK_TIMEOUT", Dialect.MICROSOFT, new SQLException("Lock request time out period exceeded.",
+				"HY000", 1222) },
+		};
+	}
+
+	@Test(dataProvider = "lockTimeouts")
+	public void testALockTimeoutIsReplayedOnlyWhereThisBackendBoundedTheWait(String name, Dialect dialect,
+			SQLException failure) {
+		assertEquals(replayReason(failure, dialect, ArmedLockBound.alreadyTighter(LockBound.ROW, 3)),
+			"a lock wait of an attempt this backend bounded", name);
+		assertNull(replayReason(failure, dialect, ArmedLockBound.none(LockBound.ROW)),
+			name + ": a wait this backend put no bound on was replayed on a clock that cannot bound it");
+	}
+
+	/**
+	 * A mysql lock wait timeout arrives in class 40, which is the conflict this loop has replayed since
+	 * #867: it keeps that reason whether or not this bound is armed, since the line should name the
+	 * strongest thing that can be said of the failure.
+	 */
+	@Test
+	public void testAMysqlLockWaitTimeoutStaysTheConflictItWas() {
+		final SQLException lockWait = new SQLException("Lock wait timeout exceeded", "40001", 1205);
+
+		assertEquals(replayReason(lockWait, MYSQL, false, false, false, Dialect.MYSQL,
+			ArmedLockBound.none(LockBound.ROW)), "a conflict");
+		assertEquals(replayReason(lockWait, MYSQL, false, false, false, Dialect.MYSQL,
+			ArmedLockBound.alreadyTighter(LockBound.ROW, 3)), "a conflict");
+	}
+
+	/**
+	 * Not while committing, for the reason a dropped connection is not replayed there: a commit that
+	 * did not answer leaves the outcome unknown, and this loop must not apply a write twice.
+	 */
+	@Test
+	public void testALockTimeoutReportedByTheCommitIsNotReplayed() {
+		final SQLException lockTimeout = new SQLException("Lock request time out period exceeded.", "HY000", 1222);
+
+		assertNull(replayReason(lockTimeout, UNKNOWN_ENGINE, true, false, false, Dialect.MICROSOFT,
+			ArmedLockBound.alreadyTighter(LockBound.ROW, 3)));
+	}
+
+	/** And never once the attempt has committed part of its own work, whatever the failure says. */
+	@Test
+	public void testALockTimeoutOfAnAttemptThatCommittedPartOfItsWorkIsNotReplayed() {
+		final SQLException lockTimeout = new SQLException("Lock request time out period exceeded.", "HY000", 1222);
+
+		assertNull(replayReason(lockTimeout, UNKNOWN_ENGINE, false, true, false, Dialect.MICROSOFT,
+			ArmedLockBound.alreadyTighter(LockBound.ROW, 3)));
+	}
+
+	/**
+	 * Read from the failure of the operation only, never from the release of the connection: the
+	 * rollback that gives a connection back runs after the outcome was decided, so a lock timeout
+	 * reported there says nothing about the statement that failed.
+	 */
+	@Test
+	public void testALockTimeoutOfTheReleaseIsNotReplayed() {
+		final SQLException rejected = new SQLException("duplicate key", "23000", 2627);
+		rejected.addSuppressed(new SQLException("Lock request time out period exceeded.", "HY000", 1222));
+
+		assertNull(replayReason(rejected, Dialect.MICROSOFT, ArmedLockBound.alreadyTighter(LockBound.ROW, 3)));
+	}
+
+	/**
+	 * sql server has one {@code LOCK_TIMEOUT} for both waits, and every DDL of this backend but the
+	 * off-write catalog drop is issued from inside a write - which armed this bound one statement
+	 * earlier. Read live, that DDL would see the 3 s of the row bound, answer "already tighter" to its
+	 * own 5 s and run at the row bound instead: {@code DDL_LOCK_TIMEOUT_PROPERTY} would govern no DDL of
+	 * a write at all. What decides it is the value the deployment set, which the row bound remembered on
+	 * the connection when it displaced it.
+	 * <p>
+	 * The readback itself is still paid, and paid live: what the DDL has to put back is the value the
+	 * session carried a statement ago - the row bound of this very write - rather than the one the
+	 * connection was borrowed with, which is why only the row bound is remembered per pooled connection.
+	 */
+	@Test
+	public void testTheDdlBoundInsideAWriteIsDecidedAgainstWhatTheDeploymentSet() throws Exception {
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				liveLockTimeout(mock(Connection.class), "-1"))) {
+			final ArmedLockBound row = storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW);
+			storage.withDdlLockBound(con, Dialect.MICROSOFT, () -> {
+				issued.add("the ddl");
+				return null;
+			});
+			storage.releaseLockBound(con, Dialect.MICROSOFT, row);
+		}
+
+		assertEquals(issued, asList(
+			// the row bound of the write, against what the session carried
+			"select @@lock_timeout", "set lock_timeout 3000",
+			// the DDL inside it, read live and armed at its own property rather than left at the row bound
+			"select @@lock_timeout", "set lock_timeout 5000",
+			"the ddl",
+			// what the session carried a statement before the DDL, which is the row bound of this write
+			"set lock_timeout 3000",
+			// and the value the deployment set, once the write is through
+			"set lock_timeout -1"));
+	}
+
+	/**
+	 * And a DDL that gives up at that bound is reported as what it is, naming its own property - the
+	 * rename of #885, which a DDL left at the row bound would lose along with the bound: it would arrive
+	 * as the bare error 1222, sending an operator to neither property.
