Orca's DP algorithms currently generate logical alternatives based only on cardinality; they do not take into account motions/partition selectors as these are physical properties handled later in the optimization process. Since DPv2 doesn't generate all possible alternatives for the optimization stage, we end up generating alternatives that do not support partition selection or can only place poor partition selectors.

This PR introduces partition knowledge into the DPv2 algorithm. If there is a possible partition selector, it will generate an alternative that considers it, in addition to the previous alternatives.

We introduce new properties, m_contain_PS  to indicate whether a SExpressionInfo contains a PS for a particular expression. We consider an expression to have a possible partition selector if the join expression columns and the partition table's partition key overlap. If they do, we mark this expression as containing a PS for a particular PT.

We consider a good PS one which is selective. Eg:
```
- DTS
- PS
   -TS
     - Pred
```

would be selective. However, if there is no selective predicate, we do not consider this as a promising PS.

For now, we add just a single alternative that satisfies this property and only consider linear trees.

This reverts commit ef887cfe.

This commit only affects cardinality estimation in ORCA when the user
sets `optimizer_damping_factor_join = 0`. It improves the square root
algorithm first introduced by commit ce453cf2.

In the original square root  algorithm, we assumed that distribution
column predicates would have some correlation with other predicates in
the join and therefore would be accordingly damped when calculating join
cardinality.

However, distribution columns are ideally unique in order to gain the
best performance for GPDB. Under this assumption, distribution columns
should not be correlated and thus needed to be treated as independent
when calculating join cardinality. This is a best guess since we do not
have a way to support correlated columns at this time.
Co-authored-by: NAshuka Xue <axue@vmware.com>
Co-authored-by: NChris Hajas <chajas@vmware.com>

We used to store them under /tmp/, we include the postmaster port number
in the file name in the hope that two clusters will not conflict with
each other on this file.

However the conflict still happen in the test src/bin/pg_basebackup.
And it can also happen if a second cluster is missed configured by
accident.  So to make things safe we also include the postmaster pid in
the domain socket path, there is no chance for two postmasters to share
the same pids.
Reviewed-by: NPaul Guo <pguo@pivotal.io>

This commit allows Orca to select plans that leverage IndexOnlyScan
node. A new GUC 'optimizer_enable_indexonlyscan' is used to enable or
disable this feature. Index only scan is disabled by default, until the
following issues are addressed:

  1) Implement cost comparison model for index only scans. Currently,
     cost is hard coded for testing purposes.
  2) Support index only scan using GiST and SP-GiST as allowed.
     Currently, code only supports index only scans on b-tree index.
Co-authored-by: NChris Hajas <chajas@vmware.com>

Add information about the compression parameters used when
a column is added to an AOCO table.

The libuv source is in the build container, rather thank in a gcp
bucket. This commit is a partial revert from getting source from a
gcp bucket and instead uses the libuv available in the build
container image.
Co-authored-by: NTyler Ramer <tramer@vmware.com>
Co-authored-by: NKris Macoskey <kmacoskey@vmware.com>
Co-authored-by: NJamie McAtamney <jmcatamney@vmware.com>

Here is the diff output of the test result.

 drop database some_database_without_tablespace;
 -DROP
 +ERROR:  database "some_database_without_tablespace" is being accessed by other users
 +DETAIL:  There is 1 other session using the database.
 drop tablespace some_basebackup_tablespace;
 -DROP
 +ERROR:  tablespace "some_basebackup_tablespace" is not empty

The reason is that after client connection to the database exits, the server
needs some time (the process might be scheduled out soon, and the operation
needs to content for the ProcArrayLock lock) to release the PGPROC in
proc_exit()->ProcArrayRemove(). During dropdb() (for database drop), postgres
will call CountOtherDBBackends() to see if there are still sessions that are
using the database by checking proc->databaseId, and it will try at most 5 sec.
This test quits the db connection of some_database_without_tablespace and then
drops the database immediately. This should be mostly fine but if the system is
in slow or in heavy load, this still could lead to test flakiness.

