In a sequence, if NO CYCLE is specified, any calls to nextval after the sequence has reached its maximum value will return an error.

The order the rows were being aggregated was not well-defined, even though
I tried to add ORDER BYs to the queries. Putting the ORDER BY inside the
aggregate does the trick.

The current plan is to use something like pg_upgrade for future in-place
upgrades. The gpupgrade mechanism will not scale to the kind of drastic
catalog and other data directory layout changes that are coming as we
merge with later PostgreSQL releases.

Kept gpupgrademirror for now. Need to check if there's some logic that's
worth saving, for a possible pg_upgrade based solution later.

It was just syntax and catalogs, you couldn't actually do anything useful
with it. Remove it, so that we have less code to maintain, until it's time
to merge this stuff from upstream again when we merge with PostgreSQL 8.4.
It's probably easier to merge this back at that point than maintain this
backported version in the meanwhile. Less effort now, until we reach that
point, and once we get to the point in 8.4 that we merge this in, we'll
have all the preceding patches applied already, so it should merge quite
smoothly.

Move gpfdist regress cases 'exttab1' and 'custom_format' to extension/gpfdist
directory.

Enhance pg_regress to call needed perl scripts from correct directory.

This splits pg_proc.h into two parts: pg_proc.h and pg_proc_gp.h. After the
split, pg_proc.h contains functions we've inherited from upstream, and is
mostly identical to the upstream version. This makes diffing and merging
with upstream easier. All the upstream functions have been removed from
pg_proc.sql, as they are no longer auto-generated. Instead of modifying
pg_proc.h, catullus.pl is now used to create pg_proc_gp.h, which contains
all the auto-generated DATA entries for the functions in pg_proc.sql.

For the upstream-inherited DATA entries in pg_proc.h, remove the columns
that are not present in upstream. This was needed again to make pg_proc.h
closer to the upstream version. Also, remove auto-generated comments and
revert whitespace changes compared to upstream.

To compensate for the missing columns, introduce a new perl script,
processs_col_defaults.pl that is run at build time, which adds the missing
columns back to any DATA rows that miss them, before they are fed to
genbki.sh. All the GPDB-added columns are annotated with a magic
GPDB_COLUMN_DEFAULT() macro, similar to the CATALOG() and DATA() macros
which tells process_col_defaults.pl what values to insert into the missing
columns.

Remove tidycat2 regression test. It tested that all the functions in
pg_catalog came from pg_proc.sql, but that is no longer true. Only the
GPDB-added ones are now in pg_proc.sql.

Copy with no errors in input file, not reusing any existing (error)
table and not specifying KEEP option in LOG ERRORS INTO clause,
will cause error table drop once done.

This reduces the diff footprint with upstream, which will make merging
easier.

I don't see any reason for adding the unique2 column to ORDER BYs. I think
this must've been some kind of a misunderstanding of the test case years
ago, and before gpdiff.pl was invented to mask out differences in row
ordering. It was a misunderstanding even back then, though, because the row
order of these queries is well-defined even without the unique2 column.

This will make merging and diffing with upstream easier.

There seems to have been a philosophy for bugbuster tests to not only clean
up after test, but also try to drop any object that the test creates before
the test, so that the test case works if it is interrupted, and you re-run
it. We don't generally require that for test cases, as pg_regress always
creates a fresh 'regression' database to run the tests in.

In any case, dropping and recreating PL/pgSQL certainly doesn't seem
appropriate, because it's installed by default.

Move the GPDB-added parts of the 'case' regression test, and the bugbuster
'case' regression test, into a new test called 'case_gp'. This way the
'case' test is almost identical to upstream version, which will ease
diffing and merging. And there's not much point keeping the bugbuster test
separate.

It wasn't actually testing gpcrondump, as the test name implied. It created
a couple of partitioned tables, inserted rows in them, and created indexes
on them. That's all. That functionality is already covered by the
'partition' test case.

We have "LIMIT 0" clauses in other tests.

It wasn't testing anything interesting. The comments described possibly
useful test scenarios, involving concurrent transactions, but the actual
queries all happened in a single backend. I think we're better off just
removing these.

All it did was some setup for the 'optquery' test, so might as well make
'optquery' self-contained, by moving the initialization there.

They were just replaced with the constants, so might as well use the
constants directly in the tests.

After creating view that contain ordered-set (WITHIN GROUP) aggregates, \d+ or pg_dump can not get the correct definition of the view back, the brackets around the ORDER BY clause disappear.

View definition is stored in pg_rewrite in the form of parsetree, get_rule_expr() parses the T_Percentile node back to query string format, it should append the brackets around ORDER BY clause.

Also, rephrase the message you get from enable_xform() etc. functions, if
the server has been built without ORCA.

Add alternative expected output files to make the regression tests pass
without --enable-orca.