+	 */
+	@Test
+	public void testADdlInsideAWriteThatGivesUpNamesItsOwnProperty() throws Exception {
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				liveLockTimeout(mock(Connection.class), "-1"))) {
+			storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW);
+
+			try {
+				storage.withDdlLockBound(con, Dialect.MICROSOFT, () -> {
+					throw new SQLException("Lock request time out period exceeded.", "HY000", 1222);
+				});
+				fail("a DDL that gave up on its lock went through");
+			} catch (SQLException e) {
+				assertTrue(e.getMessage().contains(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY), e.getMessage());
+			}
+		}
+	}
+
+	/**
+	 * And that value is put on the session, not merely decided against. A deployment giving up sooner
+	 * than the DDL bound keeps its own figure - the argument that leaves oracle alone - but "keeps what
+	 * it has" is not what the session has once the row bound of this write is on that very setting: at a
+	 * {@code LOCK_TIMEOUT} of 4 s the DDL would run at the 3 s of the row bound, tighter than either
+	 * property, and give up as the bare error 1222. What the deployment set goes on for the length of
+	 * the DDL, and the row bound of the write goes back on behind it.
+	 */
+	@Test
+	public void testADdlInsideAWriteRunsAtWhatTheDeploymentSetWhereThatIsTighter() throws Exception {
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				liveLockTimeout(mock(Connection.class), "4000"))) {
+			final ArmedLockBound row = storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW);
+			storage.withDdlLockBound(con, Dialect.MICROSOFT, () -> {
+				issued.add("the ddl");
+				return null;
+			});
+			storage.releaseLockBound(con, Dialect.MICROSOFT, row);
+		}
+
+		assertEquals(issued, asList(
+			// the row bound of the write, over a deployment looser than it
+			"select @@lock_timeout", "set lock_timeout 3000",
+			// the DDL inside it, at the value the deployment set rather than at the row bound it met
+			"select @@lock_timeout", "set lock_timeout 4000",
+			"the ddl",
+			// the row bound of the write back, and the deployment's value once the write is through
+			"set lock_timeout 3000", "set lock_timeout 4000"));
+	}
+
+	/**
+	 * The same where an operator raised the DDL bound for an index build and the deployment bounds every
+	 * lock wait of its sessions: 30 s asked for, 10 s allowed, and the row bound of the write is neither.
+	 */
+	@Test
+	public void testTheSameWhereTheDdlBoundWasRaisedForAnIndexBuild() throws Exception {
+		System.setProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY, "30");
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				liveLockTimeout(mock(Connection.class), "10000"))) {
+			storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW);
+			storage.withDdlLockBound(con, Dialect.MICROSOFT, () -> {
+				issued.add("the ddl");
+				return null;
+			});
+		}
+
+		assertEquals(issued, asList(
+			"select @@lock_timeout", "set lock_timeout 3000",
+			"select @@lock_timeout", "set lock_timeout 10000",
+			"the ddl",
+			"set lock_timeout 3000"));
+	}
+
+	/**
+	 * And a DDL that gives up under that value is left exactly as it arrived: what ended the wait is the
+	 * deployment's own {@code LOCK_TIMEOUT}, and naming this property for it would send an operator to
+	 * raise a value that governs nothing while the deployment's own is the tighter one.
+	 */
+	@Test
+	public void testADdlThatGaveUpAtTheDeploymentsValueIsNotNamedByThisProperty() throws Exception {
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				liveLockTimeout(mock(Connection.class), "4000"))) {
+			storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW);
+
+			try {
+				storage.withDdlLockBound(con, Dialect.MICROSOFT, () -> {
+					throw new SQLException("Lock request time out period exceeded.", "HY000", 1222);
+				});
+				fail("a DDL that gave up on its lock went through");
+			} catch (SQLException e) {
+				assertFalse(e.getMessage().contains(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY),
+					"a wait the deployment's own value ended was reported as this property's doing: " + e.getMessage());
+				assertEquals(e.getErrorCode(), 1222, "the failure of the engine was not handed through as it arrived");
+			}
+		}
+	}
+
+	/**
+	 * And the same of a lookup, which is the other kind of work this bound is armed around: the one
+	 * deciding each drop of a clear wraps whatever it sees in a {@code StorageRuntimeException}
+	 * ({@code isExistsTable}), so it reaches the rename by the unchecked arm rather than the checked
+	 * one. Both arms read the same thing - whether the session is carrying this bound's own figure -
+	 * and keying either of them on "a setting of ours was issued" instead would name this property for
+	 * a wait the deployment's own value ended.
+	 */
+	@Test
+	public void testALookupThatGaveUpAtTheDeploymentsValueIsNotNamedByThisPropertyEither() throws Exception {
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				liveLockTimeout(mock(Connection.class), "4000"))) {
+			storage.armLockBound(con, Dialect.MICROSOFT, LockBound.ROW);
+
+			try {
+				storage.withDdlLockBound(con, Dialect.MICROSOFT, () -> {
+					throw new StorageRuntimeException(
+						new SQLException("Lock request time out period exceeded.", "HY000", 1222));
+				});
+				fail("a lookup that gave up on its lock went through");
+			} catch (StorageRuntimeException e) {
+				assertFalse(String.valueOf(e.getMessage()).contains(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY),
+					"a wait the deployment's own value ended was reported as this property's doing: " + e.getMessage());
+			}
+		}
+	}
+
+	/**
+	 * On mysql the two waits are two variables - {@code innodb_lock_wait_timeout} for the row lock,
+	 * {@code lock_wait_timeout} for the metadata lock a DDL waits for - so what the row bound displaced
+	 * describes neither the other's session nor its default. Reading it there would leave a DDL waiting
+	 * a year because a deployment had tightened the row lock to a second.
