vldpyatkov commented on code in PR #13366:
URL: https://github.com/apache/ignite/pull/13366#discussion_r3755796372
##########
modules/calcite/src/main/java/org/apache/ignite/internal/processors/query/calcite/exec/ExecutionServiceImpl.java:
##########
@@ -581,6 +601,388 @@ private FieldsQueryCursor<List<?>>
executeDdl(RootQuery<Row> qry, DdlPlan plan)
}
}
+ /**
+ * Executes a {@code SELECT ... FOR UPDATE} plan.
+ *
+ * <ol>
+ * <li>Validates that the current transaction is PESSIMISTIC.</li>
+ * <li>Runs the inner SELECT with hidden key, value, and version columns
and materialises all rows.</li>
+ * <li>Builds cache entries from the hidden columns.</li>
+ * <li>Creates a savepoint, acquires pessimistic locks via {@code
lockTxEntries()},
+ * and releases the savepoint on success (or rolls back on
failure).</li>
+ * <li>Repeats the SELECT and lock attempt after a concurrent version
change while the deadline permits.</li>
+ * <li>Returns a cursor with only the user-visible columns (the appended
_KEY is stripped).</li>
+ * </ol>
+ */
+ private FieldsQueryCursor<List<?>> executeForUpdate(RootQuery<Row> qry,
SelectForUpdatePlan plan) {
+ GridNearTxLocal userTx = Commons.queryTransaction(qry.context(),
ctx.cache().context());
+
+ if (userTx == null || !userTx.pessimistic())
+ throw new IgniteSQLException(
+
IgniteResource.INSTANCE.selectForUpdateRequiresPessimisticTx().str(),
+ IgniteQueryErrorCode.UNSUPPORTED_OPERATION);
+
+ long waitMs = waitMillis(plan);
+
+ // Zero means that retries are limited only by the transaction or
query timeout.
+ long lockAcquisitionEndTime = waitMs > 0
+ ? U.currentTimeMillis() + waitMs
+ : waitMs < 0 ? U.currentTimeMillis() : 0L;
+
+ RootQuery<Row> selectQry = qry;
+
+ while (true) {
+ FieldsQueryCursor<List<?>> cursor = tryExecuteForUpdate(selectQry,
plan, userTx, waitMs, lockAcquisitionEndTime);
+
+ if (cursor != null)
+ return cursor;
+
+ if (lockAcquisitionEndTime != 0 && U.currentTimeMillis() >=
lockAcquisitionEndTime) {
+ throw new IgniteSQLException(
+ IgniteResource.INSTANCE.selectForUpdateLockFailed().str(),
+ IgniteQueryErrorCode.CONCURRENT_UPDATE);
+ }
+
+ // The previous query has already been closed after execution, so
retry with a fresh root query.
+ selectQry = qry.retryQuery();
+ qryReg.register(selectQry);
+ }
+ }
+
+ /**
+ * Converts the SQL lock wait value to the internal millisecond
representation.
+ *
+ * @param plan SELECT FOR UPDATE plan.
+ * @return {@code 0} for the remaining transaction/query timeout, {@code
-1} for NOWAIT,
+ * or a positive timeout in milliseconds.
+ */
+ private static long waitMillis(SelectForUpdatePlan plan) {
+ // Convert SQL waitSeconds to the internal lock-wait representation:
+ // null means use the remaining transaction time or the query timeout
and is encoded as 0;
+ // 0 requests NOWAIT and is encoded as -1; a positive value is
converted from seconds to milliseconds.
+ Long waitSeconds = plan.waitSeconds();
+
+ if (waitSeconds == null)
+ return 0L;
+ else if (waitSeconds == 0L)
+ return -1L;
+ else
+ return waitSeconds * 1000L;
+ }
+
+ /**
+ * Executes the inner SELECT and attempts to acquire transaction locks for
the selected row versions.
+ *
+ * @param qry Root query for this execution attempt.
+ * @param plan SELECT FOR UPDATE plan.
+ * @param userTx Transaction that acquires the locks.
+ * @param waitMs Lock wait time in the internal representation.
+ * @param lockAcquisitionEndTime Absolute lock acquisition deadline in
milliseconds.
+ * @return Result cursor if all required locks were acquired, or {@code
null} if at least one lock was not acquired.
+ */
+ @Nullable private FieldsQueryCursor<List<?>> tryExecuteForUpdate(
+ RootQuery<Row> qry,
+ SelectForUpdatePlan plan,
+ GridNearTxLocal userTx,
+ long waitMs,
+ long lockAcquisitionEndTime
+ ) {
+ // Run the inner SELECT (with _KEY, _VAL, _VER appended) and collect
all rows.
+ ListFieldsQueryCursor<?> innerCursor = mapAndExecutePlan(qry,
plan.innerPlan());
+
+ // TODO: IGNITE-28957 SELECT FOR UPDATE may cause OOM by materializing
the entire result set.
+ List<List<?>> rows = innerCursor.getAll();
+
+ int userColCnt = plan.userColumnCount();
+
+ if (rows.isEmpty())
+ return createResultCursor(qry, plan, rows, userColCnt);
+
+ List<Map.Entry<IgniteInternalCache<Object, Object>, Map<Object,
CacheEntry<Object, Object>>>> lockBatches =
+ collectLockBatches(plan, rows);
+
+ if (!tryAcquireLocks(userTx, lockBatches, waitMs,
lockAcquisitionEndTime))
+ return null;
+
+ return createResultCursor(qry, plan, rows, userColCnt);
+ }
+
+ /**
+ * Collects unique cache entries to lock and orders them by cache ID,
partition ID, key hash, and key bytes.
+ *
+ * @param plan SELECT FOR UPDATE plan containing the lock targets.
+ * @param rows Selected rows containing the internal lock columns.
+ * @return Ordered cache-entry batches.
+ */
+ private List<Map.Entry<IgniteInternalCache<Object, Object>, Map<Object,
CacheEntry<Object, Object>>>>
+ collectLockBatches(SelectForUpdatePlan plan, List<List<?>> rows) {
+ Map<IgniteInternalCache<Object, Object>, Map<Object,
CacheEntry<Object, Object>>> lockBatches =
+ new TreeMap<>(Comparator.comparingInt(cache ->
cache.context().cacheId()));
+
+ for (LockTarget target : plan.lockTargets()) {
+ SchemaPlus schemaPlus = schemaHolder.schema(target.schemaName());
+
+ if (schemaPlus == null)
+ throw new IgniteSQLException("Schema not found: " +
target.schemaName(),
+ IgniteQueryErrorCode.SCHEMA_NOT_FOUND);
+
+ IgniteTable igniteTable =
(IgniteTable)schemaPlus.getTable(target.tableName());
+
+ if (igniteTable == null)
+ throw new IgniteSQLException("Table not found: " +
target.tableName(),
+ IgniteQueryErrorCode.TABLE_NOT_FOUND);
+
+ GridCacheContext<Object, Object> cctx =
+ (GridCacheContext<Object,
Object>)((CacheTableDescriptor)igniteTable.descriptor()).cacheContext();
+
+ IgniteInternalCache<Object, Object> cache =
cctx.cache().keepBinary();
+ Map<Object, CacheEntry<Object, Object>> entries =
+ lockBatches.computeIfAbsent(cache, key -> new
LinkedHashMap<>());
+ int keyColumnIdx = target.keyColumnIndex();
+
+ for (List<?> row : rows) {
+ Object key = row.get(keyColumnIdx);
+
+ // An outer join has no row to lock on its non-matching side.
+ if (key == null)
+ continue;
+
+ Object val = row.get(keyColumnIdx + 1);
+ GridCacheVersion ver = (GridCacheVersion)row.get(keyColumnIdx
+ 2);
+
+ // JOINs can repeat a row, but a transaction needs only one
lock per cache key.
+ entries.put(key, new CacheEntryImplEx<>(key, val, ver));
+ }
+ }
+
+ for (Map.Entry<IgniteInternalCache<Object, Object>, Map<Object,
CacheEntry<Object, Object>>> batch :
+ lockBatches.entrySet()) {
+ Map<Object, CacheEntry<Object, Object>> entries = batch.getValue();
+
+ orderLockEntries(batch.getKey().context(), entries);
+ }
+
+ return new ArrayList<>(lockBatches.entrySet());
+ }
+
+ /**
+ * Orders entries by partition, key hash, and serialized key bytes in case
of a hash collision.
+ *
+ * @param cctx Cache context used to prepare cache keys.
+ * @param entries Entries to order.
+ */
+ private static void orderLockEntries(
+ GridCacheContext<Object, Object> cctx,
+ Map<Object, CacheEntry<Object, Object>> entries
+ ) {
+ List<LockEntry> orderedEntries = new ArrayList<>(entries.size());
+
+ for (CacheEntry<Object, Object> entry : entries.values()) {
+ KeyCacheObject key = cctx.toCacheKeyObject(entry.getKey());
+
+ orderedEntries.add(new LockEntry(entry, key));
+ }
+
+ orderedEntries.sort(Comparator.comparingInt((LockEntry entry) ->
entry.part)
+ .thenComparingInt(entry -> entry.keyHash));
+
+ for (int start = 0; start < orderedEntries.size(); ) {
Review Comment:
Probably it is a premature optimization. Moreover, we don't have sorting in
other bulk ops.
https://issues.apache.org/jira/browse/IGNITE-28958
Maybe the right way to exclude the code at all and resolve it in the other
ticket?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
