pqFlush is sending data synchronously though the socket is set
O_NONBLOCK, this incurs performance downgradation. This commit uses
pqFlushNonBlocking instead, and synchronizes the completion of
dispatching to all Gangs before query execution.

Signed-off-by: Kenan Yao<kyao@pivotal.io>

Because it makes it so much faster. I'm not sure which of these changes
exactly are required to pass down the option to installcheck-good, but
Ashwin tested this and said it works this way.

It is a footgun, but it's just too useful to not have, for things like
running the regression test, where durability on an OS crash is not a
concern.

Per discussion at https://groups.google.com/a/greenplum.org/d/msg/gpdb-dev/0HjPmUk469o/Hg2CdamaCAAJ

* Resolving var from root->parse->rtable instead of simple_rte_array.

The optimization as introduced in
afcf09dd breaks bugbuster view tests as
it appears to not have refreshed the simple_rte_array (by calling
rebuild_simple_rel_and_rte) properly. For now, we fix it by using the
root->parse->rtable.
Signed-off-by: NFoyzur Rahman <foyzur@gmail.com>

Fixes issue #1310

Currently GPDB ignores NULL values when computing select array_agg(DISTINCT a).
This doesn't comply with the SQL-2008 standards (Refer
https://www.postgresql.org/message-id/e08cc0400911150005r82bc48du726d02b680c7920c%40mail.gmail.com
for context).
This was fixed in postgres 9.0 in 34d26872.
This commit brings GPDB output to match that of upstream postgres.

This fixes issue #427.

Signed-off-by: NNikos Armenatzoglou <nikos.armenatzoglou@gmail.com>

In commit ceea7d76, I remove some redundant tests, but fumbled on changing
the expected output, and left a stray "start_ignore" marker in place. That
made the test to fail.

The test queries that were supposed to check the relfilenodes after
operations, were all in ignore-blocks, so they weren't accomplishing much.
The not-ignored parts of these didn't do anything interesting.

We have very similar test (clearly copy-pasted at some point), in
qp_gist_indexes2 and qp_gist_indexes3.

We have the same in bitmap_index test already. Except for the SELECTs on
relfilenode, but those were all in ignore-block, so they wouldn't catch
anything anyway.

The 'ao_create_alter_valid_table' test in bugbuster was completely defunct,
as the expected output just contained an error about connecting to a
non-existent database. However, the exact same test queries were included
in the schema_topology test. Schema_toplogy is huge, and some of the tests
in it require a separate output file for ORCA, which these test queries do
not, so splitting it up into smaller parts seems like a good idea.

The exact same queries are present in qp_with_clause. Some of them have
slightly different GUCs set, but I doubt we're getting any meaningful extra
coverate by testing them with different GUC settings.

Also remove 'world', as it was only used to set up test tables for
'with_clause'

According to the comment, this was supposed to test COPY, but there's no
COPY to be seen. The created tables are rather uninteresting, so just remove
them.

To make it simpler and marginally faster.

Fix wording for gp_email_to GUC and clarify the expected separator
for the list (as it has changed).  Also use email rather than e-mail
to be consistent with the online Greenplum documentation.

EnumRelationId was incorrectly in the "These tables don't need to have
their OIDs synchronized" category. In fact, there was bespoken code to
synchronize them. But it's easy enough to rely on the generic mechanism,
so do so, and reduce our diff vs. upstream a little bit.

To stop the wide-spread abuse of namespaceOid for other things.

When a "null value in column <col> violates not-null constraint" error was
thrown from the ORCA translator code, it was propagated to the rest of the
system with ERRCODE_INTERNAL_ERROR. That's not nice, if an application
wants to e.g. catch errors based on the error code.

The reason to do this right now was that my previous commit that moved
some ORCA translator code in the source files made the regression tests
to fail. That's because an error handling code adds the source file name
and line number to the error message, like (CTranslatorUtils.cpp:2627),
and we had some such line numbers memorized in the regression tests.
When the code moved, the tests broke. That's obviously too fragile, and
using a proper error code fixes that by preventing the line number from
being printed. That also removed the last difference between the ORCA
and non-ORCA output from the 'insert' regression test.

