Horton

Thread

Commits

  1. Add support for NO_INSTALLCHECK in MSVC scripts

  2. Add missing NO_INSTALLCHECK in commit_ts and test_rls_hooks

  1. Handling of REGRESS_OPTS in MSVC for regression tests

    Michael Paquier <michael@paquier.xyz> — 2018-11-26T05:43:02Z

    Hi all,

As mentioned on this thread, I have been fighting a bit with the
buildfarm when trying to introduce new PGXS options for isolation and
TAP tests:
https://www.postgresql.org/message-id/E1gR4D0-0002Yj-Jw@gemulon.postgresql.org

However it happens that we have more issues than the buildfarm failures
to begin with.  One of them it related to the way REGRESS_OPTS is
defined and handled across the whole tree for MSVC.

There are in the tree now a couple of paths using top_builddir or
top_srcdir:
$ git grep REGRESS_OPTS | grep top
contrib/dblink/Makefile:REGRESS_OPTS =
--dlpath=$(top_builddir)/src/test/regress
contrib/pg_stat_statements/Makefile:REGRESS_OPTS = --temp-config
$(top_srcdir)/contrib/pg_stat_statements/pg_stat_statements.conf 
src/test/modules/commit_ts/Makefile:REGRESS_OPTS =
--temp-config=$(top_srcdir)/src/test/modules/commit_ts/commit_ts.conf
src/test/modules/test_rls_hooks/Makefile:REGRESS_OPTS =
--temp-config=$(top_srcdir)/src/test/modules/test_rls_hooks/rls_hooks.conf

Using top_builddir for dblink is for example fine as this needs to point
out to the place where builds of pg_regress are present.  However when
it comes to configuration files we should use top_builddir, and ought to
use top_srcdir instead as VPATH would complain about things gone
missing.  So the definition that we make out of it is correct.  However
there is an issue with MSVC which thinks that REGRESS_OPTS should only
include top_builddir.  This is actually fine per-se as all MSVC tests
run with make-installcheck, and not make-check, however I think that we
should allow top_srcdir to be switched on-the-fly as much as
top_builddir as top_srcdir could be used for something else.

Another way worse issue is that MSVC scripts ignore some configuration
values because of the way values of REGRESS_OPTS are parsed.  This
causes some sets of regression tests to never run on Windows.  For
example, here is what the command generated for pg_stat_statements, and
what happens with it:
c:/pgbuildfarm/pgbuildroot/HEAD/pgsql.build/Release/pg_regress/pg_regress \
  --bindir=c:/pgbuildfarm/pgbuildroot/HEAD/pgsql.build/Release/psql \
  --dbname=contrib_regression --temp-config pg_stat_statements
[...]
============== running regression test queries        ==============

=====================
 All 0 tests passed.
=====================

This causes the tests to pass, but to run nothing as test list is eaten
out as an option value for --temp-config.  In this case, what happens is
that --temp-config is set to "pg_stat_statements", leaving the test list 
empty.  The same cannot happen with test_decoding as the buildfarm uses
a custom module to run its tests (still this module could go away with
PGXS options to control isolation tests).

Attached is a patch which makes MSVC scripts more intelligent by being
able to replace top_srcdir as well by topdir when fetching options.

Thanks to that, MSVC is able to run the tests correctly, by building a
proper command.  I think that this is a bug that should be back-patched,
because the tests just don't run, and we likely would miss regressions
because of that.

Unfortunately, if I were to commit that directly, the buildfarm would
turn immediately to red for all Windows members, because when checking
module and contrib tests an "installcheck" happens, and
shared_preload_libraries is not set in this case.  When working on the
switch to add isolation and TAP test support to PGXS, test_decoding has
been failing because the installed server did not have wal_level =
logical set up, which is one instance of that.  The buildfarm code
installing the Postgres instance on which the tests are run should
configure that properly I think, and one tweak that we could use for the
time being is to bypass tests of modules which cannot work yet, so as
the buildfarm keeps being green.  This way, this would not be a blocker
for the integration of the new PGXS infra for TAP and isolation.  And
those could be enabled back once the buildfarm code is able to set up a
proper configuration.  The following modules require some extra
configuration:
- commit_ts
- test_rls_hooks
- pg_stat_statements
- test_decoding (if its Makefile is rewritten)
- snapshot_too_old (if its Makefile is rewritten)

The buildfarm part requires Andrew Dunstan's input, and there is a bit
to sort out for the rest, so input from all is welcome.  Please note
that I would prefer if the tests which just cannot work yet are
disabled until the needed buildfarm infrastructure is needed.  Another
option could also be to switch contribcheck and modulescheck so as they
use a temporary installation, but that's a can of worms waiting to
explode as MSVC makes more complicated the search for initdb and such
(see the way upgradecheck happens for example), and this makes the tests
waaay longer to run.

Thoughts?
--
Michael

  2. Re: Handling of REGRESS_OPTS in MSVC for regression tests

    Michael Paquier <michael@paquier.xyz> — 2018-11-26T06:48:07Z

    On Mon, Nov 26, 2018 at 02:43:02PM +0900, Michael Paquier wrote:
> Another option could also be to switch contribcheck and modulescheck
> so as they use a temporary installation, but that's a can of worms
> waiting to explode as MSVC makes more complicated the search for
> initdb and such (see the way upgradecheck happens for example), and
> this makes the tests waaay longer to run.

This has been itching me, and actually it proves to not be that
complicated to achieve per the attached.  This makes all the tests from
contrib/ and src/test/modules pass with temporary installations, the
tests runs are much slower though.  This would not blow up the buildfarm
visibly, buts its code assumes that installcheck should be used, so we
could as well just introduce new options for vcregress.pl.  I am not
sure what would be the best way, still using temporary installations has
the merit to not cause any tests to run unconfigured in the future.
--
Michael

  3. Re: Handling of REGRESS_OPTS in MSVC for regression tests

    Andrew Dunstan <andrew.dunstan@2ndquadrant.com> — 2018-11-26T17:41:21Z

    On 11/26/18 12:43 AM, Michael Paquier wrote:
> Hi all,
>
> As mentioned on this thread, I have been fighting a bit with the
> buildfarm when trying to introduce new PGXS options for isolation and
> TAP tests:
> https://www.postgresql.org/message-id/E1gR4D0-0002Yj-Jw@gemulon.postgresql.org
>
> However it happens that we have more issues than the buildfarm failures
> to begin with.  One of them it related to the way REGRESS_OPTS is
> defined and handled across the whole tree for MSVC.
>
> There are in the tree now a couple of paths using top_builddir or
> top_srcdir:
> $ git grep REGRESS_OPTS | grep top
> contrib/dblink/Makefile:REGRESS_OPTS =
> --dlpath=$(top_builddir)/src/test/regress
> contrib/pg_stat_statements/Makefile:REGRESS_OPTS = --temp-config
> $(top_srcdir)/contrib/pg_stat_statements/pg_stat_statements.conf
> src/test/modules/commit_ts/Makefile:REGRESS_OPTS =
> --temp-config=$(top_srcdir)/src/test/modules/commit_ts/commit_ts.conf
> src/test/modules/test_rls_hooks/Makefile:REGRESS_OPTS =
> --temp-config=$(top_srcdir)/src/test/modules/test_rls_hooks/rls_hooks.conf
>
> Using top_builddir for dblink is for example fine as this needs to point
> out to the place where builds of pg_regress are present.  However when
> it comes to configuration files we should use top_builddir, and ought to
> use top_srcdir instead as VPATH would complain about things gone
> missing.  So the definition that we make out of it is correct.  However
> there is an issue with MSVC which thinks that REGRESS_OPTS should only
> include top_builddir.  This is actually fine per-se as all MSVC tests
> run with make-installcheck, and not make-check, however I think that we
> should allow top_srcdir to be switched on-the-fly as much as
> top_builddir as top_srcdir could be used for something else.
>
> Another way worse issue is that MSVC scripts ignore some configuration
> values because of the way values of REGRESS_OPTS are parsed.  This
> causes some sets of regression tests to never run on Windows.  For
> example, here is what the command generated for pg_stat_statements, and
> what happens with it:
> c:/pgbuildfarm/pgbuildroot/HEAD/pgsql.build/Release/pg_regress/pg_regress \
>    --bindir=c:/pgbuildfarm/pgbuildroot/HEAD/pgsql.build/Release/psql \
>    --dbname=contrib_regression --temp-config pg_stat_statements
> [...]
> ============== running regression test queries        ==============
>
> =====================
>   All 0 tests passed.
> =====================
>
> This causes the tests to pass, but to run nothing as test list is eaten
> out as an option value for --temp-config.  In this case, what happens is
> that --temp-config is set to "pg_stat_statements", leaving the test list
> empty.  The same cannot happen with test_decoding as the buildfarm uses
> a custom module to run its tests (still this module could go away with
> PGXS options to control isolation tests).
>
> Attached is a patch which makes MSVC scripts more intelligent by being
> able to replace top_srcdir as well by topdir when fetching options.
>
> Thanks to that, MSVC is able to run the tests correctly, by building a
> proper command.  I think that this is a bug that should be back-patched,
> because the tests just don't run, and we likely would miss regressions
> because of that.
>
> Unfortunately, if I were to commit that directly, the buildfarm would
> turn immediately to red for all Windows members, because when checking
> module and contrib tests an "installcheck" happens, and
> shared_preload_libraries is not set in this case.  When working on the
> switch to add isolation and TAP test support to PGXS, test_decoding has
> been failing because the installed server did not have wal_level =
> logical set up, which is one instance of that.  The buildfarm code
> installing the Postgres instance on which the tests are run should
> configure that properly I think, and one tweak that we could use for the
> time being is to bypass tests of modules which cannot work yet, so as
> the buildfarm keeps being green.  This way, this would not be a blocker
> for the integration of the new PGXS infra for TAP and isolation.  And
> those could be enabled back once the buildfarm code is able to set up a
> proper configuration.  The following modules require some extra
> configuration:
> - commit_ts
> - test_rls_hooks
> - pg_stat_statements
> - test_decoding (if its Makefile is rewritten)
> - snapshot_too_old (if its Makefile is rewritten)
>
> The buildfarm part requires Andrew Dunstan's input, and there is a bit
> to sort out for the rest, so input from all is welcome.  Please note
> that I would prefer if the tests which just cannot work yet are
> disabled until the needed buildfarm infrastructure is needed.  Another
> option could also be to switch contribcheck and modulescheck so as they
> use a temporary installation, but that's a can of worms waiting to
> explode as MSVC makes more complicated the search for initdb and such
> (see the way upgradecheck happens for example), and this makes the tests
> waaay longer to run.
>

