The repo https://github.com/greenplum-db/greenplum-database-release has
been changed default repo from master to main. This is to sync up this
change
Co-authored-by: NNing Wu <ningw@vmware.com>
Co-authored-by: NShaoqi Bai <bshaoqi@vmware.com>

Commit 4bbbb381 introduced some hardening
around concurrent drop and recreate of tables while analyzedb is running but it
failed to take into account the code around updating the last operation
performed. This commit fixes it.

The error was introduced by dc96f667.
If `set -u` was called before sourcing greenplum_path.sh with bash, an
error `ZSH_VERSION: unbound variable` would be reported.
To solve the issue, use shell syntax `{:-}` which will output an empty
value if the variable doesn't exist.

Tested with zsh, bash and dash.

On QD, it tracks whether QE wrote_xlog in the libpq connection.

The logic is, if QE writes xlog, it'll send a libpq msg to QD. But the
msg is sent in ReadyForQuery. So, before QE execute this function, the
QE may already send back results to QD. Then when QD process this
message, it does not read the new wrote_xlog value. This makes the
connection still contains the previous dispatch wrote_xlog value,
which will affect whether choosing one phase commit.

The issue only happens when the QE flush the libpq msg before the
ReadyForQuery function, hard to find a case to cover it.
I found the issue when I playing the code to send some information from
QE to QD. And it breaks the gangsize test which shows the commit info.

The generated greenplum_path.sh env file contained bash specific syntax
previously, so it errors out if the user's shell is zsh.

zsh doesn't have BASH_SOURCE. "${(%):-%x}" is the similar replacement
for zsh.
Also try to support other shells with some command combinations.
Tested with bash/zsh/dash.

* Docs - update interconnect proxy discussion to cover hostname support

* Change gp_interconnect_type -> gp_interconnect_proxy_addresses in note

Looking at GP documents, there is no indication that master dbid
must be 1. However, when CREATE_QD_DB, gpinitsystem always writes
"gp_dbid=1" into file `internal.auto.conf` even if we specify:

```
mdw~5432~/data/master/gpseg-1~2~-1
 OR
mdw~5432~/data/master/gpseg-1~0~-1
```

But catalog gp_segment_configuration can have the correct master
dbid value (2 or 0), the mismatch causes gpinitsystem hang.
Users can run into such problem for their first time to use
gpinitsystem -I.

Here we test dbid 0, because PostmasterMain() will simply check
dbid >= 0 (non-utility mode), it says:

> This value must be >= 0, or >= -1 in utility mode

It seems 0 is a valid value.

Changes:

- use specified master dbid field when CREATE_QD_DB.
- remove unused macros MASTER_DBID, InvalidDbid in C sources.
Reviewed-by: NAshwin Agrawal <aashwin@vmware.com>

If cdbcomponent_getCdbComponents() caught an error threw by
function getCdbComponents, FtsNotifyProber would be called.
But if it happened inside fts process, ftp process would hang.

Skip fts probe for fts process, after that, under the same
situation, fts process would exit and then be restarted by
postmaster.

Collect tuple relead pgstat table info from segments. Then the
auto-analyze could consider partition tables now. Since before, we don't
have accurate pgstat for partition leaf table. This kind of info is counted
through the access method on segments and we used to collect them by the
estate es_processed count on QD. So if insert into the root partition
table, we can not know how many tuples inserted into a leaf, autovacuum
never trigger auto-ANALYZE for leaf table.

The idea is, for writer QE, report current nest level xact tables pgstat
to QD through libpq at the end of a query statement. For a single
statement, it wouldn't operate too many tables, so the effort is really
small.
And on QD, retrieve and combine these tables' stat from the dispatch
result and add to current nest level xact pgstats.
Now we can remove the old pgstat collection code on QD.

The pgstat for a table could be view by query `pg_stat_all_tables_inernal`.
And now, except for the scan related counters, other counters should be accurate.
On master, the table's pgstat of scan related counters are not gathered
from segments yet, this requires extra work. The current implementation
is already enough for supporting auto-ANALYZE on partition table.
Reviewed-by: NHubert Zhang <hzhang@pivotal.io>

In some cases, some signals (like SIGQUIT) that should only be
processed by the main thread of the postmaster may be dispatched to rxThread.
So we should and it is safe to block all signals in the udp pthreads.

