I looked into bug #13756,
http://www.postgresql.org/message-id/20151105171933.14035.25039@wrigleys.postgresql.org
The cause of the problem is that gin_extract_jsonb_path() computes
a different hash for the "2" in
{"a":[ ["b",{"x":1}], ["b",{"x":2}]]}
than it does for the "2" in
{"a":[[{"x":2}]]}
And the cause of that is that it supposes that after emitting a hash
for a VALUE, it can just leave stack->hash to be reinitialized by the
next KEY or ELEM. But if the next thing is a sub-object, we propagate
the previous value's hash down into the sub-object, causing values
therein to receive different hashes than they'd have gotten without
the preceding outer-level VALUE.
The attached one-liner fixes it, but of course this means that on-disk
jsonb_path_ops indexes are possibly broken and will need to be reindexed.
I see no way around that ... does anybody else?
regards, tom lane
diff --git a/src/backend/utils/adt/jsonb_gin.c b/src/backend/utils/adt/jsonb_gin.c
index 204fb8b..b917ec2 100644
*** a/src/backend/utils/adt/jsonb_gin.c
--- b/src/backend/utils/adt/jsonb_gin.c
*************** gin_extract_jsonb_path(PG_FUNCTION_ARGS)
*** 419,425 ****
JsonbHashScalarValue(&v, &stack->hash);
/* and emit an index entry */
entries[i++] = UInt32GetDatum(stack->hash);
! /* Note: we assume we'll see KEY before another VALUE */
break;
case WJB_END_ARRAY:
case WJB_END_OBJECT:
--- 419,426 ----
JsonbHashScalarValue(&v, &stack->hash);
/* and emit an index entry */
entries[i++] = UInt32GetDatum(stack->hash);
! /* reset hash for next sub-object */
! stack->hash = stack->parent->hash;
break;
case WJB_END_ARRAY:
case WJB_END_OBJECT: