GeezerVison

Add the oracles PQS, HAVING, AGGREGATE, DISTINCT, GROUP_BY,
QUERY_PARTITIONING and DQP. DQP runs each query under GPORCA and the
PostgreSQL-based planner, and with join, aggregation and motion methods
disabled, and checks that all plans return the same rows. The CI job
now also runs TestCloudberryTLP, TestCloudberryPQS and TestCloudberryDQP.

GPORCA falls back to the PostgreSQL-based planner for queries with
ONLY or whole-row references, so the new oracles and the NoREC and
WHERE oracles no longer generate them.

Fix two places where CloudberryProvider ignored --cloudberry-oracle:
it now skips databases with empty tables when an oracle requires rows,
and enables generateOnlyKnown for PQS.

Work around the Cloudberry bugs found by these oracles:
- apache/cloudberry#2134: expect the errors of a PAX table broken by
  REINDEX DATABASE CONCURRENTLY
- apache/cloudberry#2135: do not generate CREATE STATISTICS, which can
  crash the coordinator
- apache/cloudberry#2137: do not generate SET wal_compression, which
  crashes the mirrors

Also expect the error raised when a redistributed table has an
inheritance child with an incompatible UNIQUE index.

mrigger left a comment