Fix #11006

We started hitting this on Thursday, and there's been ongoing report
from the community about this as well. While upstream is figuring out a
long term solution [1], we've been advised [2] to pin to the previous
release (v0.21.0) to avoid being blocked for hours at once.

[1]: https://github.com/telia-oss/github-pr-resource/pull/238
[2]: https://github.com/telia-oss/github-pr-resource/pull/238#issuecomment-714830491

It is often observed in CI that a test that leaves the cluster in an
inconsistent state (e.g. primary-mirror pair is not in sync, or
primary has not finished crash recovery) cause several following tests
in the schedule to fail.  What's worse, the culprit test itself may be
reported as passed because its validation criteria did not include the
state of the cluster.  This has found to mislead debugging efforts,
ultimately waste of time.

To make debugging CI failures more efficient, this patch enhances
pg_regress to perform the validation internally. If the validation
fails, further testing is aborted.

The validation is performed before running a group of tests specified
on one line in the schedule file. Also, validation is performed before
every single test, if running in serialized fashion.

When the cluster validation is found to fail, the culprit is in the
previously run test group.

This patch is built on the ground work and analysis laid out in
PR #9865 and PR #10825 by Wu Hao and Asim R P.
Reviewed-by: NAsim R P <pasim@vmware.com>

Previously, GPDB did not support the SQL "EXCLUDE [CURRENT ROW | GROUP |
TIES]" syntax in window functions. We got support for that from upstream
with the PostgreSQL v12 merge. That left the GUC obsolete and unused.

Update the 'olap_window' test accordingly. NOTE: 'olap_window' test isn't
currently run as part of the regression suite! I don't know why it's been
neglected like that, but that's not this patch's fault. The upstream
'window' test has queries with the EXCLUDE clause, so it's covered.

Reviewed-by: Jimmy Yih

GPCC want to hook query like
- create table ... as select ...
- create materialized view ... as select ...

If the $(UBUNTU_PLATFORM) is an empty string, the test command will fail,
double quotes it to fix.

```
--- mock for platform
/bin/sh: line 0: [: =: unary operator expected
```

These assertions started getting tripped in the previous commit when
adding tests, but aren't related to the Epsilon change. Rather, we're
calculating the frequency of a singleton bucket using two different
methods which causes this assertion to break down. The first method
(calculating the upper_third) assumes the singleton has 1 NDV and that there is an even distribution
across the NDVs. The second (in GetOverlapPercentage) calculates a
"resolution" that is based on Epsilon and assumes the bucket contains
some small Epsilon frequency. It results in the overlap percentage being
too high, instead it too should likely be based on the NDV.

In practice, this won't have much impact unless the NDV is very small.
Additionally, the conditional logic is based on the bounds, not
frequency. However, it would be good to align in the future so our
statistics calculations are simpler to understand and predictable.

For now, we'll remove the assertions and add a TODO. Once we align the
methods, we should add these assertions back.

When merging statistics buckets for UNION and UNION ALL queries
involving a column that maps to Double (eg: floats, numeric, time
related types), we could end up in an infinite loop. This occurred if
the bucket boundaries that we compared were within a very small value,
defined in Orca as Epsilon. While we considered that two values were
equal if they were within Epsilon, we didn't when computing whether
datum1 < datum2. Therefore we'd get into a situation where a datum
could be both equal to and less than another datum, which the logic
wasn't able to handle.

The fix is to make sure we have a hard boundary of when we consider a
datum less than another datum by including the epsilon logic in all
datum comparisons. Now, 2 datums are equal if they are within epsilon,
but datum1 is less than datum 2 only if datum1 < datum2 - epsilon.

Also add some tests since we didn't have any tests for types that mapped
to Double.

