Fix extreme skew detection in Parallel Hash Join.

macdice · macdice · commit 20d94899414a · 2024-10-17T22:01:54.000+13:00
After repartitioning the inner side of a hash join that would have exceeded the allowed size, we check if all the tuples from a parent partition moved to one child partition. That is evidence that it contains duplicate keys and later attempts to repartition will also fail, so we should give up trying to limit memory (for lack of a better fallback strategy). A thinko prevented the check from working correctly in partition 0 (the one that is partially loaded into memory already). After repartitioning, we should check for extreme skew if the *parent* partition's space_exhausted flag was set, not the child partition's. The consequence was repeated futile repartitioning until per-partition data exceeded various limits including "ERROR: invalid DSA memory alloc request size 1811939328", OS allocation failure, or temporary disk space errors. (We could also do something about some of those symptoms, but that's material for separate patches.) This problem only became likely when PostgreSQL 16 introduced support for Parallel Hash Right/Full Join, allowing NULL keys into the hash table. Repartitioning always leaves NULL in partition 0, no matter how many times you do it, because the hash value is all zero bits. That's unlikely for other hashed values, but they might still have caused wasted extra effort before giving up. Back-patch to all supported releases. Reported-by: Craig Milhiser <craig@milhiser.com> Reviewed-by: Andrei Lepikhov <lepihov@gmail.com> Discussion: https://postgr.es/m/CA%2BwnhO1OfgXbmXgC4fv_uu%3DOxcDQuHvfoQ4k0DFeB0Qqd-X-rQ%40mail.gmail.com
diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
@@ -1252,32 +1252,39 @@ ExecParallelHashIncreaseNumBatches(HashJoinTable hashtable)
 			if (BarrierArriveAndWait(&pstate->grow_batches_barrier,
 									 WAIT_EVENT_HASH_GROW_BATCHES_DECIDE))
 			{
+				ParallelHashJoinBatch *old_batches;
 				bool		space_exhausted = false;
 				bool		extreme_skew_detected = false;
 
 				/* Make sure that we have the current dimensions and buckets. */
 				ExecParallelHashEnsureBatchAccessors(hashtable);
 				ExecParallelHashTableSetCurrentBatch(hashtable, 0);
 
+				old_batches = dsa_get_address(hashtable->area, pstate->old_batches);
+
 				/* Are any of the new generation of batches exhausted? */
 				for (i = 0; i < hashtable->nbatch; ++i)
 				{
-					ParallelHashJoinBatch *batch = hashtable->batches[i].shared;
+					ParallelHashJoinBatch *batch;
+					ParallelHashJoinBatch *old_batch;
+					int			parent;
 
+					batch = hashtable->batches[i].shared;
 					if (batch->space_exhausted ||
 						batch->estimated_size > pstate->space_allowed)
-					{
-						int			parent;
-
 						space_exhausted = true;
 
+					parent = i % pstate->old_nbatch;
+					old_batch = NthParallelHashJoinBatch(old_batches, parent);
+					if (old_batch->space_exhausted ||
+						batch->estimated_size > pstate->space_allowed)
+					{
 						/*
 						 * Did this batch receive ALL of the tuples from its
 						 * parent batch?  That would indicate that further
 						 * repartitioning isn't going to help (the hash values
 						 * are probably all the same).
 						 */
-						parent = i % pstate->old_nbatch;
 						if (batch->ntuples == hashtable->batches[parent].shared->old_ntuples)
 							extreme_skew_detected = true;
 					}
