+ * 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. */ - private static final long MAX_RETRY_WINDOW_NANOS = 10L * 1000L * 1000L * 1000L; //10 s + private 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; @@ -75,15 +86,30 @@ public class JDBCStorage implements org.opends.server.backends.pluggable.spi.Sto private static final double MAX_SLEEP_ON_RETRY_MS = 1000.0; /** - * Number of links walked when classifying a failure, also a guard against a chain long enough to matter. One - * number for three chains at once - the causes, the next exceptions and the suppressed exceptions are walked - * together and counted together - so it is set well above the depth a wrapped failure of this backend reaches: - * mssql-jdbc chains every error of one message it received through {@code setNextException}, and a budget spent - * on those would never reach the cause the wrapper carries. + * Number of links walked by the questions that are not asked to the end of the chains: a guard against a chain + * long enough to matter, at the cost of what truncation costs each of them. One number for three chains at + * once - the causes, the next exceptions and the suppressed exceptions are walked together and counted + * together - so it is set well above the depth a wrapped failure of this backend reaches: mssql-jdbc chains + * every error of one message it received through {@code setNextException}, and a budget spent on those would + * never reach the cause the wrapper carries. + *
+ * For the fallbacks of {@link #conflictSummary} truncation leaves a question unanswered and nothing more: the
+ * line reporting the replay names a less precise link, and no decision moves. For
+ * {@link #isConnectionFailure} it does weaken the verdict, which is the test {@link #EVERY_LINK} exists to
+ * apply: a class 08 link past this many links leaves {@code dropped} false in {@link #write}, so
+ * {@link #distrustPool} is not called and the pool keeps handing out - unvalidated - the connections it had
+ * established before the same restart or failover. That is what this walk did before #903 and it is left as it
+ * is here rather than widened along with the two below, since nothing about the grant of #903 depends on it;
+ * the budget is pinned from both sides by {@code testTheWalkOfAFailureStopsAtItsBudget}, which is where
+ * widening it would have to start.
*/
private static final int MAX_CHAIN_LINKS = 64;
- /** The budget of {@link #failureScope}, which walks to the end of the chains: see the comment above it. */
+ /**
+ * The budget of the two walks whose verdict would weaken rather than go unnoticed under truncation:
+ * {@link #failureScope} and {@link #conflictVerdict}. See the comments above them; the {@code seen} set of
+ * the walk terminates it either way.
+ */
private static final int EVERY_LINK = Integer.MAX_VALUE;
/** SQL Server error number of the transaction picked as the deadlock victim: "Rerun the transaction". */
@@ -92,6 +118,15 @@ public class JDBCStorage implements org.opends.server.backends.pluggable.spi.Sto
/** Oracle error number of a detected deadlock: ORA-00060, reported with SQLState 61000 rather than class 40. */
private static final int ORACLE_DEADLOCK_DETECTED = 60;
+ /**
+ * MySQL error number of a lock wait timeout, ER_LOCK_WAIT_TIMEOUT: the one conflict of the set an engine
+ * reports late rather than promptly. Connector/J maps it to the same class 40 state as a deadlock, so this
+ * number is not what matches it - it only tells the two apart, and only under a MySQL driver. That it repeats
+ * the literal of {@link #MSSQL_DEADLOCK_VICTIM} is the collision {@link #conflictVerdict} keys every vendor
+ * number by the driver for, and is stated there rather than again here.
+ */
+ private static final int MYSQL_LOCK_WAIT_TIMEOUT = 1205;
+
/**
* Class 40 states that are transaction rollbacks but must not be replayed. 40003 leaves the outcome of the
* transaction unknown, so replaying an add that in fact committed would answer the client with
@@ -537,6 +572,11 @@ static int clampSeconds(int seconds) {
// below turns on a few milliseconds either side of the bound, and a mock statement cannot be made
// to take a real second without the suite taking one too. Monotonic, so that a step of the wall
// clock can neither lengthen nor shorten what a statement is measured to have taken.
+ //
+ // The retry window of write() is read off this same clock, once before the first attempt and once
+ // after each: what that window bounds is the whole run of attempts rather than each attempt on its
+ // own, which is a single startedAt outside the loop. #877 and #903 each added a clock of their own
+ // here, for the same reason and with the same body; this is the one they share.
long nanoTime() {
return System.nanoTime();
}
@@ -1267,14 +1307,26 @@ static String driverNameOf(Connection con) {
// The dialect behind a pooled connection, or null for an engine none of the statements of this
// class fit: it is left unstamped and its statistics untouched rather than fed untested SQL.
static Dialect dialectOf(Connection con) {
- final String driverName=driverNameOf(con);
- if (driverName.contains("postgres")) {
+ return dialectOf(driverNameOf(con));
+ }
+
+ /**
+ * The engine a driver class name names, or null for one this class does not recognise. Every question this
+ * class asks about the engine is keyed on this one answer - the column types, the upsert, the paging clause,
+ * whether a DDL statement commits, and the class of a conflict - so that they cannot disagree about a driver.
+ * A cascade of its own per question is how a deployment ends up given one engine's SQL and another engine's
+ * conflict class; what an unrecognised engine gets is a decision per question, taken and stated at each of
+ * them rather than falling out of the order the {@code contains} calls happen to be written in.
+ */
+ static Dialect dialectOf(String driverName) {
+ final String name=String.valueOf(driverName);
+ if (name.contains("postgres")) {
return Dialect.POSTGRES;
- }else if (driverName.contains("mysql")) {
+ }else if (name.contains("mysql")) {
return Dialect.MYSQL;
- }else if (driverName.contains("oracle")) {
+ }else if (name.contains("oracle")) {
return Dialect.ORACLE;
- }else if (driverName.contains("microsoft")) {
+ }else if (name.contains("microsoft")) {
return Dialect.MICROSOFT;
}
return null;
@@ -1654,9 +1706,12 @@ final class CatalogSession implements Closeable {
private WriteableTransactionTransactionImpl txn;
// The moment the replay window of the write() this session belongs to runs out, as
- // System.nanoTime() reads it - null where nothing above this session replays, which is the
- // importer and nothing else. Boxed rather than given a sentinel: nanoTime() is documented to
- // return an arbitrary long, so there is no reading of it that could stand for "no window".
+ // JDBCStorage.nanoTime() reads it - null where nothing above this session replays, which is the
+ // importer and nothing else. That clock and not System.nanoTime() directly: the window this is a
+ // reading of was taken from it, and the two are the same clock everywhere except the one place
+ // they would be compared as a mixed pair. Boxed rather than given a sentinel: nanoTime() is
+ // documented to return an arbitrary long, so there is no reading of it that could stand for
+ // "no window".
private Long replayWindowEndsAt;
/**
@@ -1683,7 +1738,7 @@ private long budgetDeadline() {
if (replayWindowEndsAt==null) {
return Long.MAX_VALUE;
}
- final long leftNanos=replayWindowEndsAt-System.nanoTime();
+ final long leftNanos=replayWindowEndsAt-nanoTime();
final long now=System.currentTimeMillis();
return leftNanos<=0 ? now : now+leftNanos/1_000_000L;
}
@@ -2944,10 +2999,15 @@ public
* 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
@@ -2958,11 +3018,11 @@ public
* A transaction conflict is replayable whichever phase reported it: the engine rolled the transaction back
* before it answered. It is read from the failure of the operation only, never from the release of the
- * connection - see {@link #isRetryableConflict} - since the release runs after the outcome was decided and
- * cannot make that claim for it. A connection the database dropped is replayable only while the transaction
- * had not been committed yet. A drop reported by {@code commit()} leaves the outcome unknown - the server may
+ * connection - see {@link #conflictVerdict} - since the release runs after the outcome was decided and cannot
+ * make that claim for it. A connection the database dropped is replayable only while the transaction had not been
+ * committed yet. A drop reported by {@code commit()} leaves the outcome unknown - the server may
* have committed and died before the answer reached us - and replaying a write that in fact committed applies
* it twice, which is the very reason 40003 is one of {@link #NON_REPLAYABLE_ROLLBACK_STATES}.
*
@@ -3101,17 +3192,18 @@ public void write(WriteOperation writeOperation) throws Exception {
* second time and fails with ERR_ENTRY_CONTAINER_ALREADY_REGISTERED, which masks the failure that caused the
* replay and leaves the indexes of the previous attempt behind with their configuration listeners.
*
+ * @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
*/
- static String replayReason(Throwable failure, String driver, boolean committing, boolean partlyCommitted,
+ static String replayReason(Conflict conflict, Throwable failure, boolean committing, boolean partlyCommitted,
boolean connectionClosed) {
if (partlyCommitted) {
return null;
}
- if (isRetryableConflict(failure, driver)) {
+ if (conflict!=Conflict.NONE) {
return "a conflict";
}
if (!committing && (connectionClosed || isConnectionFailure(failure))) {
@@ -3184,6 +3276,27 @@ private static SQLException firstLinkMatching(Throwable failure, boolean withThe
/** The walk above, with the number of links it is allowed to look at. */
private static SQLException firstLinkMatching(Throwable failure, boolean withTheRelease, int links,
Predicate
+ * 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.