In commit c5f6dbbe, we changed the row and cost estimates on plan nodes
to represent per-segment costs. That made some estimates worse, because
the effects of the estimate "clamping" compounds. Per my comment on the
PR back then:

> One interesting effect of this change, that explains many of the
> plan changes: If you have a table with very few rows, or e.g. a qual
> like id = 123 that matches exactly one row, the Seq/Index Scan on it
> will be marked with rows=1. It now means that we estimate that every
> segment returns one row, although in reality, only one of them will
> return a row, and the rest will return nothing. That's because the
> row count estimates are "clamped" in the planner to at least
> 1. That's not a big deal on its own, but if you then have e.g. a
> Gather Motion on top of the Scan, the planner will estimate that the
> Gather Motion returns as many rows as there are segments. If you
> have e.g. 100 segments, that's relatively a big discrepancy, with
> 100 rows vs 1. I don't think that's a big problem in practice, I
> don't think most plans are very sensitive to that kind of a
> misestimate. What do you think?
>
> If we wanted to fix that, perhaps we should stop "clamping" the
> estimates to 1. I don't think there's any fundamental reason we need
> to do it. Perhaps clamp down to 1 / numsegments instead.

But I came up with a less intrusive idea, implemented in this commit:
Most Motion nodes have a "parent" RelOptInfo, and the RelOptInfo
contains an estimate of the total number of rows, before dividing it
with the number of segments or clamping. So if the row estimate we get
from the subpath seems clamped to 1.0, we look at the row estimate on
the underlying RelOptInfo instead, and use that if it's smaller. That
makes the row count estimates better for plans that fetch a single row
or a few rows, same as they were before commit c5f6dbbe. Not all
RelOptInfos have a row count estimate, and the subpaths estimate is
more accurate if the number of rows produced by the path differs from
the number of rows in the underlying relation, e.g.  because of a
ProjectSet node, so we still prefer the subpath's estimate if it
doesn't seem clamped.
Reviewed-by: NZhenghua Lyu <zlv@pivotal.io>

Otherwise it will raise an exception "command not run yet".

It was handled in expression_tree_mutator(), which is why everything
worked, but that's not the right place. expression_tree_mutator() is
supposed to handle nodes that can appear in expressions, and
plan_tree_mutator() is supposed to handle Plan nodes.
Reviewed-by: NZhenghua Lyu <zlv@pivotal.io>

In buildGangDefinition, newGangDefinition->db_descriptors are
initialized one by one, but newGangDefinition->size was already
set to its final value. If an error was caught, its size should
be reset to the right number.

The tests in commit be5d11e2 contained a typo that caused the changes
in the Scenario "gpstart starts even if the standby host is unreachable"
to not properly cleanup after itself.  Though the test feature still
passes, this leaves a bug to be found later when more tests are added.

Append-optimized tables do not contain transaction information in
their tuples.  Therefore, pg_class.relfrozenxid must remain invalid.
This is being done correctly during table creation, however, when the
table was rewritten, the relfrozenxid was accidentally set.  Fix it
such that diff with upstream is minimised.  In particular, the function
"should_have_valid_relfozenxid" is removed.

The fixme comments that led me to this bug are also removed.

Reviewed by: Ashwin Agrawal

Postgresql libpq document:

> Note that when PQconnectStart or PQconnectStartParams returns a
> non-null pointer, you must call PQfinish when you are finished
> with it, in order to dispose of the structure and any associated
> memory blocks. **This must be done even if the connection attempt
> fails or is abandoned**.

However, cdbconn_disconnect() function did not call PQfinish when
CONNECTION_BAD, it can cause socket leaks (CLOSE_WAIT state).

It didn't log the error message before if pg_rewind fails, fix that to make
DBA/field/developer's life eaisier.

Before this:
```
20201022:15:19:10:011118 gprecoverseg:earth:adam-[INFO]:-Running pg_rewind on required mirrors
20201022:15:19:12:011118 gprecoverseg:earth:adam-[WARNING]:-Incremental recovery failed for dbid 2. You must use gprecoverseg -F to recover the segment.
20201022:15:19:12:011118 gprecoverseg:earth:adam-[INFO]:-Starting mirrors
20201022:15:19:12:011118 gprecoverseg:earth:adam-[INFO]:-era is 0406b847bf226356_201022151031
```

