Github user sureshsubbiah commented on a diff in the pull request:
https://github.com/apache/incubator-trafodion/pull/1232#discussion_r139228319
--- Diff: core/sql/sqlcomp/nadefaults.cpp ---
@@ -1336,6 +1336,8 @@ SDDkwd__(EXE_DIAGNOSTIC_EVENTS, "OFF"),
SDDui___(EXE_MEMORY_FOR_PROBE_CACHE_IN_MB, "100"),
+ SDDui___(EXE_MEMORY_FOR_UNPACK_ROWS_IN_MB, "100"),
--- End diff --
A default value of 100 MB maybe too small. I understand that this is good
for mixed workloads, but do consider than plans with Unpack, especially when
used for insert/upsert are simple. Unpack is always serial and part of master
exe. Often there is only one in a query. The cost of having a low value here
seems to be that queue length from unpack could be limited by this when we are
loading a table with wide rows (even if actual data in row is not wide). We
know from performance runs that queue length is critical to upsert load
performance. This is a hunch on my part and could be misguided. It will be good
to performance test for upsert of table with wide rows, if you think that is
worthwhile.
---