+ */
+ enum Conflict {
+ /** Not a conflict: no replay resolves it. */
+ NONE,
+ /** A conflict reported as soon as the engine detects it, however long the attempt waited to reach it. */
+ PROMPT,
+ /**
+ * A conflict under a driver none of the four engines is recognised in. Whether the engine bounded the
+ * wait that preceded it is not something this class can tell, and the grant of {@link #PROMPT} rests on
+ * knowing that it did not, so an unrecognised engine is refused it: see {@link #classOf}.
+ */
+ UNKNOWN_ENGINE,
+ /** A conflict an engine reports only once a lock wait timeout of its own has elapsed. */
+ AFTER_LOCK_WAIT
+ }
+
+ /**
+ * The verdict of a failure nothing asked about, handed to the questions {@link #write} asks of a partly
+ * committed attempt: every one of them is answered by that flag alone, so its chains are never walked. It is
+ * not a claim that the failure carries no conflict - it may carry one, and is refused a replay either way.
+ */
+ private static final ConflictVerdict NOT_CLASSIFIED=new ConflictVerdict(Conflict.NONE, null);
+
+ /**
+ * The class of the conflict a failure carries and the link that class was read from, which are one answer
+ * rather than two: the line reporting a replay names the link the decision was taken on, and a summary that
+ * walked the chains again to find it could name a different one - see {@link #conflictSummary}.
+ */
+ static final class ConflictVerdict {
+ final Conflict conflict;
+ /** Null where the failure carries no conflict at all, which is what {@link Conflict#NONE} says. */
+ final SQLException link;
+
+ ConflictVerdict(Conflict conflict, SQLException link) {
+ this.conflict=conflict;
+ this.link=link;
+ }
+ }
+
+ /**
+ * Returns the class of the conflict the given failure carries, or {@link Conflict#NONE} if it carries none -
+ * which is what decides whether replaying the operation can resolve it - together with the link that class was
+ * read from. One walk of the chains that keeps the strongest class it meets, rather than one walk per class
+ * asked in the right order: asking per class is what let the two walks of the earlier form be given different
+ * budgets, and the ordering of an added class is then a rule its author has to find rather than one the enum
+ * states - see {@link Conflict}.
*
* The conflict is looked up along every chain of the failure, for the reason {@link #isConnectionFailure} walks
* them all: it reaches this class wrapped - a deadlock in {@code put} arrives as
* {@code StorageRuntimeException(SQLException)}, and a caller such as {@code EntryContainer.addEntry} may wrap it
* once more - and a driver reports the error that says what happened as the next exception of a generic one at
- * least as often as it reports it as the cause.
+ * least as often as it reports it as the cause. The suppressed links of the release are left out of it, for the
+ * reason {@link #replayReason} gives: the release runs after the outcome was decided.
+ *
+ * The strongest class in those chains wins rather than the first one found: a wrapper that carries a class
+ * 40 state of its own but no vendor number would otherwise downgrade the {@link Conflict#AFTER_LOCK_WAIT} of
+ * the {@link SQLException} it wraps, and hand a wait the engine already bounded a replay it does not need.
+ * That rule is deliberately not restricted to the wrapper it was introduced for, although the walk reaches
+ * links that are not ancestors of the operative failure - a deadlock whose chain also carries a lock wait
+ * timeout is classed by the timeout and loses the grant. The two errors are not equally costly to get wrong:
+ * granting a replay to a wait the engine had already bounded pays that bound a second time, while refusing
+ * one to a deadlock costs a replay the window was about to refuse anyway, wherever the bound that sibling
+ * names is longer than the window. So the class is read the conservative way, and the whole chain of a
+ * failure is evidence for it.
+ *
+ * Every link is looked at, rather than {@link #MAX_CHAIN_LINKS} of them, for the reason {@link #failureScope}
+ * walks to the end: the verdict weakens under truncation rather than simply going unnoticed. An
+ * {@link Conflict#AFTER_LOCK_WAIT} link past the budget with a bare class 40 link inside it comes back
+ * {@link Conflict#PROMPT}, and truncation there does not lose a replay - it grants one, which is the single
+ * thing this classification exists to refuse. The {@code seen} set terminates the walk regardless.
*
* The standard class 40 states carry the conflict of most engines - 40P01 for PostgreSQL, 40001 for SQL Server
* and for MySQL, whose driver replaces the server side HY000 of a deadlock and of a lock wait timeout with
@@ -3266,28 +3449,135 @@ static long retryDelayMillis(int attempt) {
* reports a deadlock as ORA-00060 with SQLState 61000, and gives 1205 to a fatal "not a data file" error that
* no replay can resolve, while 1205 is exactly the deadlock victim of SQL Server. The SQL Server number is
* matched beyond its class 40 state because a deployment may add {@code xopenStates=true} to its connection
- * URL, which reports the same deadlock as 42000. MySQL needs no number of its own, since its driver has already
- * mapped both conditions into class 40; see {@link #NON_REPLAYABLE_ROLLBACK_STATES} for the two class 40 states
- * that are excluded from that match.
+ * URL, which reports the same deadlock as 42000. MySQL needs no number of its own for the match, since its
+ * driver has already mapped both conditions into class 40 - its number is read by {@link #classOf} alone, and
+ * only to tell the two apart; see {@link #NON_REPLAYABLE_ROLLBACK_STATES} for the two class 40 states that are
+ * excluded from that match.
+ *
+ * The walk still stops as soon as its answer is final, but the class it stops at is the strongest one this
+ * engine can report - {@link #ceilingOf} - rather than the strongest one the enum declares. Only MySQL reports
+ * an {@link Conflict#AFTER_LOCK_WAIT}, so a walk stopping at that constant never stops early on the other
+ * three engines, nor under a driver none of them is recognised in: it reads every link of every failed write,
+ * a plain {@code 23000} from adding an entry that is already there included, on the driver whose chains are
+ * longest. The dialect is resolved once here for the same reason - {@link #classOf} and {@link #isConflict}
+ * would otherwise read it off the driver name twice for every link walked.
+ */
+ static ConflictVerdict conflictVerdict(Throwable failure, String driver) {
+ final Dialect dialect=dialectOf(driver);
+ final Conflict ceiling=ceilingOf(dialect);
+ final Conflict[] strongest={Conflict.NONE};
+ final SQLException[] link=new SQLException[1];
+ walkLinks(failure, WITHOUT_THE_RELEASE, EVERY_LINK, e -> {
+ final Conflict conflict=classOf(e, dialect);
+ if (conflict.compareTo(strongest[0])>0) {
+ strongest[0]=conflict;
+ link[0]=e;
+ }
+ return strongest[0]==ceiling;
+ });
+ return new ConflictVerdict(strongest[0], link[0]);
+ }
+
+ /**
+ * 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
+ * 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.
*/
- static boolean isRetryableConflict(Throwable t, String driver) {
- // without the suppressed exceptions, unlike isConnectionFailure(): a conflict is replayed whichever phase
- // reported it, on the strength of the engine having rolled the transaction back before it answered - and
- // the release of the connection runs after the outcome was decided and cannot make that claim. A class 40
- // raised there would otherwise replay a transaction commit() left in doubt, which is what the committing
- // guard of replayReason() exists to prevent
- return firstLinkMatching(t, WITHOUT_THE_RELEASE, e -> isConflict(e, driver))!=null;
+ static Conflict ceilingOf(Dialect dialect) {
+ if (dialect==null) {
+ return Conflict.UNKNOWN_ENGINE;
+ }
+ return dialect==Dialect.MYSQL ? Conflict.AFTER_LOCK_WAIT : Conflict.PROMPT;
+ }
+
+ /**
+ * Returns the class of a single failure. The vendor number only refines a failure {@link #isConflict} has
+ * already matched and never widens that match, which the engines colliding on 1205 do not allow: the number is
+ * read here to tell the late conflict of MySQL from the deadlock its driver reports under the same state.
+ *
+ * A conflict raised under a driver {@link #dialectOf(String)} did not recognise is
+ * {@link Conflict#UNKNOWN_ENGINE}: still replayed, since replayability is what the class 40 state says and it
+ * says it whatever the engine, but not granted the replay past the window. The grant rests on knowing that
+ * the wait preceding the conflict was not bounded by the engine, and of an unrecognised engine that is not
+ * known. It is a MySQL-wire-compatible driver - MariaDB Connector/J, an Aurora- or Percona-branded one -
+ * that makes the difference concrete: it reports a lock wait timeout as 1205 under class 40 exactly as
+ * Connector/J does, this class would read the number only under a name carrying {@code mysql}, and granting
+ * a free replay there buys a second full {@code innodb_lock_wait_timeout}. Such a deployment does reach this
+ * code: a backend created under {@code com.mysql.cj.jdbc} and later opened through one of those drivers
+ * issues no DDL at all - every {@code create table} and {@code create index} of
+ * {@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.
+ */
+ private static Conflict classOf(SQLException e, Dialect dialect) {
+ if (!isConflict(e, dialect)) {
+ return Conflict.NONE;
+ }
+ if (dialect==null) {
+ return Conflict.UNKNOWN_ENGINE;
+ }
+ return dialect==Dialect.MYSQL && e.getErrorCode()==MYSQL_LOCK_WAIT_TIMEOUT
+ ? Conflict.AFTER_LOCK_WAIT : Conflict.PROMPT;
}
- private static boolean isConflict(SQLException e, String driver) {
+ /**
+ * Whether another attempt is still allowed: the bounds half of the decision {@link #write} takes after every
+ * attempt, asked of a failure {@link #replayReason} has already found worth replaying and made here apart
+ * from the clock so that it can be tested without a database. It is asked of the conflict class rather than
+ * of the failure because not every replayable failure carries one - a connection the database dropped is
+ * replayed on the evidence of the drop, and would be refused by a bound that first insisted on a class 40
+ * state - and because {@code write()} has already read that class off the failure once.
+ *
+ * Replays are bounded by {@link #MAX_RETRIES} and by {@link #RETRY_WINDOW_NANOS} against the time elapsed
+ * since the first attempt began, with the one grant {@link #grantedPastTheWindow} states on top of them.
+ */
+ static boolean replayableWithin(int attempt, long elapsedNanos, Conflict conflict) {
+ if (attempt>=MAX_RETRIES) {
+ return false;
+ }
+ if (grantedPastTheWindow(attempt, elapsedNanos, conflict)) {
+ return true;
+ }
+ return elapsedNanos
+ * {@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.
+ */
+ static boolean grantedPastTheWindow(int attempt, long elapsedNanos, Conflict conflict) {
+ return attempt==1 && conflict==Conflict.PROMPT && elapsedNanos>=RETRY_WINDOW_NANOS;
+ }
+
+ private static boolean isConflict(SQLException e, Dialect dialect) {
final String state=String.valueOf(e.getSQLState());
if (state.startsWith("40") && !NON_REPLAYABLE_ROLLBACK_STATES.contains(state)) {
return true;
}
- final String driverName=String.valueOf(driver);
- if (driverName.contains("oracle")) {
+ if (dialect==Dialect.ORACLE) {
return e.getErrorCode()==ORACLE_DEADLOCK_DETECTED;
- } else if (driverName.contains("microsoft")) {
+ } else if (dialect==Dialect.MICROSOFT) {
return e.getErrorCode()==MSSQL_DEADLOCK_VICTIM;
}
return false;
@@ -3301,11 +3591,16 @@ private static boolean isConflict(SQLException e, String driver) {
* suppressed into it, and naming the state of the rejected statement instead would describe a replay that did
* not happen. Falls back to the first SQLException of the failure, and to the failure itself where it carries
* none.
+ *
+ * Asked of the verdict rather than of the failure and the driver, since {@link #write} - the only caller - has
+ * had the failure classified already: a form taking those two would walk the chains a second time to reach the
+ * verdict this one is handed.
*/
- static String conflictSummary(Throwable failure, String driver) {
- // asked in the order replayReason() asks it, and of the same chains, so that the line names the link the
- // decision was taken on rather than one that merely resembles it
- SQLException named=firstLinkMatching(failure, WITHOUT_THE_RELEASE, e -> isConflict(e, driver));
+ static String conflictSummary(ConflictVerdict verdict, Throwable failure) {
+ // the link the class was read from, handed over by the walk that read it rather than looked up again in
+ // the order that walk happens to use: repeated by hand, the two drift, and the line then names a link
+ // that merely resembles the one the decision was taken on
+ SQLException named=verdict.link;
if (named==null) {
named=firstLinkMatching(failure, WITH_THE_RELEASE, JDBCStorage::saysTheConnectionIsGone);
}
@@ -3359,7 +3654,7 @@ static byte[] db2real(byte[] db) {
* Casting the parameter back to char keeps the comparison seekable.
*/
static String hashParam(Connection con) {
- return driverNameOf(con).contains("microsoft") ? "cast(? as char(128))" : "?";
+ return dialectOf(con)==Dialect.MICROSOFT ? "cast(? as char(128))" : "?";
}
class ReadableTransactionImpl implements ReadableTransaction {
@@ -3570,19 +3865,20 @@ private void commitStatement(String sql, boolean ddl) throws SQLException {
/** Whether this engine commits the transaction before a DDL statement whether asked to or not. */
private boolean commitsBeforeDdl() {
- final String driverName=driverNameOf(con);
- return driverName.contains("mysql") || driverName.contains("oracle");
+ final Dialect dialect=dialectOf(con);
+ return dialect==Dialect.MYSQL || dialect==Dialect.ORACLE;
}
String getTableDialect() {
- if (driverNameOf(con).contains("oracle")) {
+ final Dialect dialect=dialectOf(con);
+ if (dialect==Dialect.ORACLE) {
return "h char(128),k raw(2000),v blob,primary key(h,k)";
- }else if (driverNameOf(con).contains("mysql")) {
+ }else if (dialect==Dialect.MYSQL) {
return "h char(128),k varbinary(255),v longblob,primary key(h,k)";
- }else if (driverNameOf(con).contains("microsoft")) {
+ }else if (dialect==Dialect.MICROSOFT) {
return "h char(128),k varbinary(max),v image,primary key(h)";
}
- return "h char(128),k bytea,v bytea,primary key(h,k)";
+ return "h char(128),k bytea,v bytea,primary key(h,k)"; // postgres, and an unrecognised engine with it
}
@Override
@@ -3616,9 +3912,9 @@ public void openTree(TreeName treeName, boolean createOnDemand) {
}
}
// CursorImpl iterates with "where k>? order by k" batches: primary key (h,k) cannot serve them
- final String driverName=driverNameOf(con);
+ final Dialect dialect=dialectOf(con);
final String tableName=getTableName(treeName);
- if (driverName.contains("postgres")) {
+ if (dialect==Dialect.POSTGRES) {
try {
// asked although postgresql has "create index if not exists": that statement commits
// whether it creates anything or not, and this is the engine of every default
@@ -3630,7 +3926,7 @@ public void openTree(TreeName treeName, boolean createOnDemand) {
}catch (SQLException e) {
throw new StorageRuntimeException(e);
}
- }else if (driverName.contains("mysql")) {
+ }else if (dialect==Dialect.MYSQL) {
try {
if (!isExistsIndex(tableName,"k_"+tableName.substring("opendj_".length()))) { // mysql has no "create index if not exists"
commitStatement("create index k_"+tableName.substring("opendj_".length())+" on "+tableName+" (k)", true);
@@ -3638,7 +3934,7 @@ public void openTree(TreeName treeName, boolean createOnDemand) {
}catch (SQLException e) {
throw new StorageRuntimeException(e);
}
- }else if (driverName.contains("oracle")) {
+ }else if (dialect==Dialect.ORACLE) {
try {
// oracle has no "create index if not exists"; unquoted identifiers are stored in uppercase
if (!isExistsIndex(tableName.toUpperCase(Locale.ROOT),"k_"+tableName.substring("opendj_".length()))) {
@@ -4065,29 +4361,29 @@ public void put(TreeName treeName, ByteSequence key, ByteSequence value) {
}
boolean upsert(TreeName treeName, ByteSequence key, ByteSequence value) throws SQLException {
- final String driverName=driverNameOf(con);
- if (driverName.contains("postgres")) { //postgres upsert
+ final Dialect dialect=dialectOf(con);
+ if (dialect==Dialect.POSTGRES) { //postgres upsert
try (final PreparedStatement statement = con.prepareStatement("insert into " + getTableName(treeName) + " (h,k,v) values (?,?,?) ON CONFLICT (h, k) DO UPDATE set v=excluded.v")) {
statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray())));
statement.setBytes(2, real2db(key.toByteArray()));
statement.setBytes(3, value.toByteArray());
return (execute(statement, bound) == 1 && statement.getUpdateCount() > 0);
}
- }else if (driverName.contains("mysql")) { //mysql upsert
+ }else if (dialect==Dialect.MYSQL) { //mysql upsert
try (final PreparedStatement statement = con.prepareStatement("insert into " + getTableName(treeName) + " (h,k,v) values (?,?,?) as new ON DUPLICATE KEY UPDATE v=new.v")) {
statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray())));
statement.setBytes(2, real2db(key.toByteArray()));
statement.setBytes(3, value.toByteArray());
return (execute(statement, bound) == 1 && statement.getUpdateCount() > 0);
}
- }else if (driverName.contains("oracle")) { //ANSI MERGE without ;
+ }else if (dialect==Dialect.ORACLE) { //ANSI MERGE without ;
try (final PreparedStatement statement = con.prepareStatement("merge into " + getTableName(treeName) + " old using (select ? h,? k,? v from dual) new on (old.h=new.h and old.k=new.k) WHEN MATCHED THEN UPDATE SET old.v=new.v WHEN NOT MATCHED THEN INSERT (h,k,v) VALUES (new.h,new.k,new.v)")) {
statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray())));
statement.setBytes(2, real2db(key.toByteArray()));
statement.setBytes(3, value.toByteArray());
return (execute(statement, bound) == 1 && statement.getUpdateCount() > 0);
}
- }else if (driverName.contains("microsoft")) { //ANSI MERGE with ; WITH (HOLDLOCK) makes the upsert atomic: without it SQL Server MERGE can race two concurrent NOT MATCHED inserts of the same key into a PRIMARY KEY violation. UPDLOCK is required on top of it: with HOLDLOCK alone the search phase takes a shared lock that the WHEN MATCHED update then has to convert to an exclusive one, so two concurrent upserts of the same key deadlock on the conversion; an update lock is taken right away and makes the second transaction wait instead. h is cast back to char so that the join can seek the primary key instead of scanning the whole table under those locks, see hashParam()
+ }else if (dialect==Dialect.MICROSOFT) { //ANSI MERGE with ; WITH (HOLDLOCK) makes the upsert atomic: without it SQL Server MERGE can race two concurrent NOT MATCHED inserts of the same key into a PRIMARY KEY violation. UPDLOCK is required on top of it: with HOLDLOCK alone the search phase takes a shared lock that the WHEN MATCHED update then has to convert to an exclusive one, so two concurrent upserts of the same key deadlock on the conversion; an update lock is taken right away and makes the second transaction wait instead. h is cast back to char so that the join can seek the primary key instead of scanning the whole table under those locks, see hashParam()
