Fix crash when converting to Hypercore TAM

The tuple sort state used for compression was created with a reference
to a tuple descriptor in a relcache entry. If this relcache entry is
invalidated, the tuple sort state becomes corrupt. This caused
occasional crashes when converting a chunk to using Hypercore TAM
(i.e., compressing).

Fix this by always making a copy of the tuple descriptor when creating
the tuple sort state.

tsl/src/compression/compression.c

@@ -560,7 +560,10 @@ compression_create_tuplesort_state(CompressionSettings *settings, Relation rel)
 													 &nulls_first[n]);
 	}
 
-	return tuplesort_begin_heap(tupdesc,
+	/* Make a copy of the tuple descriptor so that it is allocated on the same
+	 * memory context as the tuple sort instead of pointing into the relcache
+	 * entry that could be blown away. */
+	return tuplesort_begin_heap(CreateTupleDescCopy(tupdesc),
 								n_keys,
 								sort_keys,
 								sort_operators,

tsl/src/hypercore/hypercore_handler.c

@@ -3464,6 +3464,14 @@ convert_to_hypercore_finish(Oid relid)
 		return;
 	}
 
+#ifdef USE_ASSERT_CHECKING
+	/* Blow away relation cache to test that the tuple sort state works across
+	 * relcache invalidations. Previously there was sometimes a crash here
+	 * because the tuple sort state had a reference to a tuple descriptor in
+	 * the relcache. */
+	RelationCacheInvalidate(false);
+#endif
+
 	Chunk *chunk = ts_chunk_get_by_relid(conversionstate->relid, true);
 	Relation relation = table_open(conversionstate->relid, AccessShareLock);
 	TupleDesc tupdesc = RelationGetDescr(relation);