It was not working correctly anyway on 32-bit x86 systems, where
sizeof(int) < sizeof(Datum), because in Greenplum sizeof(Datum) == 8
even on 32-bit platforms. This was revealed by failures on the 'misc'
regression test, which tested this the V0 support.

The 'gp_optimizer' test also creates a table called 'sales', which started
to clash with the 'sales' table created by the new 'decode_expr' test.
Rename the test table used in decode_expr, and also drop it at the end of
the test.

Not sure why ORCA causes the error, but it seems to be intended behaviour,
so I'm not going to try fixing that right now.

The gphdfs test was concerned with testing gphdfs. It's useless without
gphdfs itself, and was causing the main regression suite to fail, if gphdfs
was not installed. Likewise, move the gphdfs-related parts of exttab1 test.

Looks like all the 'functional' test was testing was DECODE() expressions,
so rename the test, and move it to the main test suite. While we're at it,
clean up the test case to some extent. Fix the missing semicolons from some
queries, which caused the subsequent DROP to fail. Remove unused functions
and emp and dept tables. Add some comments explaining what each test does.
Remove use of multi-byte characters from test case, so that it works
regardless of encoding.

Remove unnecessary ORDER BY clauses that are not present in the upstream
version. Disable COPY BINARY test, as GPDB doesn't support BINARY copy.
Finally, adjust the error messages in the expected output to match what
you get nowadays.

DROP TABLE cleanup commands have been added to some other tests, so that
the extra user tables don't appear in the output of the 'misc' test. It's
not very nice of the 'misc' test to list all user tables, but this
nevertheless seems less likely to cause merge conflicts than ripping out
that part of the test altogether. There are still differences in the list
of user tables though.

Revert some unnecessary changes compared to upstream: there's no need for
ORDER BYs in queries, gpdiff.pl will mask out row order differences. That
needed some fixes to atmsort.pl though, to make it smarter about reordering
COPY TO STDOUT results. It used to only deal with the COPY (SELECT ...) TO
STDOUT variant, and particularly named tables (COPY .*test1 TO STDOUT). Make
it handle all COPY TO STDOUT commands. Also make it smarter about detecting
the end of COPY TO STDOUT output. In addition to -- or ERROR, also treat any
SELECT, UPDATE, CREATE etc. command as the end of result. In the passing,
remove the "-- copy_stdout" command from atmsort.pl, as it was unused.

No need to label the functions as immutable or with NO SQL or MODIFIES SQL
DATA, and update the error message in the expected output about
cross-segment access within functions.

One of the tests gives a different error than on upstream: an empty line
gives "invalid integer" error in upstream, but "missing data" on GPDB. I'm
not quite sure what's causing that, but both errors seem reasonable for that
case (the "missing data" is perhaps even better), so let's just memorize
that difference.

The upstream 'copy2' regression test exercised this, but we haven't noticed
because we've disabled that test. I spotted this while trying to re-enable
it. Nevertheless, let's add an explicit test for this.

Integers and numerics don't need quoting in CSV-format COPY, but they should
still be quoted if the FORCE QUOTE option is used.

We had memorized the incorrect output as expected output of the GPDB-added
parts at the end of the 'copy' regression test. Fix that, but also add
an explicit test case for this.

The shell on my laptop produces an error like:

ERROR:  external table ext_stderr1 command ended with error. sh: 1: ...

while the expected output is:

ERROR:  external table ext_stderr1 command ended with error. sh: ...

(Note the missing line number, "1: ").

Add a substitution pattern to the test case to mask that out. There was
already a similar pattern for shells that print "line 1:".

Many of the test queries were doubled: run first within an ignore-block,
and then for real. That seems pointless, so remove the duplicates. Also,
fix typos in comments.

This hopefully fixes test failures that we've occasionally seen in the
buildfarm, where the mechanism in gpdiff.pl that masks out result set
ordering differences gets confused by the duplicated queries in the
ignore-blocks.

Exchanging a default partition is currently not allowed because there is
no validation on the data being exchanged. This commit add a GUC to
enable this if the user chooses to do so.

operator strategy numbers, ie, GiST and GIN.  This is almost cosmetic
enough to not need a catversion bump, but since the opr_sanity regression
test has to change in sync with the catalog entry, I figured I'd better
do one.

identify long-running transactions. Since we already need to record
the transaction-start time (e.g. for now()), we don't need any
additional system calls to report this information.

Catversion bumped, initdb required.

vacuum/analyze timestamp columns at the end, rather than at a random
spot in the middle as in the original patch.  This was deemed more usable
as well as less likely to break existing application code.  initdb forced
accordingly.  In passing, remove former kluge for initializing
pg_stat_file()'s pg_proc entry --- bootstrap mode was fixed recently
so that this can be done without any hacks, but I overlooked this usage.