try (final PreparedStatement statement = con.prepareStatement("merge into " + getTableName(treeName) + " WITH (HOLDLOCK, UPDLOCK) old using (select cast(? as char(128)) h,? k,? v) new on (old.h=new.h and old.k=new.k) WHEN MATCHED THEN UPDATE SET old.v=new.v WHEN NOT MATCHED THEN INSERT (h,k,v) VALUES (new.h,new.k,new.v);")) {
statement.setString(1, key2hash.get(ByteBuffer.wrap(key.toByteArray())));
statement.setBytes(2, real2db(key.toByteArray()));
@@ -4204,7 +4500,7 @@ public CursorImpl(boolean isReadOnly, Connection con, TreeName treeName, Stateme
// of #873 reads the shared tree, and reading a tree must not put it up for removal
this.tableName=readTableName(treeName);
this.batchBound=batchBound;
- this.limitClause=((CachedConnection)con).parent.getClass().getName().contains("mysql")
+ this.limitClause=dialectOf(con)==Dialect.MYSQL
? " limit ?,?" : " offset ? rows fetch next ? rows only";
}
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 2c339a9ceb..0aff8f7d3f 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
@@ -17,10 +17,12 @@
import org.forgerock.opendj.server.config.server.JDBCBackendCfg;
import org.opends.server.DirectoryServerTestCase;
+import org.opends.server.backends.jdbc.JDBCStorage.Conflict;
import org.opends.server.backends.pluggable.spi.AccessMode;
import org.opends.server.backends.pluggable.spi.StorageRuntimeException;
import org.opends.server.backends.pluggable.spi.TreeName;
import org.opends.server.backends.pluggable.spi.WriteOperation;
+import org.opends.server.backends.pluggable.spi.WriteableTransaction;
import org.opends.server.types.DirectoryException;
import org.testng.annotations.AfterClass;
import org.testng.annotations.BeforeClass;
@@ -40,6 +42,7 @@
import java.sql.Statement;
import java.util.Properties;
import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicBoolean;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicReference;
import java.util.function.Predicate;
@@ -60,16 +63,22 @@
import static org.mockito.Mockito.times;
import static org.mockito.Mockito.verify;
import static org.mockito.Mockito.when;
+import static org.opends.server.backends.jdbc.JDBCStorage.Conflict.AFTER_LOCK_WAIT;
+import static org.opends.server.backends.jdbc.JDBCStorage.Conflict.NONE;
+import static org.opends.server.backends.jdbc.JDBCStorage.Conflict.PROMPT;
+import static org.opends.server.backends.jdbc.JDBCStorage.Conflict.UNKNOWN_ENGINE;
import static org.testng.Assert.assertEquals;
import static org.testng.Assert.assertFalse;
import static org.testng.Assert.assertNull;
+import static org.testng.Assert.assertSame;
import static org.testng.Assert.assertTrue;
import static org.testng.Assert.fail;
/**
- * Tests how a failure is classified - as a transaction conflict, or as a connection the database dropped - which
- * is what decides whether {@link JDBCStorage#write} replays the operation, and how long it waits before it does;
- * and what {@code write()} itself does with that verdict, replay and pool alike.
+ * Tests how a failure is classified - as a transaction conflict of one class or another, or as a connection the
+ * database dropped - which is what decides whether {@link JDBCStorage#write} replays the operation, whether its
+ * first replay is granted regardless of the clock, how long the replays may go on for and how long it waits
+ * before each of them; and what {@code write()} itself does with that verdict, replay and pool alike.
*
* Runs without a database: the failures the drivers report are reproduced as synthetic
* {@link SQLException}s carrying the same vendor error number and SQLState, and the writes that carry them run
@@ -84,6 +93,8 @@ public class JDBCStorageRetryTest extends DirectoryServerTestCase
private static final String MYSQL = "com.mysql.cj.jdbc.ConnectionImpl";
private static final String ORACLE = "oracle.jdbc.driver.T4CConnection";
private static final String POSTGRES = "org.postgresql.jdbc.PgConnection";
+ /** A MySQL-wire-compatible driver, whose class name carries no engine this backend recognises. */
+ private static final String MARIADB = "org.mariadb.jdbc.Connection";
/** 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");
@@ -121,6 +132,16 @@ interface mysqlConnection 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,
+ * and a backend created under {@code com.mysql.cj.jdbc} opens through it: every DDL of {@code openTree()} is
+ * guarded by a catalog read, so an existing backend issues none of it.
+ */
+ interface mariadbConnection extends Connection
+ {
+ }
+
/** A failure whose cause chain is a cycle, to check that walking it terminates. */
private static final class SelfCausedException extends RuntimeException
{
@@ -133,62 +154,192 @@ public synchronized Throwable getCause()
}
}
+ /**
+ * The failures the engines report, and the class each of them belongs to: {@link Conflict#NONE} for a failure no
+ * replay resolves, and otherwise how promptly the engine reporting it does so, which is all the class decides.
+ */
@DataProvider
public Object[][] failures()
{
return new Object[][] {
- // SQL Server picking a transaction as the deadlock victim: the failure this retry exists for
- { "mssql deadlock victim", sql(1205, "40001"), MSSQL, true },
+ // SQL Server picking a transaction as the deadlock victim: the failure this retry exists for. It is reported
+ // as promptly as the deadlock monitor runs, but only once the victim has waited out a lock wait of its own,
+ // which SQL Server leaves unbounded - the wait belongs to the attempt, not to the reporting
+ { "mssql deadlock victim", sql(1205, "40001"), MSSQL, PROMPT },
// a deployment may add xopenStates=true to its connection URL, which reports the same deadlock as 42000
- { "mssql deadlock victim, xopenStates", sql(1205, "42000"), MSSQL, true },
+ { "mssql deadlock victim, xopenStates", sql(1205, "42000"), MSSQL, PROMPT },
// the conflict of most other engines is carried by the SQLState, under a vendor number of their own
- { "postgres serialization failure", sql(0, "40001"), POSTGRES, true },
- { "postgres deadlock detected", sql(0, "40P01"), POSTGRES, true },
+ { "postgres serialization failure", sql(0, "40001"), POSTGRES, PROMPT },
+ { "postgres deadlock detected", sql(0, "40P01"), POSTGRES, PROMPT },
// Connector/J replaces the server side HY000 of both conditions with 40001, so neither needs a number here
- { "mysql deadlock", sql(1213, "40001"), MYSQL, true },
- // not a deadlock, but transient in the same way and equally resolved by a replay
- { "mysql lock wait timeout", sql(1205, "40001"), MYSQL, true },
+ { "mysql deadlock", sql(1213, "40001"), MYSQL, PROMPT },
+ // not a deadlock, but transient in the same way and equally resolved by a replay - and the one conflict of
+ // them all that an engine reports only after a lock wait timeout of its own, innodb_lock_wait_timeout
+ { "mysql lock wait timeout", sql(1205, "40001"), MYSQL, AFTER_LOCK_WAIT },
// the rollback a MySQL group replication conflict reports, error 3101, which the driver maps to 40000
- { "mysql group replication rollback", sql(3101, "40000"), MYSQL, true },
+ { "mysql group replication rollback", sql(3101, "40000"), MYSQL, PROMPT },
// Oracle maps ORA-00060 to SQLState 61000, so only its error number identifies the deadlock
- { "oracle deadlock detected", sql(60, "61000"), ORACLE, true },
+ { "oracle deadlock detected", sql(60, "61000"), ORACLE, PROMPT },
// the conflict reaches JDBCStorage.write() wrapped, so the whole cause chain has to be walked
- { "wrapped once", new StorageRuntimeException(sql(1205, "40001")), MSSQL, true },
+ { "wrapped once", new StorageRuntimeException(sql(1205, "40001")), MSSQL, PROMPT },
{ "wrapped twice",
new DirectoryException(OTHER, raw("unchecked"), new StorageRuntimeException(sql(1205, "40001"))), MSSQL,
- true },
+ PROMPT },
// the vendor numbers collide across engines, so they must not be matched driver-independently:
// ORA-01205 "not a data file" is fatal, and no replay resolves it
- { "oracle not a data file", sql(1205, "64000"), ORACLE, false },
+ { "oracle not a data file", sql(1205, "64000"), ORACLE, NONE },
// and a lock wait timeout is a MySQL number: 1205 means nothing of the kind to PostgreSQL
- { "postgres unrelated 1205", sql(1205, "22001"), POSTGRES, false },
+ { "postgres unrelated 1205", sql(1205, "22001"), POSTGRES, NONE },
// two class 40 states are rollbacks that a replay must not repeat: 40003 leaves the outcome of the
// transaction unknown, and 40002 is an integrity constraint violation that a replay would only hit again
- { "statement completion unknown", sql(0, "40003"), POSTGRES, false },
- { "transaction integrity constraint violation", sql(0, "40002"), POSTGRES, false },
- // ... but the state of a conflict is still matched whatever vendor number carries it
- { "class 40 is driver independent", sql(0, "40001"), null, true },
+ { "statement completion unknown", sql(0, "40003"), POSTGRES, NONE },
+ { "transaction integrity constraint violation", sql(0, "40002"), POSTGRES, NONE },
+ // ... but the state of a conflict is still matched whatever vendor number carries it, which is what makes
+ // an unrecognised engine replayable at all. Which class of conflict it is cannot be told, though: the
+ // grant rests on the engine not having bounded the wait already, and of this engine that is not known
+ { "class 40 is driver independent", sql(0, "40001"), null, UNKNOWN_ENGINE },
+ // the case that costs: a MySQL-wire-compatible driver reports innodb_lock_wait_timeout as 1205 under
+ // class 40 exactly as Connector/J does, and reading the number only under a name carrying "mysql" would
+ // hand it the grant - a second full 50 s wait, the one thing the window exists to refuse
+ { "mysql wire compatible lock wait timeout", sql(1205, "40001"), MARIADB, UNKNOWN_ENGINE },
+ { "class 40 with 1205, no driver", sql(1205, "40001"), null, UNKNOWN_ENGINE },
// nothing a replay can resolve
- { "primary key violation", sql(2627, "23000"), MSSQL, false },
- { "syntax error", sql(102, "S0001"), MSSQL, false },
- { "no SQLState", sql(0, null), MSSQL, false },
- { "not a SQLException", new IllegalStateException("connection closed"), MSSQL, false },
- { "wrapped, not a conflict", new StorageRuntimeException(sql(2627, "23000")), MSSQL, false },
- { "no failure at all", null, MSSQL, false },
+ { "primary key violation", sql(2627, "23000"), MSSQL, NONE },
+ { "syntax error", sql(102, "S0001"), MSSQL, NONE },
+ { "no SQLState", sql(0, null), MSSQL, NONE },
+ { "not a SQLException", new IllegalStateException("connection closed"), MSSQL, NONE },
+ { "wrapped, not a conflict", new StorageRuntimeException(sql(2627, "23000")), MSSQL, NONE },
+ { "no failure at all", null, MSSQL, NONE },
// a vendor number is never matched without a driver to key it off, since the engines collide on it
- { "unknown driver", sql(1205, "HY000"), null, false },
- { "cyclic cause chain", new SelfCausedException(), MSSQL, false },
+ { "unknown driver", sql(1205, "HY000"), null, NONE },
+ // the state the MySQL server itself gives a lock wait timeout, before Connector/J remaps it to class 40: the
+ // number read to date that conflict refines a match its state has already made, and never makes one of its own
+ { "mysql lock wait timeout, server state", sql(1205, "HY000"), MYSQL, NONE },
+ { "cyclic cause chain", new SelfCausedException(), MSSQL, NONE },
+
+ // the chain is walked to its end, not stopped at its first conflict: a hop carrying a bare class 40 state
+ // is a conflict by itself, and returning it would hand the lock wait timeout it wraps - a wait MySQL has
+ // already bounded - the replay that only the conflicts nothing bounds are granted
+ { "lock wait timeout under a bare class 40 wrapper", sql(0, "40001", sql(1205, "40001")), MYSQL,
+ AFTER_LOCK_WAIT },
+ { "bare class 40 wrapper over a deadlock", sql(0, "40001", sql(1213, "40001")), MYSQL, PROMPT },
};
}
+ /**
+ * Whether a failure is a conflict at all decides that it is replayed; which class of conflict it is decides
+ * only whether its first replay is granted regardless of the clock.
+ */
@Test(dataProvider = "failures")
- public void testIsRetryableConflict(String name, Throwable failure, String driver, boolean expected)
+ public void testConflictClass(String name, Throwable failure, String driver, Conflict expected)
+ {
+ assertEquals(conflictOf(failure, driver), expected, name);
+ }
+
+ /**
+ * Which replays happen. The window bounds them from the first attempt, with one grant: a conflict its engine
+ * reports promptly is given its first replay whatever the clock says, since the wait charged to the attempt
+ * that hit it is unbounded and no window survives it. A conflict the engine reported only after a lock wait
+ * timeout of its own gets no such grant - that wait is bounded already, and repeating it is what the window
+ * refuses.
+ */
+ @DataProvider
+ public Object[][] replays()
+ {
+ return new Object[][] {
+ // the engine asked for the transaction to be rerun after a wait nothing here bounds, and no clock denies
+ // that first rerun. The failure of run 33010633197 is the case: SQL Server leaves the lock wait unbounded,
+ // so its deadlock monitor picked a victim ~12 s into the first attempt, and master replayed it zero times
+ { "deadlock reported after a long lock wait", 1, seconds(12), sql(1205, "40001"), MSSQL, true },
+ { "deadlock reported later than any window", 1, seconds(600), sql(1205, "40001"), MSSQL, true },
+ // and exactly at the window, which is the boundary the grant is decided on: elapsed >= window, not > it.
+ // The rows around this one bracket that point without standing on it, and a > there would refuse the first
+ // replay of #903 to every conflict reported at the window to the nanosecond
+ { "deadlock at the window, first attempt", 1, seconds(10), sql(1205, "40001"), MSSQL, true },
+
+ // the grant is one replay, not an exemption: from the second attempt on the window governs, so that a
+ // conflict which never clears is failed rather than never returned
+ { "deadlock within the window", 2, seconds(9), sql(1205, "40001"), MSSQL, true },
+ { "deadlock at the window", 2, seconds(10), sql(1205, "40001"), MSSQL, false },
+ // the same elapsed time that was granted on attempt 1 is refused on attempt 2: one grant, and only one
+ { "deadlock past the window", 2, seconds(12), sql(1205, "40001"), MSSQL, false },
+ // a MySQL deadlock is reported as promptly as any other engine reports one, so it is granted the same
+ { "mysql deadlock, first attempt", 1, seconds(12), sql(1213, "40001"), MYSQL, true },
+ { "mysql deadlock, past the window", 2, seconds(12), sql(1213, "40001"), MYSQL, false },
+
+ // MySQL reports a lock wait timeout only after innodb_lock_wait_timeout, 50 s by default: that wait is
+ // bounded by the engine, so the window is measured against it from the first attempt and a second 50 s wait
+ // is refused - which is the whole reason the window was introduced
+ { "mysql lock wait timeout at the default 50 s", 1, seconds(50), sql(1205, "40001"), MYSQL, false },
+ { "mysql lock wait timeout past the window", 1, seconds(12), sql(1205, "40001"), MYSQL, false },
+ // ... and a deployment that tuned innodb_lock_wait_timeout below the window still gets its replays
+ { "mysql lock wait timeout tuned under the window", 1, seconds(3), sql(1205, "40001"), MYSQL, true },
+ { "mysql lock wait timeout, second attempt within", 2, seconds(6), sql(1205, "40001"), MYSQL, true },
+ { "mysql lock wait timeout, second attempt at the window", 2, seconds(10), sql(1205, "40001"), MYSQL, false },
+ // the same 1205 under a MySQL-wire-compatible driver, which is what a backend created under Connector/J
+ // and opened through MariaDB Connector/J reports: the window governs it from the first attempt too, since
+ // a grant here would buy the second innodb_lock_wait_timeout the rows above refuse
+ { "mysql wire compatible lock wait timeout", 1, seconds(12), sql(1205, "40001"), MARIADB, false },
+ { "mysql wire compatible conflict within the window", 1, seconds(3), sql(1205, "40001"), MARIADB, true },
+
+ // the attempt count bounds every class, whatever the window has left. It is the bound that rarely fires:
+ // reaching it takes ten attempts inside a 10 s window, which only a conflict reported in milliseconds
+ // leaves room for - a conflict preceded by a wait longer than the window stops at two attempts, the
+ // granted one included, and the line reporting the replay names the window for that reason
+ { "last attempt left", 9, 0L, sql(1205, "40001"), MSSQL, true },
+ { "attempts exhausted", 10, 0L, sql(1205, "40001"), MSSQL, false },
+ // a failure carrying no conflict class at all still passes these bounds: what makes it replayable is
+ // replayReason(), which write() asks first, and a dropped connection carries no class 40 state
+ { "a drop, which no class describes", 1, 0L, sql(2627, "23000"), MSSQL, true },
+ };
+ }
+
+ /**
+ * Composed the way {@code write()} composes it: the class is read off the failure once, and the bounds are
+ * asked of the class. Whether the failure is worth replaying at all is {@code replayReason()}, tested apart.
+ */
+ @Test(dataProvider = "replays")
+ public void testReplayable(String name, int attempt, long elapsedNanos, Throwable failure, String driver,
+ boolean expected)
+ {
+ assertEquals(JDBCStorage.replayableWithin(attempt, elapsedNanos, conflictOf(failure, driver)),
+ expected, name);
+ }
+
+ /**
+ * The grant is reported rather than inferred: the line reporting a replay says "granted past it" only where
+ * the loop really took that branch. Pinned apart from {@link #testReplayable} because the two agree today by
+ * construction - a replay past the window can only be the grant - and it is that coincidence, not the claim,
+ * that a later change to the bounds would take away.