After this:
```
20201022:15:33:31:019577 gprecoverseg:earth:adam-[INFO]:-Running pg_rewind on required mirrors
20201022:15:33:31:019577 gprecoverseg:earth:adam-[WARNING]:-pg_rewind: fatal: could not find common ancestor of the source and target cluster's timelines
20201022:15:33:31:019577 gprecoverseg:earth:adam-[WARNING]:-Incremental recovery failed for dbid 2. You must use gprecoverseg -F to recover the segment.
20201022:15:33:31:019577 gprecoverseg:earth:adam-[INFO]:-Starting mirrors
20201022:15:33:31:019577 gprecoverseg:earth:adam-[INFO]:-era is 0406b847bf226356_201022151031
```

This conforms our code to the best practice that each polymorphic
function should have exactly one of "virtual", "override", and "final"
specifier. I made this change using the following invocation:

clang-tidy -checks '-*,modernize-use-override'

Once we've made this change, ideally this practice can be enforced in
CI. We can just run clang-tidy with a single "modernize-use-override"
check to start with. Or we can see if our compilers are helpful enough.
Fortunately Clang already issues warnings (turned into errors by
-Werror) when we have inconsistent use of "override", and GCC appears to
have something similar (-Wsuggest-override) in versions 9.2 and later.

Now that we are explicitly declaring copy-assignment operators and copy
constructors as deleted, we should also make them public -- private and
=delete don't make much sense in combination, and this results in the
best diagnostics in practice [1].

In the previous commit where private unimplemented special member
functions are changed to be declared as deleted, we left the new
declarations private. This follows up by moving those deleted functions
to the public section of their classes.

I made this change guided by tooling: even though it doesn't offer a
fix, clang-tidy is useful enough to emit diagnostics that look like the
following:

/home/pivotal/workspace/gpdb/src/backend/gporca/libgpopt/include/gpopt/base/CColRef.h:76:2: warning: deleted member function should be public [modernize-use-equals-delete]
        CColRef(const CColRef &) = delete;
        ^

The process involved of making this change is a tale of much sorrow,
blood and tears. Suffice it to say 20 different regular expressions were
involved, and I'm edging closer to a PTSD when looking at backslashes.

References:
[1] https://abseil.io/tips/143#summary

This commit replaces two pre-C++ 11 practices with their modern, more
intent-expressive equivalents: "disallowed" copy and assignments, and
pairs of empty braces for special member function (destructors, default
constructors, and etc) bodies.

"Disallowed" copy constructors / assignment
-------------------------------------------
Old: private, unimplemented copy constructors and copy-assignment
operators in classes. These are usually paired with a semi-descriptive
comment.

New: publicly declare them as "= delete".

Fun fact: we had 13 spellings in comments on disallowed copy
constructors:

01. disable copy ctor
02. hidden copy ctor
03. inaccessible copy ctor
04. no copy ctor
05. no default copy ctor
06. private copy ctor
07. private no copy ctor
08. disabled copy constructor
09. no copy constructor
10. private copy constructor
11. copy c'tor - not defined
12. no copy c'tor
13. private copy c'tor

This commit removes all of them, because the new code ("~T() = delete")
already clearly expresses the intent of disallowing copy and assignment.

To keep the history clear, this commit leaves the declaration of
"prohibited" functions private. A forthcoming commit will wholesale
change them to public.

Defaulted special member functions
----------------------------------
Old:
struct A {
  A() {}
  ~A();
};
A::~A() {}

New:
struct A {
  A() = default;
  ~A();
};
A::~A() = default;

Replacing empty braces with defaulting not only makes them more clear,
they also enable more opportunities in compiler optimizations, as e.g.
some defaulted functions might be recognized as trivial.

Most of this commit is produced by running clang-tidy with an invocation
like the following (plus some CMake and shell tricks):

clang-tidy-12 -header-filter 'gpdbcost|gpopt|gpos|naucrates' -checks '-*,modernize-use-equals-delete,modernize-use-equals-default'

The tool uses a slightly conservative heuristic to detect a large
portion of the two outdated patterns above and rewrite them into using
"= delete" and "= default". Making the "= delete" functions public, is
sadly a FIXME item, so we'll have to do it by hand (in a forthcoming
commit).