It's not the buildfarm that is broken. Both contribcheck() and 
modulescheck() in vcregress.pl run in installcheck mode, i.e. with an 
already running database. That makes the tests run faster, because 
setting up and breaking down instances is expensive, especially on 
Windows. We've got far too profligate with that in recent years, to the 
extent that some of my Windows buildfarm animals take many hours to 
complete full runs these days.

Note that TAP tests are not a problem here. It's up to the TAP scripts 
to set up and break down postgres instances they need, with whatever 
config options are required, such as shared preload libraries. It's only 
modules using pg_regress that could be a problem.

A simple way to proceed might be to have vcregress.pl honor the 
NO_INSTALLCHECK flag. Currently this is only used by pg_stat_statements, 
but we could add it anywhere required. In vcregress,pl, I would probably 
do this by having fetchTests() check for the flag and return an empty 
set of tests if it's found, and in turn have subdircheck() do nothing if 
it has an empty set of tests.


cheers


andrew


-- 
Andrew Dunstan                https://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

  4. Re: Handling of REGRESS_OPTS in MSVC for regression tests

    Michael Paquier <michael@paquier.xyz> — 2018-11-27T01:27:17Z

    On Mon, Nov 26, 2018 at 12:41:21PM -0500, Andrew Dunstan wrote:
> It's not the buildfarm that is broken. Both contribcheck() and
> modulescheck() in vcregress.pl run in installcheck mode, i.e. with an
> already running database. That makes the tests run faster, because setting
> up and breaking down instances is expensive, especially on Windows. We've
> got far too profligate with that in recent years, to the extent that some of
> my Windows buildfarm animals take many hours to complete full runs these
> days.

I have noticed the performance impact when switching to check for those
targets.  It was bad.

> Note that TAP tests are not a problem here. It's up to the TAP scripts to
> set up and break down postgres instances they need, with whatever config
> options are required, such as shared preload libraries. It's only modules
> using pg_regress that could be a problem.

Indeed.