+	 */
+	@Test
+	public void testTheDdlBoundOfAMysqlWriteIsDecidedAgainstItsOwnVariable() throws Exception {
+		try (final CachedConnection con = new CachedConnection("jdbc:mock",
+				answering(mock(Connection.class),
+					"select @@session.innodb_lock_wait_timeout", "1",
+					"select @@session.lock_wait_timeout", "31536000"))) {
+			storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW);
+			storage.withDdlLockBound(con, Dialect.MYSQL, () -> {
+				issued.add("the ddl");
+				return null;
+			});
+		}
+
+		assertEquals(issued, asList(
+			// the row bound: this session gives up sooner than ours would, so it keeps what it has
+			"select @@session.innodb_lock_wait_timeout",
+			// and the metadata lock is bounded all the same, against the variable that bounds it
+			"select @@session.lock_wait_timeout", "set session lock_wait_timeout=5",
+			"the ddl",
+			"set session lock_wait_timeout=31536000"));
+	}
+
+	/**
+	 * The latches these bounds warn through are per bound, not per storage: they are armed by different
+	 * code on different paths, and an open whose DDL bound this session would not take says nothing
+	 * about the writes behind it. Through one latch, the first open of a backend would silence every
+	 * write of it - which is what these two assertions, taken together, keep from happening.
+	 */
+	@Test
+	public void testTheWarningLatchesAreOnePerBound() throws Exception {
+		final Connection con = refusingTheSetting(mock(Connection.class), "31536000");
+
+		storage.armLockBound(con, Dialect.MYSQL, LockBound.DDL);
+
+		assertTrue(storage.lockBoundNotSetWarned.get(LockBound.DDL).get(),
+			"a setting the session refused was not reported for the bound that was armed");
+		assertFalse(storage.lockBoundNotSetWarned.get(LockBound.ROW).get(),
+			"the DDL bound of an open silenced the row lock bound of every write behind it");
+
+		storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW);
+
+		assertTrue(storage.lockBoundNotSetWarned.get(LockBound.ROW).get(),
+			"the row lock bound of a write said nothing of its own");
+	}
+
+	/**
+	 * The same, for the moment throttling what a bound left behind says - and asked of the row bound,
+	 * which is the one a single shared moment would leave unsaid: what the two of them have to be is
+	 * one per bound, in both directions.
+	 */
+	@Test
+	public void testTheLatchOfABoundLeftBehindIsOnePerBoundToo() throws Exception {
+		final Connection con = refusingToGiveTheValueBack(mock(Connection.class), "31536000");
+
+		storage.releaseLockBound(con, Dialect.MYSQL, storage.armLockBound(con, Dialect.MYSQL, LockBound.ROW));
+
+		assertTrue(storage.lockBoundLeftBehindWarned.get(LockBound.ROW).get() != 0,
+			"a bound that could not be taken off was not reported for the bound that was armed");
+		assertEquals(storage.lockBoundLeftBehindWarned.get(LockBound.DDL).get(), 0L,
+			"the row lock bound of a write silenced the DDL bound of every open of this backend");
+	}
+
+	private static String replayReason(SQLException failure, Dialect dialect, ArmedLockBound rowLock) {
+		return replayReason(failure, UNKNOWN_ENGINE, false, false, false, dialect, rowLock);
+	}
+
+	/**
+	 * The two questions {@code write()} asks after a failed attempt, composed here the way it composes
+	 * them: the conflict class is read off the failure once and handed to the reason, rather than being
+	 * asked for again.
+	 */
+	private static String replayReason(SQLException failure, String driver, boolean committing,
+			boolean partlyCommitted, boolean connectionClosed, Dialect dialect, ArmedLockBound rowLock) {
+		return JDBCStorage.replayReason(JDBCStorage.conflictVerdict(failure, driver).conflict, failure, committing,
+			partlyCommitted, connectionClosed, dialect, rowLock);
+	}
+
+	/**
+	 * A connection recording every session statement it is given, and answering the readback of the
+	 * setting with the value a session of that engine carries.
+	 */
+	private Connection recording(final Connection con, final String carries) throws SQLException {
+		final Statement statement = mock(Statement.class);
+		when(statement.execute(anyString())).thenAnswer(invocation -> {
+			issued.add((String) invocation.getArguments()[0]);
+			return false;
+		});
+		when(statement.executeQuery(anyString())).thenAnswer(invocation -> {
+			issued.add((String) invocation.getArguments()[0]);
+			final ResultSet carried = mock(ResultSet.class);
+			when(carried.next()).thenReturn(true, false);
+			when(carried.getString(1)).thenReturn(carries);
+			return carried;
+		});
+		when(con.createStatement()).thenReturn(statement);
+		return con;
+	}
+
+	/**
+	 * A connection answering each readback its own value, which is what an engine whose two waits are
+	 * two variables does: every other fixture here answers one value to every query, so a case over one
+	 * of them cannot tell the value of one setting from the value of the other.
+	 *
+	 * @param answers the query and the value it is answered with, in pairs
+	 */
+	private Connection answering(final Connection con, final String... answers) throws SQLException {
+		final Statement statement = mock(Statement.class);
+		when(statement.execute(anyString())).thenAnswer(invocation -> {
+			issued.add((String) invocation.getArguments()[0]);
+			return false;
+		});
+		when(statement.executeQuery(anyString())).thenAnswer(invocation -> {
+			final String query = (String) invocation.getArguments()[0];
+			issued.add(query);
+			final ResultSet carried = mock(ResultSet.class);
+			when(carried.next()).thenReturn(true, false);
+			for (int i = 0; i < answers.length; i += 2) {
+				if (answers[i].equals(query)) {
+					when(carried.getString(1)).thenReturn(answers[i + 1]);
+					return carried;
+				}
+			}
+			throw new SQLException("this fixture answers no " + query, "42000");
+		});
+		when(con.createStatement()).thenReturn(statement);
+		return con;
+	}
+
+	/**
+	 * A sql server connection answering the readback with what the last {@code set lock_timeout} left on
+	 * it - a live session rather than a fixed value, which is what a bound armed inside another one
+	 * meets.
