Re: Error from array_agg when table has many rows

Richard Guo <guofenglinux@gmail.com> writes:
> We are performing deserialization during the final phase of the
> aggregation on data of type RECORD but we fail to provide a valid
> typmod (array_agg_deserialize() uses -1 as the typmod when calling the
> receiveproc).

> I haven't verified it, but I suspect it's related to 16fd03e95.

Yeah.  I don't think there is any way for array_agg_deserialize to
know the correct typmod, so what we have to do is disable using
partial aggregation in this case.  Fortunately there's a
policy-setting function that can be taught that, as attached.

            regards, tom lane

diff --git a/src/backend/parser/parse_agg.c b/src/backend/parser/parse_agg.c
index 9e567f3cc45..0ac8966e30f 100644
--- a/src/backend/parser/parse_agg.c
+++ b/src/backend/parser/parse_agg.c
@@ -2052,7 +2052,7 @@ resolve_aggregate_transtype(Oid aggfuncid,
 
 /*
  * agg_args_support_sendreceive
- *        Returns true if all non-byval of aggref's arg types have send and
+ *        Returns true if all non-byval types of aggref's args have send and
  *        receive functions.
  */
 bool
@@ -2067,6 +2067,15 @@ agg_args_support_sendreceive(Aggref *aggref)
         TargetEntry *tle = (TargetEntry *) lfirst(lc);
         Oid            type = exprType((Node *) tle->expr);
 
+        /*
+         * RECORD is a special case: it has typsend/typreceive functions, but
+         * record_recv only works if passed the correct typmod to identify the
+         * specific anonymous record type.  array_agg_deserialize cannot do
+         * that, so we have to disclaim support for the case.
+         */
+        if (type == RECORDOID)
+            return false;
+
         typeTuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(type));
         if (!HeapTupleIsValid(typeTuple))
             elog(ERROR, "cache lookup failed for type %u", type);