This issue could be simulated using gdb. Let's poll until database drop
commands succeeds for the affected database.  It seems that drop database sql
command could not be in transaction block so I could not use plpgsql to
implement, instead I use dropdb utility and bash command to implement that.
Reviewed-by: NAsim R P <pasim@vmware.com>

None of the mock tests in the backend call into GPORCA functions, so
no need to generate mock objects for them. Saves some time and space when
running mock tests.
Reviewed-by: NAshuka Xue <axue@pivotal.io>

Co-authored-by: NTyler Ramer <tramer@vmware.com>
Co-authored-by: NJamie McAtamney <jmcatamney@vmware.com>

Gphostcache has numerous issues, and has been a pain point for some
time. For this reason, we are removing it.

This commit moves the useful function of gphostcache - the hostname
deduping - to the gparray class, where a list of deduped hostnames is
returned from gpArray.get_hostlist().

There is a FIXME of correctly adding hostname to a newly added or
recovered mirror. The hostname resolution from address was incorrect and
faulty in its logic - an IP address never requires a hostname associated
with it. However, the hostname field in gp_segment_configuration should
be populated somehow - we recommend a "hostname" field addition to any
configuration files that require it. For now, we simple set the
"hostname" to "address" which ultimately delivers the same functionality
as the gphostcache implementation.
Co-authored-by: NTyler Ramer <tramer@vmware.com>
Co-authored-by: NJamie McAtamney <jmcatamney@vmware.com>

We skipped these during the PostgreSQL 9.6 merge. Time to pay down this
little technical debt.

We had copied the latest version of the t/002_pg_dump.pl test file from
upstream REL9_6_STABLE branch. That included the test change from upstream
commit e961341cc, but we didn't have the corresponding code changes yet.
Revert the test change, so that it passes. We'll get the bug fix, along
with the test change again, when we continue with the upstream merge, but
until then let's just keep the test in sync with the code.
Reviewed-by: NAshwin Agrawal <aashwin@vmware.com>

In some cases, merge stats logic for root partition table may consume
very high memory usage in CacheMemoryContext.
This may lead to `Canceling query because of high VMEM usage` when
concurrently ANALYZE partition tables.

For example, there are several root partition tables and they both have
thousands of leaf tables. And these tables are all wide tables that may
contain hundreds of columns.
So when analyze()/auto_stats() leaf tables concurrently,
`leaf_parts_analyzed` will consume lots of memory(catalog catch for
pg_statistic and pg_attribute) under
CacheMemoryContext for each backend, which may hit the protect VMEM
limit.
In `leaf_parts_analyzed`, a single backend's leaf table analysis for a
root partition table, it may add cache entries up to
number_of_leaf_tables * number_of_columns tuples from pg_statistic and
number_of_leaf_tables * number_of_columns tuples from pg_arrtibute.
Set guc `optimizer_analyze_root_partition` or
`optimizer_analyze_enable_merge_of_leaf_stats` to false could skip merge
stats for root table and `leaf_parts_analyzed` will not execute.

To resolve this issue:
1. When checking whether merge stats are available for a root table in
`leaf_parts_analyzed`, check whether all leaf tables are ANALYZEd first,
if they're still un-ANALYZE leaf table exists, return quickly to avoid touch
columns' pg_attribute and pg_statistic per leaf table(this will save lots of time).
And also don't rely on system catalog cache and use the
index to fetch the stats tuple to avoid one-time cache usage(in common cases).

2. When merging a stats in `merge_leaf_stats`, don't rely on system
catalog cache and use the index to fetch the stats tuple.

There are side-effects for not rely on system catalog cache(which are all **rare** situations).
1. If insert/update/copy several leaf tables which under **same
root partition** table in **same session** and all leaf tables are **analyzed**
will be much slower since auto_stats will call `leaf_parts_analyzed` once the leaf
table gets updated, and we don't rely on system catalog cache now.
(`set optimizer_analyze_enable_merge_of_leaf_stats=false` could avoid
this)

2. ANALYZE the same root table several times in the same session is much
slower than before since we don't rely on system catalog cache.

Seems this solution improves the performance for ANALYZE, and
it also makes ANALYZE won't hit the memory issue anymore.

We must install libuv on the PR pipeline to compile with ic-proxy
enabled.  ICW tests are still run in ic-udpifc mode.

Fix explain analyze error handling

We used to mark the GUC gp_interconnect_proxy_addresses as
PGC_POSTMASTER, so the cluster must be restarted to reload this setting,
this can be a problem during gpexpand: the cluster expansion itself is
online, but to configure the proxy addresses for the new segments a
restart is needed.

Now we changed it to PGC_SIGHUP, so the setting can be reloaded on
SIGHUP.

Also changed the setting from a developer option to a normal one.

The code will be compiled with ic-proxy enabled, but the tests are still
ran in the default ic-udpifc mode.
Authored-by: NDaniel Gustafsson <dgustafsson@pivotal.io>

Build gpdb binaries with --enable-ic-proxy, and include libuv in the
gpdb packages.

On an ARM64 machine, CREATE EXTENSION pxf_fdw fails with :
ERROR:  the log_errors option cannot be set without reject_limit

In pxf_fdw_validator(), the variable log_errors is declared bool, but
it is initialized with -1. Since bool seems to be now a built-in type,
it's definition is implementation dependent, and it's possible that on
ARM, it is defined as unsigned char. Because, through debugger it
could be seen that log_errors's value is 255 when it was assigned -1.
And because (log_errors != -1) condition returns true, we get the
error "log_errors option cannot be set without reject_limit" even when
log_errors option was not specified while creating the extension.
Due to this, all pxf_fdw tests were failing on ARM64.

If log_errors is specified, set log_errors to true rather than to
defGetBoolean(def) value. And rename log_errors to log_errors_set to
reflect that its purpose is not to store the log_errors value, but
rather to denote whether log_errors option was specified.

Enlarge the sleep time for a query which will be canceled later
to avoid slow execution fails the test.

Normally the test's running time should not get affected since the
sleep query will get terminate immediately.

During testing, I encountered an incremental gprecoverseg hang issue.
Incremental gprecoverseg is based on pg_rewind.  pg_rewind launches a single
mode postgres process and quits after crash recovery if the postgres instance
was not cleanly shut down - this is used to ensure that the postgres is in a
consistent state before doing incremental recovery. I found that the single
mode postgres hangs with the below stack.

\#1  0x00000000008cf2d6 in PGSemaphoreLock (sema=0x7f238274a4b0, interruptOK=1 '\001') at pg_sema.c:422
\#2  0x00000000009614ed in ProcSleep (locallock=0x2c783c0, lockMethodTable=0xddb140 <default_lockmethod>) at proc.c:1347
\#3  0x000000000095a0c1 in WaitOnLock (locallock=0x2c783c0, owner=0x2cbf950) at lock.c:1853
\#4  0x0000000000958e3a in LockAcquireExtended (locktag=0x7ffde826aa60, lockmode=3, sessionLock=0 '\000', dontWait=0 '\000', reportMemoryError=1 '\001', locallockp=0x0) at lock.c:1155
\#5  0x0000000000957e64 in LockAcquire (locktag=0x7ffde826aa60, lockmode=3, sessionLock=0 '\000', dontWait=0 '\000') at lock.c:700
\#6  0x000000000095728c in LockSharedObject (classid=1262, objid=1, objsubid=0, lockmode=3) at lmgr.c:939
\#7  0x0000000000b0152b in InitPostgres (in_dbname=0x2c769f0 "template1", dboid=0, username=0x2c59340 "gpadmin", out_dbname=0x0) at postinit.c:1019
\#8  0x000000000097b970 in PostgresMain (argc=5, argv=0x2c51990, dbname=0x2c769f0 "template1", username=0x2c59340 "gpadmin") at postgres.c:4820
\#9  0x00000000007dc432 in main (argc=5, argv=0x2c51990) at main.c:241

It tries to hold the lock for template1 on pg_database with lockmode 3 but
it conflicts with the lock with lockmode 5 which was held by a recovered dtx
transaction in startup RecoverPreparedTransactions(). Typically the dtx
transaction comes from "create database" (by default the template database is
template1).

Fixing this by using the postgres database for single mode postgres execution.
The postgres database is commonly used in many background worker backends like
dtx recovery, gdd and ftsprobe. With this change, we do not need to worry
about "create database" with template postgres, etc since they won't succeed,
thus avoid the lock conflict.

We may be able to fix this in InitPostgres() by bypassing the locking code in
single mode but the current fix seems to be safer.  Note InitPostgres()
locks/unlocks some other catalog tables also but almost all of them are using
lock mode 1 (except mode 3 pg_resqueuecapability per debugging output).  It
seems that it is not usual in real scenario to have a dtx transaction that
locks catalog with mode 8 which conflicts with mode 1.  If we encounter this
later we need to think out a better (might not be trivial) solution for this.
For now let's fix the issue we encountered at first.

Note in this patch the code fixes in buildMirrorSegments.py and twophase.c are
not related to this patch. They do not seem to be strict bugs but we'd better
fix them to avoid potential issues in the future.
Reviewed-by: NAshwin Agrawal <aashwin@vmware.com>
Reviewed-by: NAsim R P <pasim@vmware.com>

Previously it used max_prepared_xacts for shared snapshot slot number. The
reason that it does not use MaxBackends, per comment, is that ideally on QE we
want to use QD MaxBackends for the slot number, and note usually QE MaxBackends
should be greater than QD MaxBackends due to potential multiple gangs per
query. The code previously used max_prepared_xacts finally for the shared
snapshot slot number calculation. That is not correctly given we have read-only
query, and we have one-phase commit now.  Let's use MaxBackends for shared
snapshot slot number calculation for safety though this might waste some memory.
Reviewed-by: Nxiong-gang <gxiong@pivotal.io>

Previously we assign it as max_prepared_xacts. It is used to initialize some
2pc related shared memory. For example the array shmCommittedGxactArray is
created with this length and that array is used to collect not-yet "forgotten"
distributed transactions during master/standby recovery, but the array length
might be problematic since:

1. If master max_prepared_xacts is equal to segment max_prepared_xacts as
usual.  It is possible some distributed transactions use just partial gang so
the total distributed transactions might be larger (and even much larger) than
max_prepared_xacts. The document says max_prepared_xacts should be greater than
max_connections but there is no code to enforce that.

2. Also it is possible that master max_prepared_xacts might be different than
segment max_prepared_xacts (although the document does not suggest it there is
no code to enforce that),

To fix that we use MaxBackends for the gxact number on master. We may just use
guc max_connections (MaxBackends includes number for autovacuum workers and bg
workers additionally besides guc max_connections), but I'm conservatively using
MaxBackends,  since this issue is annoying - standby can not recover due to the
FATAL message as below even after postgres reboot unless we temporarily
increase the guc max_prepared_transactions value.

2020-07-17 16:48:19.178667
CST,,,p33652,th1972721600,,,,0,,,seg-1,,,,,"FATAL","XX000","the limit of 3
distributed transactions has been reached","It should not happen. Temporarily
increase max_connections (need postmaster reboot) on the postgres (master or
standby) to work around this issue and then report a bug",,,,"xlog redo at
0/C339BA0 for Transaction/DISTRIBUTED_COMMIT: distributed commit 2020-07-17
16:48:19.101832+08 gid = 1594975696-0000000009, gxid =
9",,0,,"cdbdtxrecovery.c",571,"Stack trace:

1    0xb3a30f postgres errstart (elog.c:558)
2    0xc3da4d postgres redoDistributedCommitRecord (cdbdtxrecovery.c:565)
3    0x564227 postgres <symbol not found> (xact.c:6942)
4    0x564671 postgres xact_redo (xact.c:7080)
5    0x56fee5 postgres StartupXLOG (xlog.c:7207)
Reviewed-by: Nxiong-gang <gxiong@pivotal.io>

We need that in more than one test.
Reviewed-by: Nxiong-gang <gxiong@pivotal.io>

ExecChooseHashTableSize() is a hot function which is not only called by executor,
but also by planner. Planner will call this function when calcualting cost for
each join path. The number of join path grow exponentially with the number of
table. As a result, do not using elog(LOG) to avoid generating too many logs.