We could probably do the same for other errors thrown from ORCA, but I'll
leave that for another day.

Commit 38686122 changed the function to not give a WARNING, but neglected
the comment that says it does.

Libraries like libssl libkrb5 are not included in the compile flags, so add those
missing libraries to make codegen unittest pass.

1. Remove TestPSQL from bugbuster
The whole test is completely enclosed in a start/end_ignore block, we can not
verify the correctness except it crash the whole service. Moreover,
allow_system_table_mods has been tested by other cases in ICG, so this test case
is useless for now.

2. Remove useless codecoverage case from bugbuster

3. Remove gp_persistent_table from bugbuster
gp_enable_inheritance is not invalid under current code base, so Remove this
case

4. merge DMLstmnt from bugbuster to icg

A bunch of functions and classes that are not used anywhere.

Ignore AccessMethodProcedureRelationId as pg_amproc do not need
OID synchronized between master and segments.

The NUM_EXTRA_OIDS_FOR_BITMAP macro was only used for generating a
set of extra relfilenodes for when running vacuuming a bitmap index.
With the new relfilenode handling in 88f0623e it's no loger required.

The test is failing with ORCA, because the expected output file hasn't been
updated for the spelling of "at OPEN". The regular expected output file was
updated in commit 05590cbf, but the ORCA-specific one was not.

There is no meaningful difference between the ORCA and non-ORCA outputs,
so just remove the alternative output file.

To test, backport test case for CREATE CAST from PostgreSQL 8.4. We don't
apparently have any coverage for it previously. The upstream commit for the
test case was:

commit a1d2e165
Author: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date:   Fri Oct 31 09:17:17 2008 +0000

    Add test case for CREATE CAST.

Reported by Yandong Yao

There was a race condition in the way relfilenodes were chosen, because
QE nodes chose relfilenodes for existing relations, e.g. at REINDEX or
TRUNCATE, independently of the master, while for newly created tables,
the relfilenode was chosen by the master. To fix:

1. If the OID of a newly-created table is already in use as relfilenode
   of a different table in a QE segment, use a different relfilenode.
   (This shouldn't happen in the dispatcher, for the same reasons it
   cannot happen in a single-server PostgreSQL instance)

2. Use a small cache of values recently used for a relfilenode, to close
   a race condition between checking if a relfilenode is in use, and
   actually creating the file

Signed-off-by: NNikos Armenatzoglou <nikos.armenatzoglou@gmail.com>

* Adding explicit casting for concatenation operator with integer
operands.
Signed-off-by: NFoyzur Rahman <foyzur@gmail.com>

* Explicit casting to convert integer to string to make opperf green.
Signed-off-by: NFoyzur Rahman <foyzur@gmail.com>

* Making expression tree mutation unconditional

We previously did not consider that we need to mutate an expression tree
only if there is a pseudo column. This introduces an executor crash in
non-assert build as executor does not know how to interpret a pseudo
column. Moreover, in assert build we fail an assert.

The conditional we are removing  was introduced to optimize performance
by commit a36436ea, but a quick performance
run shows that performance is not noticeably affected by removing the
optimization. [#131785597]
Signed-off-by: NFoyzur Rahman <foyzur@gmail.com>

* Adding PR comments.

syscache.h was previously included from parse_oper.h, but I removed that
in commit 8d77fb99. Turns out that gpdbwrappers.c was relying on that
indirect include.

Having an "extern" declaration directly in a C file is a bad habit, because
that can fail in surprising ways if the datatype of the variable is changed.
In thise case, we don't really need to check the variable at all, because
these functions are only used from ORCA.

Now that we capture and dispatch OIDs separately, we can revert these
changes vs. upstream.

I wasn't 100% sure if the change in bmbulkdelete() was OK. We no longer
update the relNode in the relcache entry. However, running this in a
debugger shows that reindex_index() updates the relcache entry itself,
so setting it in bmbulkdelete should not be necessary. Also, the
regression tests are happy with this.