+ */
+ @Test
+ public void testTheGrantIsTheOnlyReplayPastTheWindow()
{
- assertEquals(JDBCStorage.isRetryableConflict(failure, driver), expected, name);
+ // the grant, and the only shape of it: the first replay of a conflict reported past the window
+ assertTrue(JDBCStorage.grantedPastTheWindow(1, seconds(12), PROMPT), "the conflict of #903 was not granted");
+ // inside the window nothing is granted - the window itself allows the replay, and the line says nothing
+ assertFalse(JDBCStorage.grantedPastTheWindow(1, seconds(9), PROMPT), "a replay inside the window was granted");
+ // and past the first attempt, or for a wait the engine already bounded, there is no grant at all
+ assertFalse(JDBCStorage.grantedPastTheWindow(2, seconds(12), PROMPT), "a second replay was granted");
+ assertFalse(JDBCStorage.grantedPastTheWindow(1, seconds(12), AFTER_LOCK_WAIT), "a bounded wait was granted");
+ assertFalse(JDBCStorage.grantedPastTheWindow(1, seconds(12), UNKNOWN_ENGINE),
+ "an engine whose wait cannot be vouched for was granted");
+ assertFalse(JDBCStorage.grantedPastTheWindow(1, seconds(12), NONE), "a failure carrying no conflict");
+
+ // every replay the bounds allow past the window is that grant, which is what lets the line name it
+ for (int attempt = 1; attempt < 12; attempt++)
+ {
+ for (Conflict conflict : Conflict.values())
+ {
+ final boolean pastTheWindow = JDBCStorage.replayableWithin(attempt, seconds(11), conflict);
+ assertEquals(pastTheWindow, JDBCStorage.grantedPastTheWindow(attempt, seconds(11), conflict),
+ "attempt " + attempt + " of a " + conflict + " conflict past the window");
+ }
+ }
}
@DataProvider
@@ -251,9 +402,9 @@ public void testIsConnectionFailure(String name, Throwable failure, boolean expe
public void testADroppedConnectionIsReplayedOnlyBeforeTheCommit()
{
final SQLException dropped = sql(0, "08006");
- assertEquals(JDBCStorage.replayReason(dropped, POSTGRES, false, false, false),
+ assertEquals(replayReason(dropped, POSTGRES, false, false, false),
"a connection the database dropped");
- assertNull(JDBCStorage.replayReason(dropped, POSTGRES, true, false, false),
+ assertNull(replayReason(dropped, POSTGRES, true, false, false),
"an in doubt transaction was replayed");
}
@@ -265,10 +416,10 @@ public void testADroppedConnectionIsReplayedOnlyBeforeTheCommit()
public void testAConnectionTheDriverClosedIsADroppedOne()
{
final SQLException killed = sql(596, "S0001");
- assertNull(JDBCStorage.replayReason(killed, MSSQL, false, false, false), "S0001 was replayed on its own");
- assertEquals(JDBCStorage.replayReason(killed, MSSQL, false, false, true),
+ assertNull(replayReason(killed, MSSQL, false, false, false), "S0001 was replayed on its own");
+ assertEquals(replayReason(killed, MSSQL, false, false, true),
"a connection the database dropped");
- assertNull(JDBCStorage.replayReason(killed, MSSQL, true, false, true),
+ assertNull(replayReason(killed, MSSQL, true, false, true),
"an in doubt transaction was replayed");
}
@@ -281,9 +432,9 @@ public void testAConnectionTheDriverClosedIsADroppedOne()
@Test
public void testAnAttemptThatCommittedPartOfItsWorkIsNotReplayed()
{
- assertNull(JDBCStorage.replayReason(sql(0, "40001"), POSTGRES, false, true, false), "a conflict was replayed");
- assertNull(JDBCStorage.replayReason(sql(0, "08006"), POSTGRES, false, true, false), "a drop was replayed");
- assertNull(JDBCStorage.replayReason(sql(596, "S0001"), MSSQL, false, true, true), "a drop was replayed");
+ assertNull(replayReason(sql(0, "40001"), POSTGRES, false, true, false), "a conflict was replayed");
+ assertNull(replayReason(sql(0, "08006"), POSTGRES, false, true, false), "a drop was replayed");
+ assertNull(replayReason(sql(596, "S0001"), MSSQL, false, true, true), "a drop was replayed");
}
/**
@@ -297,8 +448,8 @@ public void testAnAttemptThatCommittedPartOfItsWorkIsNotReplayed()
public void testAConflictIsNotReadFromTheReleaseOfTheConnection()
{
final SQLException onRelease = suppressing(sql(2627, "23000"), sql(0, "40000"));
- assertFalse(JDBCStorage.isRetryableConflict(onRelease, POSTGRES), "a conflict was read from the release");
- assertNull(JDBCStorage.replayReason(onRelease, POSTGRES, true, false, false),
+ assertEquals(conflictOf(onRelease, POSTGRES), NONE, "a conflict was read from the release");
+ assertNull(replayReason(onRelease, POSTGRES, true, false, false),
"a transaction the commit left in doubt was replayed");
// the same shape carrying a drop instead: read, since the release is where a drop is stated at all
@@ -328,16 +479,16 @@ public void testTheConnectionIsAskedWhetherTheDriverClosedIt() throws Exception
public void testAConflictIsReplayedFromEitherPhase()
{
final SQLException conflict = sql(0, "40001");
- assertEquals(JDBCStorage.replayReason(conflict, POSTGRES, false, false, false), "a conflict");
- assertEquals(JDBCStorage.replayReason(conflict, POSTGRES, true, false, false), "a conflict");
+ assertEquals(replayReason(conflict, POSTGRES, false, false, false), "a conflict");
+ assertEquals(replayReason(conflict, POSTGRES, true, false, false), "a conflict");
}
/** Everything else fails the operation, as it did before either replay existed. */
@Test
public void testAFailureOfTheStatementIsNotReplayed()
{
- assertNull(JDBCStorage.replayReason(sql(2627, "23000"), MSSQL, false, false, false));
- assertNull(JDBCStorage.replayReason(sql(2627, "23000"), MSSQL, true, false, false));
+ assertNull(replayReason(sql(2627, "23000"), MSSQL, false, false, false));
+ assertNull(replayReason(sql(2627, "23000"), MSSQL, true, false, false));
}
/** The delay grows with the attempt, so that the replays outlast a contention lasting more than a few ms. */
@@ -367,7 +518,7 @@ public void testRetryDelayGrowsAndStaysBounded()
@Test
public void testConflictSummaryNamesTheStateAndTheNumber()
{
- final String summary = JDBCStorage.conflictSummary(
+ final String summary = conflictSummary(
new DirectoryException(OTHER, raw("unchecked"), new StorageRuntimeException(sql(1205, "40001"))), POSTGRES);
assertTrue(summary.contains("40001"), summary);
assertTrue(summary.contains("1205"), summary);
@@ -382,26 +533,51 @@ public void testConflictSummaryNamesTheStateAndTheNumber()
@Test
public void testConflictSummaryNamesTheFailureTheReplayWasDecidedOn()
{
- final String summary = JDBCStorage.conflictSummary(
+ final String summary = conflictSummary(
new StorageRuntimeException(suppressing(sql(2627, "23000"), sql(0, "08006"))), POSTGRES);
assertTrue(summary.contains("08006"), summary);
assertFalse(summary.contains("23000"), summary);
}
+ /**
+ * The line names the link the class was decided on, which is not always the first conflict of the chain:
+ * {@code conflictVerdict()} keeps the most specific class it finds, so a wrapper carrying a bare class 40 state
+ * is walked past to the lock wait timeout underneath it. Naming the wrapper would print "error 0" for a replay
+ * whose whole bound was chosen by the 1205 it never shows. Asked of one verdict per chain, the way
+ * {@code write()} asks it: the class and the link are one answer, and two walks could disagree about them.
+ */
+ @Test
+ public void testConflictSummaryNamesTheLinkTheClassWasDecidedOn()
+ {
+ final SQLException lateUnderAWrapper = sql(0, "40001", sql(1205, "40001"));
+ final JDBCStorage.ConflictVerdict late = JDBCStorage.conflictVerdict(lateUnderAWrapper, MYSQL);
+ assertEquals(late.conflict, AFTER_LOCK_WAIT);
+ final String summary = JDBCStorage.conflictSummary(late, lateUnderAWrapper);
+ assertTrue(summary.contains("1205"), summary);
+
+ // the same chain under a driver that gives 1205 no such meaning is a prompt conflict, and the first link
+ // of it is the one the decision was taken on
+ final SQLException sameChain = sql(0, "40001", sql(1205, "40001"));
+ final JDBCStorage.ConflictVerdict prompt = JDBCStorage.conflictVerdict(sameChain, POSTGRES);
+ assertEquals(prompt.conflict, PROMPT);
+ final String firstLink = JDBCStorage.conflictSummary(prompt, sameChain);
+ assertTrue(firstLink.contains("error 0"), firstLink);
+ }
+
/** A failure carrying no SQLException at all, and a cyclic cause chain, still have to yield something loggable. */
@Test
public void testConflictSummaryTerminatesWithoutASQLException()
{
- assertTrue(JDBCStorage.conflictSummary(new IllegalStateException("connection closed"), POSTGRES).contains("closed"));
- assertTrue(JDBCStorage.conflictSummary(new SelfCausedException(), POSTGRES).contains("SelfCausedException"));
- assertEquals(JDBCStorage.conflictSummary(null, POSTGRES), "null");
+ assertTrue(conflictSummary(new IllegalStateException("connection closed"), POSTGRES).contains("closed"));
+ assertTrue(conflictSummary(new SelfCausedException(), POSTGRES).contains("SelfCausedException"));
+ assertEquals(conflictSummary(null, POSTGRES), "null");
}
/** A statement that carries neither a conflict nor a drop is still the one the summary names. */
@Test
public void testConflictSummaryFallsBackToTheFirstFailureOfTheChain()
{
- final String summary = JDBCStorage.conflictSummary(new StorageRuntimeException(sql(2627, "23000")), POSTGRES);
+ final String summary = conflictSummary(new StorageRuntimeException(sql(2627, "23000")), POSTGRES);
assertTrue(summary.contains("23000"), summary);
// the fallback names the statement, not the rollback of the release behind it: this is where a replay
@@ -409,7 +585,7 @@ public void testConflictSummaryFallsBackToTheFirstFailureOfTheChain()
// and the walk reaches the suppressed exceptions of a failure before its cause
final StorageRuntimeException killedSession = new StorageRuntimeException(sql(596, "S0001"));
killedSession.addSuppressed(sql(0, "25P02"));
- final String decidedOnTheConnection = JDBCStorage.conflictSummary(killedSession, MSSQL);
+ final String decidedOnTheConnection = conflictSummary(killedSession, MSSQL);
assertTrue(decidedOnTheConnection.contains("S0001"), decidedOnTheConnection);
assertFalse(decidedOnTheConnection.contains("25P02"), decidedOnTheConnection);
}
@@ -427,6 +603,112 @@ public void testTheWalkOfAFailureStopsAtItsBudget()
assertFalse(JDBCStorage.isConnectionFailure(chainEndingInADrop(65)), "a drop past the budget was walked to");
}
+ /**
+ * The class of a conflict is the one walk that budget must not bound, and it is the reason
+ * {@code failureScope()} does not bound its own either: truncation does not leave this verdict unanswered, it
+ * weakens it. A lock wait timeout past the budget, with a bare class 40 link inside it, would come back
+ * {@link Conflict#PROMPT} and be handed the one replay the class exists to refuse - a second full
+ * {@code innodb_lock_wait_timeout}. Truncation here grants a replay rather than losing one.
+ */
+ @Test
+ public void testTheConflictClassIsReadFromEveryLinkOfTheChain()
+ {
+ // a bare class 40 wrapper, then 64 links of a rejected statement, then the timeout that decided the class
+ final SQLException bareClass40 = sql(0, "40001");
+ SQLException tail = bareClass40;
+ for (int link = 0; link < 64; link++)
+ {
+ tail = chained(tail, sql(2627, "23000")).getNextException();
+ }
+ chained(tail, sql(1205, "40001"));
+
+ final JDBCStorage.ConflictVerdict verdict = JDBCStorage.conflictVerdict(bareClass40, MYSQL);
+ assertEquals(verdict.conflict, AFTER_LOCK_WAIT,
+ "a lock wait timeout past MAX_CHAIN_LINKS came back as a conflict the window does not bound");
+ assertFalse(JDBCStorage.replayableWithin(1, seconds(12), verdict.conflict),
+ "and was granted the replay past the window");
+ // the line reporting a replay names that same link, since one walk produced both
+ assertTrue(JDBCStorage.conflictSummary(verdict, bareClass40).contains("1205"));
+ }
+
+ /**
+ * The widening that reading every link brings, which is the one behaviour change of #903 outside the grant: a
+ * chain whose only conflict-bearing link sits past the budget was not a conflict at all on master - the
+ * truncating walk never reached it - so {@code replayReason()} answered null and the operation was failed.
+ * The case above pins the *class* of such a chain, since it puts a bare class 40 state at the head that the
+ * truncating walk already matched; this one pins that the conflict is found at all.
+ */
+ @Test
+ public void testAConflictOnlyPastTheBudgetIsFoundAtAll()
+ {
+ // 64 links of a rejected statement - the whole budget - and the conflict on the 65th
+ final SQLException head = sql(2627, "23000");
+ SQLException tail = head;
+ for (int link = 2; link <= 64; link++)
+ {
+ tail = chained(tail, sql(2627, "23000")).getNextException();
+ }
+ chained(tail, sql(0, "40001"));
+
+ assertEquals(conflictOf(head, POSTGRES), PROMPT,
+ "a conflict whose only link sits past MAX_CHAIN_LINKS was not found at all");
+ assertEquals(replayReason(head, POSTGRES, false, false, false), "a conflict",
+ "and the operation carrying it was not replayed");
+ }
+
+ /**
+ * What that walk stops at is the strongest class the engine of its driver can report, not the last constant of
+ * {@link Conflict}: only MySQL reports a lock wait timeout of its own, so a walk stopping at
+ * {@link Conflict#AFTER_LOCK_WAIT} never stops early on the other three engines and reads every link of every
+ * failed write on the driver whose chains are longest. Pinned from both sides - no failure of an engine is
+ * classed above its own ceiling, and every ceiling is reached by some failure - since a ceiling set too low
+ * would end the walk on a class weaker than the chain carries, and one set too high never ends it early.
+ */
+ @Test
+ public void testTheWalkStopsAtTheStrongestClassItsEngineCanReport()
+ {
+ // one shape per branch of the classification: a bare class 40, the two numbers the engines collide on, a
+ // deadlock of oracle, the state MySQL reports before its driver remaps it, and a failure that is no conflict
+ final SQLException[] shapes = {
+ sql(0, "40001"), sql(1205, "40001"), sql(1213, "40001"), sql(60, "61000"), sql(1205, "HY000"),
+ sql(2627, "23000") };
+
+ for (String driver : new String[] { POSTGRES, MYSQL, ORACLE, MSSQL, MARIADB, null })
+ {
+ final Conflict ceiling = JDBCStorage.ceilingOf(JDBCStorage.dialectOf(driver));
+ Conflict strongest = NONE;
+ for (SQLException shape : shapes)
+ {
+ final Conflict conflict = conflictOf(shape, driver);
+ assertTrue(conflict.compareTo(ceiling) <= 0,
+ driver + " classed " + shape.getSQLState() + "/" + shape.getErrorCode() + " as " + conflict
+ + ", above the ceiling its walk stops at");
+ strongest = conflict.compareTo(strongest) > 0 ? conflict : strongest;
+ }
+ assertEquals(strongest, ceiling, "the walk of " + driver + " stops at a class it can never reach");
+ }
+
+ // and the walk really stops there: a link behind a conflict already at its engine's ceiling is not looked
+ // at. This is what makes reading every link affordable - write() classifies before it knows the failure is
+ // replayable at all, so a plain 23000 from adding an entry that is already there reaches this walk too
+ final AtomicBoolean walkedPast = new AtomicBoolean();
+ final SQLException behindTheCeiling = new SQLException("synthetic failure", "23000", 2627)
+ {
+ @Override
+ public String getSQLState()
+ {
+ walkedPast.set(true);
+ return super.getSQLState();
+ }
+ };
+ assertEquals(conflictOf(sql(0, "40P01", behindTheCeiling), POSTGRES), PROMPT);
+ assertFalse(walkedPast.get(), "an engine reporting no lock wait timeout of its own walked past its ceiling");
+
+ // under mysql the same head is not the ceiling - a lock wait timeout could still be behind it - so it is
+ assertEquals(conflictOf(sql(0, "40001", behindTheCeiling), MYSQL), PROMPT);
+ assertTrue(walkedPast.get(), "mysql stopped before the link a lock wait timeout could have been on");
+ }
+
/**
* Opening a tree that is already there commits nothing, so the attempt stays replayable. The create table is
* guarded by a catalog read, and so is the create index on every engine but postgresql, so on an existing
@@ -515,7 +797,7 @@ public void testCreatingATreeTakesTheAttemptOutOfTheReplay() throws Exception
}
catch (StorageRuntimeException expected)
{
- assertTrue(JDBCStorage.isRetryableConflict(expected, POSTGRES), "the conflict was not the failure raised");
+ assertEquals(conflictOf(expected, POSTGRES), PROMPT, "the conflict was not the failure raised");
}
assertEquals(attempts.get(), 1, "a transaction that committed part of its work was replayed");
verify(statements).executeUpdate();
@@ -746,7 +1028,7 @@ public void testACreateIndexMysqlCommittedBeforeTakesTheAttemptOutOfTheReplay()
}
catch (StorageRuntimeException expected)
{
- assertTrue(JDBCStorage.isRetryableConflict(expected, MYSQL), "the conflict was not the failure raised");
+ assertEquals(conflictOf(expected, MYSQL), PROMPT, "the conflict was not the failure raised");
}
assertEquals(attempts.get(), 1, "an attempt that committed part of its work was replayed");
verify(engineConnection).prepareStatement(startsWith("create index k_"));
@@ -843,6 +1125,22 @@ private static ResultSet noRows() throws SQLException
return rs;
}
+ /**
+ * A mock of the given connection type, with the name every engine branch of {@code JDBCStorage} is keyed on
+ * asserted rather than assumed. The name of a mock is derived from the type it mocks, so a renamed fixture -
+ * or a Mockito that names its mocks differently - would move a case into another engine's branch with no test
+ * saying so, and there are cases no count of attempts would catch that in.