Commit 362c48b6 added test
concurrent_drop_truncate_tablespace but it never made it to the
schedule.
Co-authored-by: NAlexandra Wang <lewang@pivotal.io>

Our implementations of memory pools have a hidden dependency on _the_
global memory pool manager: typically GPOS_NEW and GPOS_DELETE will
reach for the memory pool manager singleton. This makes GPOS_DELETE on a
memory pool manager undefined behavior because we call member functions
on an object after its destructor finishes.

On the Postgres 12 merge branch, this manifests itself in a crash during
initdb. More concerning is that it only crashed when we set max
connections and shared buffers to a specific number.

GPOS_ASMFP() used x86_64 assembly instructions to get current frame
address. This obviously doesn't compile on other architectures like
ARM64. So instead use __builtin_frame_address(), which is available
in gcc and presumably clang. Since gcc and clang are the two most
common compilers, and since we don't want to support GPORCA on exotic
architectures and compilers, don't bother to use any other way to get
the frame address.

Let configure fail if __builtin_frame_address() is not found, but
don't do this check if gporca is disabled.

GPORCA's CStackDescriptor::Backtrace() uses frame address. But there
is also gp_backtrace() in the backend code that has similar
functionality. This commit does not merge these two places. But it
prepares the infrastructure to do the merge, including a new macro
HAVE__BUILTIN_FRAME_ADDRESS defined in pg_config.h.

Discussion: https://groups.google.com/a/greenplum.org/forum/#!topic/gpdb-dev/FgaR_4sGYrkReviewed-by: NHeikki Linnakangas <hlinnakangas@pivotal.io>

* docs - update utility docs with IP/hostname information.

Add information to gpinitsystem, gpaddmirrors, and gpexpand ref. docs
--Information about using hostnames vs. IP addresses
--Information about configuring hosts that are configured with mulitple NICs

Also updated some examples in gpinitsystem

* docs - review comment updates. Add more information from dev.

* docs - change examples to show valid configurations that support failorver.
Also fix typos and minor edits.

* docs - updates based on review comments.

Add a bool flag 'delim_off' for CopyStateData to indicate if delimiter is set to OFF or not.

* docs - add information for SSL with standby master

--SSL file should not be in $MASTER_DATA_DIRECTORY

Also
--Add not about not using NULL ciphers
--Correct default directory for SSL files to $MASTER_DATA_DIRECTORY

* docs - review comment updates

That's how the locking works for all other tables, including inherited
tables, and for partitions in PostgreSQL v10 partitioning. But in GPDB,
we were intentionally releasing the lock too early, to save memory when
working on huge partition hierarchies I believe. But that's a lousy
tradeoff, we shouldn't skimp on safety just to save some memory. If
you run out of lock memory as a result, you need to bump up
max_locks_per_tranactions. Memory is cheap, and if you're working with
tens of thousands of partitions, you can afford reserving some memory for
the lock manager.

Fixes https://github.com/greenplum-db/gpdb/issues/5919Reviewed-by: NZhenghua Lyu <zlv@pivotal.io>
Reviewed-by: NAsim R P <pasim@vmware.com>

It was used to implement dedup plan which has been
refactored by the commit 9628a332.

So in this commit we remove these unused functions.

The failed test case is to test the command "copy lineitem to '/tmp/abort.csv'"
can be cancelled after COPY is dispatched to QEs. To verify this, it checks that
/tmp/abort.csv has fewer rows than lineitem.

The cancel logical in codes is:

QD dispatched the COPY command to QEs, then if QD get a cancel interrupt, it
sends a cancel request to QEs, however, the QD will keep receiving data from
QEs even QD already get a cancel interrupt. QD relies on QEs to receive the
cancel request and explicitly stop copying data to QD.

Obviously, QEs may already have copied out all data to QDs before they
get cancel requests, so the test case cannot guarantee /tmp/aborted.csv
has fewer rows than lineitem.

To fix this, we just verify the COPY command can be aborted with message
'ERROR:  canceling statement due to user request', the count
verification looks pointless here.