+	 */
+	private Connection liveLockTimeout(final Connection con, final String initially) throws SQLException {
+		final AtomicReference<String> carried = new AtomicReference<>(initially);
+		final Statement statement = mock(Statement.class);
+		when(statement.execute(anyString())).thenAnswer(invocation -> {
+			final String sql = (String) invocation.getArguments()[0];
+			issued.add(sql);
+			if (sql.startsWith("set lock_timeout ")) {
+				carried.set(sql.substring("set lock_timeout ".length()));
+			}
+			return false;
+		});
+		when(statement.executeQuery(anyString())).thenAnswer(invocation -> {
+			issued.add((String) invocation.getArguments()[0]);
+			final ResultSet rows = mock(ResultSet.class);
+			when(rows.next()).thenReturn(true, false);
+			when(rows.getString(1)).thenReturn(carried.get());
+			return rows;
+		});
+		when(con.createStatement()).thenReturn(statement);
+		return con;
+	}
+
+	/** A connection that says what it carries and will not take the setting of a bound at all. */
+	private Connection refusingTheSetting(final Connection con, final String carries) throws SQLException {
+		final Statement statement = recording(con, carries).createStatement();
+		doAnswer(invocation -> {
+			issued.add((String) invocation.getArguments()[0]);
+			throw new SQLException("this session takes no such setting", "42000");
+		}).when(statement).execute(anyString());
+		return con;
+	}
+
+	/** A connection that takes the bound and will not take back the value that bound displaced. */
+	private Connection refusingToGiveTheValueBack(final Connection con, final String carries) throws SQLException {
+		final Statement statement = recording(con, carries).createStatement();
+		final AtomicBoolean bound = new AtomicBoolean();
+		doAnswer(invocation -> {
+			issued.add((String) invocation.getArguments()[0]);
+			if (!bound.compareAndSet(false, true)) {
+				throw new SQLException("the connection went before the value could be given back", "08006");
+			}
+			return false;
+		}).when(statement).execute(anyString());
+		return con;
+	}
+}
diff --git a/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCStorageRetryTest.java b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCStorageRetryTest.java
index 3c64867..45b0ebb 100644
--- a/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCStorageRetryTest.java
+++ b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/JDBCStorageRetryTest.java
@@ -15,6 +15,7 @@
  */
 package org.opends.server.backends.jdbc;
 
+import org.forgerock.opendj.ldap.ByteString;
 import org.forgerock.opendj.server.config.server.JDBCBackendCfg;
 import org.opends.server.DirectoryServerTestCase;
 import org.opends.server.backends.jdbc.JDBCStorage.Conflict;
@@ -40,6 +41,8 @@
 import java.sql.SQLNonTransientConnectionException;
 import java.sql.SQLRecoverableException;
 import java.sql.Statement;
+import java.util.ArrayList;
+import java.util.List;
 import java.util.Locale;
 import java.util.Properties;
 import java.util.concurrent.TimeUnit;
@@ -49,6 +52,8 @@
 import java.util.function.Predicate;
 import java.util.logging.Logger;
 
+import static java.util.Collections.emptyList;
+import static java.util.Collections.frequency;
 import static org.forgerock.i18n.LocalizableMessage.raw;
 import static org.forgerock.opendj.ldap.ResultCode.OTHER;
 import static org.mockito.Mockito.any;
@@ -56,7 +61,9 @@
 import static org.mockito.Mockito.anyInt;
 import static org.mockito.Mockito.anyString;
 import static org.mockito.Mockito.atLeastOnce;
+import static org.mockito.Mockito.doAnswer;
 import static org.mockito.Mockito.doNothing;
+import static org.mockito.Mockito.doReturn;
 import static org.mockito.Mockito.doThrow;
 import static org.mockito.Mockito.eq;
 import static org.mockito.Mockito.mock;
@@ -98,6 +105,14 @@
   /** A MySQL-wire-compatible driver, whose class name carries no engine this backend recognises. */
   private static final String MARIADB = "org.mariadb.jdbc.Connection";
 
+  /**
+   * What the cases below arm around a failure they classify: nothing. They are about the other classes of
+   * failure - a conflict, a connection the database dropped - and a row lock bound this backend did not put on
+   * is the state every one of them was written under (#915).
+   */
+  private static final JDBCStorage.ArmedLockBound NO_ROW_LOCK_BOUND =
+      JDBCStorage.ArmedLockBound.none(JDBCStorage.LockBound.ROW);
+
   /** The tree every write of this test opens; the table name behind it is a hash of this name. */
   private static final TreeName TREE = new TreeName("dc=example,dc=com", "id2entry");
 
@@ -134,6 +149,11 @@
   {
   }
 
+  /** sql server, whose one {@code LOCK_TIMEOUT} is read back and put back around every write. */
+  interface microsoftConnection extends Connection
+  {
+  }
+
   /**
    * A MySQL-wire-compatible driver none of the four engines is recognised in - MariaDB Connector/J, an Aurora-
    * or Percona-branded one. It reports a lock wait timeout as 1205 under class 40 exactly as Connector/J does,
@@ -1138,6 +1158,358 @@
   }
 
   /**
+   * The wait a write takes for a row lock another session holds is bounded on the session of the attempt
+   * ({@code ROW_LOCK_TIMEOUT_PROPERTY}, #915), and the failure the engine ends that wait with is replayed. On
+   * postgres that failure is 55P03, which is no conflict - nothing was rolled back and the blocker still holds
+   * the lock - and what makes it worth replaying is that the wait it cost fits inside the replay window, which
+   * is the only thing that lets a clock bound these replays at all (#903).
+   * <p>
+   * The bound is armed per attempt: on postgres it is a {@code set local}, which the rollback of the attempt
+   * that failed discards along with everything else the attempt did.