+ */
+ private static Connection mockOfEngine(Class engine)
+ {
+ final Connection con = mock(engine);
+ final String engineName = engine.getSimpleName().replace("Connection", "");
+ assertTrue(JDBCStorage.driverNameOf(con).contains(engineName),
+ "a mock of " + engine.getSimpleName() + " reaches no " + engineName + " branch: "
+ + JDBCStorage.driverNameOf(con));
+ return con;
+ }
+
/**
* The connection the tree catalog of a storage of this test is written on: it opens one straight through the
* driver, for the reason a stamp opens one of its own - the caller of openTree() is holding a pooled
@@ -872,11 +1170,7 @@ private Connection catalogConnection() throws Exception
private JDBCStorage storageOverAnEngine(Class engine, boolean theIndex,
Connection... behind) throws Exception
{
- final Connection con = mock(engine);
- final String engineName = engine.getSimpleName().replace("Connection", "");
- assertTrue(JDBCStorage.driverNameOf(con).contains(engineName),
- "a mock of " + engine.getSimpleName() + " reaches no " + engineName + " branch: "
- + JDBCStorage.driverNameOf(con));
+ final Connection con = mockOfEngine(engine);
engineConnection = con;
// the connections behind it answer the connects the pool does not make: the tree catalog is read and
// written on one of its own, straight through the driver, since the caller of openTree() is holding a
@@ -999,11 +1293,155 @@ public Logger getParentLogger()
}
}
+ /**
+ * 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(Throwable failure, String driver, boolean committing, boolean partlyCommitted,
+ boolean connectionClosed)
+ {
+ return JDBCStorage.replayReason(conflictOf(failure, driver), failure, committing, partlyCommitted,
+ connectionClosed);
+ }
+
+ /**
+ * The class of a failure, composed of the pieces {@code write()} composes it of. Forms of this and of
+ * {@link #conflictSummary} taking a failure and a driver used to live in {@code JDBCStorage} with no caller of
+ * their own in {@code src/main}, which left the classification javadoc hanging off methods production never
+ * called and made every test asking both questions walk the chains twice. The convenience is a test's, so it
+ * is written here.
+ */
+ private static Conflict conflictOf(Throwable failure, String driver)
+ {
+ return JDBCStorage.conflictVerdict(failure, driver).conflict;
+ }
+
+ /** The line reporting a replay, composed the way {@code write()} composes it: one walk, then the summary of it. */
+ private static String conflictSummary(Throwable failure, String driver)
+ {
+ return JDBCStorage.conflictSummary(JDBCStorage.conflictVerdict(failure, driver), failure);
+ }
+
private static SQLException sql(int errorCode, String sqlState)
{
return new SQLException("synthetic failure", sqlState, errorCode);
}
+ private static SQLException sql(int errorCode, String sqlState, Throwable cause)
+ {
+ return new SQLException("synthetic failure", sqlState, errorCode, cause);
+ }
+
+ private static long seconds(long seconds)
+ {
+ return TimeUnit.SECONDS.toNanos(seconds);
+ }
+
+ /**
+ * How {@link JDBCStorage#write} drives the two decisions above, which the cases before this one cannot see:
+ * they are handed an elapsed time and an attempt number rather than producing them. The clock is scripted and
+ * advances a fixed step per attempt - not per read of it - so that the timeline the loop sees depends on what
+ * it does rather than on how often it asks the time: a read added anywhere in {@code write()} leaves every row
+ * of this provider answering exactly as it does now.
+ *
+ * Between them the rows pin the three lines the rest of the file would let a refactor take away. A single
+ * {@code startedAt} outside the retry loop is what makes the window bound the whole run rather than each
+ * attempt: moved inside, every attempt is measured against its own start, sees the step and nothing more, and
+ * replays to MAX_RETRIES. The grant of the first replay is what issue #903 is about: without it an attempt
+ * that alone outlasts the window leaves the loop with no replay at all. And the class the grant is asked of is
+ * read off the failure of this very run, rather than off a driver the loop does not carry: the last two rows
+ * fail as plainly as the first two and are replayed no times at all.
+ */
+ @DataProvider
+ public Object[][] writeRuns()
+ {
+ return new Object[][] {
+ // a step under the window, so the window is what ends the run: attempt 1 is granted its replay at 4 s,
+ // attempt 2 is inside the window at 8 s, attempt 3 is past it at 12 s. With startedAt inside the loop every
+ // attempt measures 4 s, never reaches the window, and the run goes to MAX_RETRIES instead
+ { "the window bounds the run, not the attempt", postgresConnection.class, sql(0, "40P01"), 4L, PROMPT, 3 },
+ // a step past the window, so only the grant can produce a second attempt: remove it and the run ends on
+ // the first. This is the row that pins the grant end to end, and the row above is the one that pins
+ // startedAt - at 12 s a per-attempt startedAt also stops at two attempts, and at 4 s the window alone
+ // already allows the replay of attempt 1. Neither row is redundant
+ { "the first replay is granted past the window", postgresConnection.class, sql(0, "40P01"), 12L, PROMPT, 2 },
+ // the same step, and the same class 40 state, for the one conflict the engine had already bounded: a
+ // single attempt. The predicate rows pin that decision, but only these rows pin that write() hands the
+ // predicate the class of its own failure - dropped on the way to replayableWithin() and the run above
+ // stays green while this one buys a second innodb_lock_wait_timeout
+ { "a lock wait timeout is granted no replay", mysqlConnection.class, sql(1205, "40001"), 12L,
+ AFTER_LOCK_WAIT, 1 },
+ // and the driver that reports that same timeout under a name this backend does not recognise: no grant
+ // there either, since what the grant rests on - the engine having bounded nothing - is unknown of it.
+ // No number of attempts separates this row from the one above: AFTER_LOCK_WAIT and UNKNOWN_ENGINE are
+ // refused the grant and bounded by the window identically, so the mysql row misread as unrecognised
+ // produces exactly the count expected of it. That is why each row names the class its engine gives its
+ // failure and the case asserts it of the mock it actually built
+ { "an unrecognised engine is granted no replay", mariadbConnection.class, sql(1205, "40001"), 12L,
+ UNKNOWN_ENGINE, 1 },
+ };
+ }
+
+ @Test(dataProvider = "writeRuns")
+ public void testWriteDrivesTheRetryLoop(String name, Class engine,
+ final SQLException conflict, final long stepSeconds, Conflict expectedClass, int expectedAttempts)
+ throws Exception
+ {
+ final AtomicInteger attempts = new AtomicInteger();
+ // the class name of the mock is what write() reads the engine off, asserted here the way
+ // storageOverAnEngine() asserts it
+ final Connection connection = mockOfEngine(engine);
+ // which branch of the classification the row reaches, asserted rather than inferred from the count: the
+ // count cannot tell AFTER_LOCK_WAIT from UNKNOWN_ENGINE, since both are refused the grant and bounded by
+ // the window alike, so a fixture renamed out of the mysql branch would leave that row green
+ assertEquals(conflictOf(conflict, JDBCStorage.driverNameOf(connection)), expectedClass,
+ name + ": the mock does not reach the branch the row names");
+
+ // a url of its own, as storageOver() gives every fixture of this file: getConnection() is overridden below,
+ // but distrustPool() is not, and a null one would reach ConcurrentHashMap.merge(null, ...) rather than the
+ // assertion under test the moment a row of this provider scripts a connection failure
+ final JDBCBackendCfg cfg = mock(JDBCBackendCfg.class);
+ when(cfg.getDBDirectory()).thenReturn(StubDriver.PREFIX + pools.incrementAndGet());
+
+ final JDBCStorage storage = new JDBCStorage(cfg, null)
+ {
+ @Override
+ Connection getConnection()
+ {
+ return connection;
+ }
+
+ @Override
+ long nanoTime()
+ {
+ // the attempts made are what moves this clock, so the run is the same however often write() reads it
+ return seconds(stepSeconds * attempts.get());
+ }
+ };
+ storage.accessMode = AccessMode.READ_WRITE;
+
+ StorageRuntimeException thrown = null;
+ try
+ {
+ storage.write(new WriteOperation()
+ {
+ @Override
+ public void run(WriteableTransaction txn)
+ {
+ attempts.incrementAndGet();
+ throw new StorageRuntimeException(conflict);
+ }
+ });
+ }
+ catch (StorageRuntimeException e)
+ {
+ thrown = e;
+ }
+
+ assertSame(thrown != null ? thrown.getCause() : null, conflict, name + ": the conflict reaches the caller");
+ assertEquals(attempts.get(), expectedAttempts, name + ": attempts made");
+ }
+
/** The second failure as the next exception of the first, the way a driver chains the errors of one message. */
private static SQLException chained(SQLException first, SQLException next)
{