The current setup in CMake (and Makefile's too, but that's an even
harder problem) leads to equivalent-but-not-identical header include
paths (-I) for the same directory, e.g. -Ilibgpopt/include vs
-Ilibgpdbcost/../libgpopt/include .

This confuses Clang-based tooling into identifying multiple paths for
the same header -- the difference being the extra sibling directory
followed by dot-dot, or "niece" directory followed by two levels of
dot-dot, and so forth -- as different headers. That, in turn, undermines
the conflict resolution and edit deduplication features in Clang's
refactoring engine, leading to duplicate edits when applying FixIt's.

This commit applies some fairly simple fixes to spell the sibling
directories in a way that generates consistent include paths, so that
Clang tooling are more functional. The Makefile's are left unchanged as
that is a lot more difficult to make "right". One can argue that we
_might_ want to instead transform the intermediate representation of
Clang's "replacement" YAML files, but that's left for another day.

During the development of a forthcoming commit that replaces private
unimplemented special member functions with explicitly deleted ones, we
got a surprising improvement of compiler diagnostics: Clang started to
see a lot of unused private fields that are masked by the fact that
there were unimplemented functions in the class. This makes sense
because the ostensibly unused field _could be_ used by a method whose
definition compiler hasn't seen in the current translation unit -- it
also suggests that the compiler would be better equipped at detecting
this if we had used whole-program analysis.

A sample of the errors looks like the following:

In file included from CMiniDumper.cpp:14:
In file included from ../../../../../../src/backend/gporca/libgpos/include/gpos/error/CErrorContext.h:17:
../../../../../../src/backend/gporca/libgpos/include/gpos/error/CMiniDumper.h:32:15: error: private field 'm_mp' is not used [-Werror,-Wunused-private-field]
        CMemoryPool *m_mp;
                     ^
1 error generated.

There are 12 such warnings, and this commit fixes them all. Note that
code mortality is a chain reaction: oftentimes, a variable (including
parameter) is only live because it's passed to another variable. While I
was at it, I also performed chained removal of variables and parameters
that became dead after removing the dead fields.

We have a bunch of classes with private fields that are used only in
debug builds, some of them are probably opportunities for complete
removal. This is exposed by an upcoming commit that enforces the use of
"=delete" and "=default" throughout the codebase.

I've attempted to solve this by adding the GPOS_ASSERTS_ONLY attribute,
but GCC doesn't like that, throwing errors like the following:

In file included from ../src/backend/gporca/libgpos/include/gpos/common/clibwrapper.h:22,
                 from ../src/backend/gporca/libgpos/include/gpos/error/CMessage.h:21,
                 from ../src/backend/gporca/libgpos/include/gpos/error/CMessageTable.h:14,
                 from ../src/backend/gporca/libgpos/include/gpos/error/CMessageRepository.h:14,
                 from ../src/backend/gporca/libgpos/src/_api.cpp:15:
../src/backend/gporca/libgpos/include/gpos/attributes.h:15:43: error: ‘unused’ attribute ignored [-Werror=attributes]
   15 | #define GPOS_UNUSED __attribute__((unused))
      |                                           ^
../src/backend/gporca/libgpos/include/gpos/attributes.h:21:27: note: in expansion of macro ‘GPOS_UNUSED’
   21 | #define GPOS_ASSERTS_ONLY GPOS_UNUSED
      |                           ^~~~~~~~~~~
../src/backend/gporca/libgpos/include/gpos/memory/CAutoMemoryPool.h:56:31: note: in expansion of macro ‘GPOS_ASSERTS_ONLY’
   56 |  ELeakCheck m_leak_check_type GPOS_ASSERTS_ONLY;
      |                               ^~~~~~~~~~~~~~~~~
cc1plus: all warnings being treated as errors

So we're back to the good ol' #ifdef GPOS_DEBUG.

This is the first of a pair of manual changes. In the immediately
following commit I'll remove the remaining (unconditionally) unused
private fields.

Seems resource group run_tests task didn't configure with Python
before and seems image doesn't have PYTHON, as changing the configure
default failed the job. Hence, explicitely adding --without-python to
retain old functionality with changed default.

This allows us to reduce code duplication of workload SQL scripts