+   */
+  @Test
+  public void testARowLockWaitThisBackendBoundedIsReplayed() throws Exception
+  {
+    final JDBCStorage storage = storageOverAnEngine(postgresConnection.class, true);
+    final List<String> sessionStatements = sessionStatementsOf(engineConnection);
+    when(statements.executeUpdate()).thenThrow(sql(0, "55P03")).thenReturn(1);
+    final AtomicInteger attempts = new AtomicInteger();
+
+    storage.write(txn -> {
+      attempts.incrementAndGet();
+      txn.openTree(TREE, false);
+      txn.put(TREE, ByteString.valueOfUtf8("dc=example,dc=com"), ByteString.valueOfUtf8("an entry"));
+    });
+
+    assertEquals(attempts.get(), 2, "a row lock wait this backend bounded was not replayed");
+    assertEquals(sessionStatements.stream().filter("set local lock_timeout = 3000"::equals).count(), 2L,
+        "the bound was not armed once per attempt: " + sessionStatements);
+  }
+
+  /**
+   * And where nothing bounds that wait it is not replayed, which is what a deployment asks for by setting the
+   * property to 0: this loop must not take again a wait that has no end of its own, and the failure is left
+   * exactly as unreplayable as it was before #915 - the behaviour every engine had, and oracle still has.
+   */
+  @Test
+  public void testARowLockWaitNothingBoundedIsNotReplayed() throws Exception
+  {
+    System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, "0");
+    try
+    {
+      final JDBCStorage storage = storageOverAnEngine(postgresConnection.class, true);
+      final List<String> sessionStatements = sessionStatementsOf(engineConnection);
+      when(statements.executeUpdate()).thenThrow(sql(0, "55P03")).thenReturn(1);
+      final AtomicInteger attempts = new AtomicInteger();
+
+      try
+      {
+        storage.write(txn -> {
+          attempts.incrementAndGet();
+          txn.openTree(TREE, false);
+          txn.put(TREE, ByteString.valueOfUtf8("dc=example,dc=com"), ByteString.valueOfUtf8("an entry"));
+        });
+        fail("a lock wait this backend put no bound on was replayed");
+      }
+      catch (StorageRuntimeException expected)
+      {
+        assertEquals(attempts.get(), 1, "an unbounded lock wait was taken a second time");
+      }
+      assertEquals(sessionStatements, emptyList(), "a bound turned off was armed all the same");
+    }
+    finally
+    {
+      System.clearProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY);
+    }
+  }
+
+  /**
+   * A DDL of an attempt commits, and that commit is the end of the transaction the row lock bound was armed
+   * around: on postgres the {@code set local} goes with it by itself, and on the two engines whose setting is
+   * the session's {@code commitStatement()} is what takes it off. What follows such a DDL is a write that has
+   * committed part of its work, so it is out of the replay whatever it waits for - and a bound on a wait
+   * nothing can replay only fails a write at 3 s where it used to wait for the blocker and go through.
+   * <p>
+   * The order is what says it: the value is put back before the statement behind the DDL is issued, rather
+   * than in the finally of the attempt, which runs after the whole write is through.
+   */
+  @Test
+  public void testTheRowLockBoundComesOffWhereADdlOfTheAttemptCommits() throws Exception
+  {
+    final JDBCStorage storage = storageOverAnEngine(mysqlConnection.class, false);
+    final List<String> issued = statementsOf(engineConnection, "50");
+
+    storage.write(txn -> {
+      txn.openTree(TREE, true); // a create index, which commits
+      txn.put(TREE, ByteString.valueOfUtf8("dc=example,dc=com"), ByteString.valueOfUtf8("an entry"));
+    });
+
+    final int boundArmed = issued.indexOf("set session innodb_lock_wait_timeout=3");
+    final int boundOff = issued.indexOf("set session innodb_lock_wait_timeout=50");
+    final int theDdl = indexOfFirst(issued, "create index k_");
+    final int behindTheDdl = indexOfFirst(issued, "insert into ");
+    assertTrue(boundArmed >= 0 && theDdl > boundArmed, "the attempt issued no bounded DDL: " + issued);
+    assertTrue(boundOff > theDdl, "the row lock bound was not taken off where the DDL committed: " + issued);
+    assertTrue(behindTheDdl > boundOff,
+        "the rest of the write ran under a bound its DDL had already taken out of the replay: " + issued);
+    // and taken off once: the transaction records that it carries none, so the finally of write() has
+    // nothing left to give back and the value is not put back over whatever the rest of the write left
+    assertEquals(frequency(issued, "set session innodb_lock_wait_timeout=50"), 1,
+        "the value the row lock bound displaced was put back twice: " + issued);
+  }
+
+  /**
+   * And it is not put back on a connection the driver reports closed: there is nothing there to give it back
+   * to, and the setting would fail on a dead session and be reported as a bound left behind - a database
+   * restart under write load would read as a stream of stranded bounds for connections that are gone.
+   * <p>
+   * Which is what the guard is keyed on, rather than on the attempt being classified as a dropped connection:
+   * the case below holds the other side of that.
+   */
+  @Test
+  public void testTheRowLockBoundIsNotPutBackOnAConnectionTheDatabaseDropped() throws Exception
+  {
+    final JDBCStorage storage = storageOverAnEngine(mysqlConnection.class, false);
+    final List<String> issued = statementsOf(engineConnection, "50");
+    // the state the driver leaves such a connection in, which is what tells it from a connection a failure of
+    // this class merely passed through: mssql-jdbc closes the connection for any error of severity 20 and
+    // above before it throws, and connector/j closes one the server hung up on
+    when(engineConnection.isClosed()).thenReturn(true);
+
+    try
+    {
+      storage.write(txn -> {
+        throw new StorageRuntimeException(sql(0, "08006"));
+      });
+      fail("the drop of the connection was swallowed");
+    }
+    catch (StorageRuntimeException expected)
+    {
+      assertTrue(JDBCStorage.isConnectionFailure(expected), "the failure this case rests on is not a drop");
+    }
+
+    assertTrue(issued.contains("set session innodb_lock_wait_timeout=3"),
+        "the write armed no row lock bound: " + issued);
+    assertEquals(frequency(issued, "set session innodb_lock_wait_timeout=50"), 0,
+        "the bound was given back on a connection the database had dropped: " + issued);
+  }
+
+  /**
+   * While a connection that only this attempt's classification made look dropped is given its value back. A
+   * write enrolling a tree opens a connection of the catalog's own, and a database refusing that connect
+   * answers a state of class 08 - mysql answers its connection limit with 08004, sql server with 08S01 - which
+   * puts the attempt in the replay as a dropped connection with the pooled connection of the write untouched.
+   * That connection does go back to the pool: the rollback went through, the validation of the next borrow is
+   * {@code isValid()}, and a live session passes it. On the two engines whose bound is a session setting,
+   * skipping the restore there hands the next borrow this backend's own 3 s - a read among them, and a read
+   * that gives up at a lock wait has no replay to absorb it.
+   */
+  @Test
+  public void testTheRowLockBoundIsPutBackOnALiveConnectionARefusedCatalogConnectMadeLookDropped() throws Exception
+  {
+    final JDBCStorage storage = storageOverAnEngine(mysqlConnection.class, false);
+    final List<String> issued = statementsOf(engineConnection, "50");
+
+    try
+    {
+      storage.write(txn -> {
+        throw new StorageRuntimeException(new SQLException("Too many connections", "08004", 1040));
+      });
+      fail("the refusal was swallowed");
+    }
+    catch (StorageRuntimeException expected)
+    {
+      assertTrue(JDBCStorage.isConnectionFailure(expected), "the failure this case rests on is not read as a drop");
+    }
+
+    // every attempt of the replay, rather than one of them: what the case is about is that no attempt leaves
+    // its bound on a connection it hands back, and a refusal of this class is replayed like any other drop
+    final int armed = frequency(issued, "set session innodb_lock_wait_timeout=3");
+    assertTrue(armed > 0, "the write armed no row lock bound: " + issued);
+    assertEquals(frequency(issued, "set session innodb_lock_wait_timeout=50"), armed,
+        "the bound was left on a connection the driver still reports open: " + issued);
+  }
+
+  /**
+   * A DDL told to wait as this backend waited before its bound existed ({@code DDL_LOCK_TIMEOUT_PROPERTY} at 0)
+   * is not left waiting at the row bound of the write it is issued from. On postgres that bound is a
+   * {@code set local}, which bounds every lock wait of the transaction, so the create table of an open would
+   * give up at 3 s - as the bare vendor error, nothing of ours being armed around it - and be replayed as a
+   * lock wait of the attempt until the window was spent. The bound comes off in front of the DDL instead.
+   * <p>
+   * Off the write and not off each of its DDL: the open of a tree that is not there issues two - the create
+   * table and the create index behind it - and the second has no bound of this backend left to lift.
+   */
+  @Test
+  public void testADdlToldToWaitDoesNotWaitAtTheRowBoundOnPostgres() throws Exception
+  {
+    System.setProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY, "0");
+    try
+    {
+      final JDBCStorage storage = storageOverAnEngine(postgresConnection.class, false);
+      tablesAreNotThere(engineConnection);
+      final List<String> issued = statementsOf(engineConnection, "0");
+
+      storage.write(txn -> txn.openTree(TREE, true));
+
+      final int bound = issued.indexOf("set local lock_timeout = 3000");
+      final int lifted = issued.indexOf("set local lock_timeout to default");
+      final int theDdl = indexOfFirst(issued, "create index ");
+      assertTrue(bound >= 0, "the write armed no row lock bound: " + issued);
+      assertTrue(lifted > bound, "the row lock bound was not taken off in front of an unbounded DDL: " + issued);
+      assertTrue(theDdl > lifted, "the DDL ran under the row bound of the write that issued it: " + issued);
+      assertTrue(indexOfFirst(issued, "create table ") >= 0, "the open of this case created no table: " + issued);
+      assertEquals(frequency(issued, "set local lock_timeout to default"), 1,
+          "the row lock bound was lifted once per DDL rather than once per write: " + issued);
+    }
+    finally
+    {
+      System.clearProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY);
+    }
+  }
+
+  /**
+   * The same on sql server, where the setting is the session's and what puts the wait back where the deployment
+   * left it is the value the row bound displaced - here the -1 that waits forever, which is what this backend
+   * waited before either bound existed. The DDL is the create table of a tree that is not there: this engine
+   * indexes nothing behind it, {@code k} being a {@code varbinary(max)} no index key column can hold.
+   */
+  @Test
+  public void testADdlToldToWaitDoesNotWaitAtTheRowBoundOnSqlServer() throws Exception
+  {
+    System.setProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY, "0");
+    try
+    {
+      final JDBCStorage storage = storageOverAnEngine(microsoftConnection.class, false);
+      tablesAreNotThere(engineConnection);
+      final List<String> issued = statementsOf(engineConnection, "-1");
+
+      storage.write(txn -> txn.openTree(TREE, true));
+
+      final int bound = issued.indexOf("set lock_timeout 3000");
+      final int lifted = issued.indexOf("set lock_timeout -1");
+      final int theDdl = indexOfFirst(issued, "create table ");
+      assertTrue(bound >= 0, "the write armed no row lock bound: " + issued);
+      assertTrue(lifted > bound, "the deployment's own value was not put back in front of an unbounded DDL: "
+          + issued);
+      assertTrue(theDdl > lifted, "the DDL ran under the row bound of the write that issued it: " + issued);
+    }
+    finally
+    {
+      System.clearProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY);
+    }
+  }
+
+  /**
+   * While a DDL that has a bound of its own is left exactly where it was: it arms that bound over the row one
+   * and puts it back afterwards, so taking the row bound off in front of it would be a round trip per DDL -
+   * about 25 of them per suffix of an open - buying nothing.
+   */
+  @Test
+  public void testABoundedDdlLeavesTheRowBoundOfTheWriteWhereItIs() throws Exception
+  {
+    final JDBCStorage storage = storageOverAnEngine(postgresConnection.class, false);
+    final List<String> issued = statementsOf(engineConnection, "0");
+
+    storage.write(txn -> txn.openTree(TREE, true));
+
+    assertTrue(issued.contains("set local lock_timeout = 3000"), "the write armed no row lock bound: " + issued);
+    assertEquals(frequency(issued, "set local lock_timeout to default"), 0,
+        "a DDL with a bound of its own took the row bound off in front of itself: " + issued);
+  }
+
+  /**
+   * And a DDL that did give up at its own bound inside a write is replayed like any other lock wait of a
+   * bounded attempt - on the copy of the bound that attempt ran under, which is what {@code write()} keeps
+   * for the replay while {@code commitStatement()} takes the transaction's copy off at the DDL that ends it.
+   * Decided on the transaction's copy instead, such a failure would be thrown at the first attempt.
+   */
+  @Test
+  public void testADdlLockWaitInsideAWriteIsReplayedOnTheBoundTheAttemptRanUnder() throws Exception
+  {
+    final JDBCStorage storage = storageOverAnEngine(postgresConnection.class, false);
+    sessionStatementsOf(engineConnection); // the row bound of the attempt: the fixture takes no session statement
+    when(statements.executeUpdate()).thenThrow(sql(0, "55P03")).thenReturn(0);
+    final AtomicInteger attempts = new AtomicInteger();
+
+    storage.write(txn -> {
+      attempts.incrementAndGet();
+      txn.openTree(TREE, true);
+    });
+
+    assertEquals(attempts.get(), 2, "a DDL that gave up at its own bound inside a bounded attempt was not replayed");
+  }
+
+  /** Where the first statement of the attempt starting with the given text was issued, or -1. */
+  private static int indexOfFirst(List<String> issued, String startsWith)
+  {
+    for (int i = 0; i < issued.size(); i++)
+    {
+      if (issued.get(i).startsWith(startsWith))
+      {
+        return i;
+      }
+    }
+    return -1;
+  }
+
+  /**
+   * Answers the metadata of a connection with a database holding none of the tables asked about, so that an
+   * open creates them: the fixture of {@code storageOverAnEngine()} answers that every table asked about is
+   * there, which is the existing backend most cases here are about.
+   */
+  private static void tablesAreNotThere(Connection con) throws Exception
+  {
+    final ResultSet none = mock(ResultSet.class);
+    when(none.next()).thenReturn(false);
+    // the metadata is taken out of the chain first, and the stubbing is a doReturn: asking when() for the
+    // value of a call already stubbed with an answer would run that answer here, and it stubs a result set
+    // of its own as it goes
+    final DatabaseMetaData metaData = con.getMetaData();
+    doReturn(none).when(metaData).getTables(any(), any(), any(), any());
+  }
+
+  /**
+   * Records the session statements a connection is given - the bound of an attempt among them - on a fixture
+   * whose {@code createStatement()} otherwise refuses them. Stubbed through {@code doReturn}, since asking
+   * {@code when()} for the value of a call already stubbed to throw would raise that throw here.
+   */
+  private static List<String> sessionStatementsOf(Connection con) throws Exception
+  {
+    final List<String> issued = new ArrayList<>();
+    final Statement statement = mock(Statement.class);
+    when(statement.execute(anyString())).thenAnswer(invocation -> {
+      issued.add((String) invocation.getArguments()[0]);
+      return false;
+    });
+    doReturn(statement).when(con).createStatement();
+    return issued;
+  }
+
+  /**
+   * The same, with the statements of the work itself in the very same list and the readback of a session
+   * setting answered: what a case reads off this is the order the two were issued in, which is what a bound
+   * armed around a transaction and taken off inside it can only be pinned by.
+   */
+  private List<String> statementsOf(Connection con, String carries) throws Exception
+  {
+    final List<String> issued = sessionStatementsOf(con);
+    final Statement statement = con.createStatement();
+    when(statement.executeQuery(anyString())).thenAnswer(invocation -> {
+      issued.add((String) invocation.getArguments()[0]);
+      final ResultSet carried = mock(ResultSet.class);
+      when(carried.next()).thenReturn(true, false);
+      when(carried.getString(1)).thenReturn(carries);
+      return carried;
+    });
+    doAnswer(invocation -> {
+      issued.add((String) invocation.getArguments()[0]);
+      return statements;
+    }).when(con).prepareStatement(anyString());
+    return issued;
+  }
+
+  /**
    * mysql commits before a DDL statement whether asked to or not, so a create index that failed there has
    * committed everything the transaction did before it just as surely as one that succeeded: the attempt is out
    * of the replay whatever the failure says.
@@ -1432,7 +1804,7 @@
       boolean connectionClosed)
   {
     return JDBCStorage.replayReason(conflictOf(failure, driver), failure, committing, partlyCommitted,
-        connectionClosed);
+        connectionClosed, JDBCStorage.dialectOf(driver), NO_ROW_LOCK_BOUND);
   }
 
   /**
diff --git a/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/TestCase.java b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/TestCase.java
index 3cbdbfe..f3b3500 100644
--- a/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/TestCase.java
+++ b/opendj-server-legacy/src/test/java/org/opends/server/backends/jdbc/TestCase.java
@@ -291,12 +291,18 @@
 	 * Oracle asserts the other half of the same contract: nothing of ours is set there, because its
 	 * own {@code ddl_lock_timeout} gives up at once - so the drop still fails rather than hanging, and
 	 * the failure names the engine's own doing rather than a property of ours that armed nothing.
+	 * <p>
+	 * The bound is set looser than the row lock bound of the write this drop runs inside, deliberately:
+	 * sql server has one {@code LOCK_TIMEOUT} for both waits, so a DDL bound tighter than the row one
+	 * passes this case whether this backend armed it or merely inherited the row bound of the write
+	 * around it - and inheriting it is exactly what would leave this property governing nothing and the
+	 * failure naming neither property (#915).
 	 */
 	@Test(timeOut = 120000)
 	public void testTheDdlGivesUpOnALockAnotherSessionHolds() throws Exception {
 		final TreeName tree = new TreeName("testDdlLockBound", "tree");
 		final JDBCStorage storage = new JDBCStorage(createBackendCfg(), null);
-		System.setProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY, "2");
+		System.setProperty(JDBCStorage.DDL_LOCK_TIMEOUT_PROPERTY, "7");
 		try {
 			storage.open(AccessMode.READ_WRITE);
 			storage.write(new WriteOperation() {
@@ -728,6 +734,93 @@
 			});
 	}
 
+	/**
+	 * A write queued behind a row lock another session holds gives up at
+	 * {@code ROW_LOCK_TIMEOUT_PROPERTY} and is replayed, and the replay goes through once that session
+	 * lets the rows go: the wait ends inside the replay window instead of consuming it, which is what
+	 * left the operation with no replay at all before (#903, #915).
+	 * <p>
+	 * On an engine with no session setting for that wait - oracle - the same write waits the blocker
+	 * out on its first attempt and needs no replay, which is what "left alone" means there. That is
+	 * asserted rather than skipped: it is the difference this bound makes, read from the one place it
+	 * shows.
+	 */
+	@Test(timeOut = 300000)
+	public void testAWriteBlockedByARowLockIsReplayedInsideTheWindow() throws Exception {
+		final int boundSeconds = 2;
+		final TreeName tree = new TreeName("testRowLockBound", "tree");
+		final JDBCStorage storage = new JDBCStorage(createBackendCfg(), null);
+		System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, Integer.toString(boundSeconds));
+		final ExecutorService releasing = Executors.newSingleThreadExecutor();
+		try {
+			storage.open(AccessMode.READ_WRITE);
+			storage.write(new WriteOperation() {
+				@Override
+				public void run(WriteableTransaction txn) throws Exception {
+					txn.openTree(tree, true);
+					txn.put(tree, key(1), value(1));
+				}
+			});
+
+			// a session of its own, taking an exclusive lock on the row the write below wants: an update
+			// that changes nothing takes the same lock as one that does, and it parses on all four engines
+			try (final Connection holder = DriverManager.getConnection(getJdbcUrl())) {
+				holder.setAutoCommit(false);
+				try (final PreparedStatement lock = holder.prepareStatement(
+						"update " + storage.getTableName(tree) + " set v=v")) {
+					lock.executeUpdate();
+				}
+				// held for twice the bound, so that the first attempt of the write gives up on it and the
+				// replay finds the row free - both inside the replay window, which is what this is about
+				final Future<?> released = releasing.submit(new Callable<Void>() {
+					@Override
+					public Void call() throws Exception {
+						Thread.sleep(boundSeconds * 2 * 1000L);
+						holder.rollback(); // the rows go back exactly as they were
+						return null;
+					}
+				});
+
+				final AtomicInteger attempts = new AtomicInteger();
+				final long startedAt = System.nanoTime();
+				storage.write(new WriteOperation() {
+					@Override
+					public void run(WriteableTransaction txn) throws Exception {
+						attempts.incrementAndGet();
+						txn.put(tree, key(1), value(2));
+					}
+				});
+				final long elapsedMillis = (System.nanoTime() - startedAt) / 1000000L;
+				released.get(); // the failure of the holder, if it had one, rather than a stuck rollback
+
+				if (dialect() == JDBCStorage.Dialect.ORACLE) {
+					assertEquals(attempts.get(), 1, "oracle has no session setting for a row lock enqueue, so the"
+						+ " write should have waited the holder out on its first attempt (" + elapsedMillis + " ms)");
+				} else {
+					assertTrue(attempts.get() >= 2, "the write was not replayed: it waited " + elapsedMillis
+						+ " ms for a row lock that this engine was told to give up on after " + boundSeconds + " s");
+					assertTrue(elapsedMillis >= boundSeconds * 1000L - JDBCStorage.CLOCK_SLACK_MILLIS,
+						"the write came back after " + elapsedMillis + " ms, before the bound it was given:"
+						+ " something other than the row lock ended its first attempt");
+				}
+			}
+		} finally {
+			System.clearProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY);
+			releasing.shutdownNow();
+			try {
+				storage.write(new WriteOperation() {
+					@Override
+					public void run(WriteableTransaction txn) throws Exception {
+						txn.deleteTree(tree);
+					}
+				});
+			} catch (Exception ignored) {
+			} finally {
+				storage.close();
+			}
+		}
+	}
+
 	private interface BlockedOperation {
 		void run(JDBCStorage storage, TreeName tree) throws Exception;
 	}
@@ -807,6 +900,11 @@
 					for (final JDBCStorage.StatementBound each : JDBCStorage.StatementBound.values()) {
 						System.setProperty(each.property, each == bound ? Integer.toString(boundSeconds) : "0");
 					}
+					// and the row lock bound of a write is off here for the same reason: it is tighter
+					// than the bound under test, so it - and not the statement bound this case is about -
+					// would be what ended the wait, and the failure would name no property of this class
+					// (#915). testAWriteBlockedByARowLockIsReplayedInsideTheWindow covers that layer
+					System.setProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY, "0");
 					// the monotonic clock, which is what timedOut() measures the bound with: a step of
 					// the wall clock can neither lengthen nor shorten what the assertions below allow
 					final long startedAt = System.nanoTime();
@@ -868,6 +966,7 @@
 			for (final JDBCStorage.StatementBound each : JDBCStorage.StatementBound.values()) {
 				System.clearProperty(each.property);
 			}
+			System.clearProperty(JDBCStorage.ROW_LOCK_TIMEOUT_PROPERTY);
 			try {
 				storage.write(new WriteOperation() {
 					@Override

--
Gitblit v1.10.0