> A simple way to proceed might be to have vcregress.pl honor the
> NO_INSTALLCHECK flag. Currently this is only used by pg_stat_statements, but
> we could add it anywhere required. In vcregress.pl, I would probably do this
> by having fetchTests() check for the flag and return an empty set of tests
> if it's found, and in turn have subdircheck() do nothing if it has an empty
> set of tests.

Okay, let's do so by supporting correctly NO_INSTALLCHECK.  My other
refactoring work can also live with that.  Returning an empty list via
fetchTests() and bypass a run if nothing is present looks fine to me.
One extra thing is that modules/commit_ts and modules/test_rls_hooks are
missing NO_INSTALLCHECK, so we would need to add that part to be
completely correct.  I would also be inclined to back-patch both parts,
however for my stuff I could live with this patch only on HEAD, and
anybody willing to use installcheck on commit_ts and test_rls_hooks may
be surprised to not be able to do that anymore after the minor release.
It still looks incorrect to me though to be able to run installcheck on
those modules though...

Attached is a proposal of patch, which works as I would expect with
modulescheck and contribcheck.  How does that look?
--
Michael

  5. Re: Handling of REGRESS_OPTS in MSVC for regression tests

    Michael Paquier <michael@paquier.xyz> — 2018-11-27T21:10:49Z

    On Tue, Nov 27, 2018 at 10:27:17AM +0900, Michael Paquier wrote:
> Okay, let's do so by supporting correctly NO_INSTALLCHECK.  My other
> refactoring work can also live with that.  Returning an empty list via
> fetchTests() and bypass a run if nothing is present looks fine to me.
> One extra thing is that modules/commit_ts and modules/test_rls_hooks are
> missing NO_INSTALLCHECK, so we would need to add that part to be
> completely correct.  I would also be inclined to back-patch both parts,
> however for my stuff I could live with this patch only on HEAD, and
> anybody willing to use installcheck on commit_ts and test_rls_hooks may
> be surprised to not be able to do that anymore after the minor release.
> It still looks incorrect to me though to be able to run installcheck on
> those modules though...
> 
> Attached is a proposal of patch, which works as I would expect with
> modulescheck and contribcheck.  How does that look?

If possible, I would like to move on with this stuff.  To keep things
green in the buildfarm all the time, I would like to do that with two
independent steps:
1) Add NO_INSTALLCHECK to modules/commit_ts and modules/test_rls_hook.
2) Add support for NO_INSTALLCHECK in the MSVC scripts.

Are there any objections?  I could do a back-patch as well to keep
things consistent across branches if there are voices in favor of that,
but that's not necessary for the upcoming Makefile cleanup with the new
set of PGXS options.
--
Michael

  6. Re: Handling of REGRESS_OPTS in MSVC for regression tests

    Andrew Dunstan <andrew.dunstan@2ndquadrant.com> — 2018-11-27T22:59:34Z

    On 11/27/18 4:10 PM, Michael Paquier wrote:
> On Tue, Nov 27, 2018 at 10:27:17AM +0900, Michael Paquier wrote:
>> Okay, let's do so by supporting correctly NO_INSTALLCHECK.  My other
>> refactoring work can also live with that.  Returning an empty list via
>> fetchTests() and bypass a run if nothing is present looks fine to me.
>> One extra thing is that modules/commit_ts and modules/test_rls_hooks are
>> missing NO_INSTALLCHECK, so we would need to add that part to be
>> completely correct.  I would also be inclined to back-patch both parts,
>> however for my stuff I could live with this patch only on HEAD, and
>> anybody willing to use installcheck on commit_ts and test_rls_hooks may
>> be surprised to not be able to do that anymore after the minor release.
>> It still looks incorrect to me though to be able to run installcheck on
>> those modules though...
>>
>> Attached is a proposal of patch, which works as I would expect with
>> modulescheck and contribcheck.  How does that look?
> If possible, I would like to move on with this stuff.  To keep things
> green in the buildfarm all the time, I would like to do that with two
> independent steps:
> 1) Add NO_INSTALLCHECK to modules/commit_ts and modules/test_rls_hook.
> 2) Add support for NO_INSTALLCHECK in the MSVC scripts.
>
> Are there any objections?  I could do a back-patch as well to keep
> things consistent across branches if there are voices in favor of that,
> but that's not necessary for the upcoming Makefile cleanup with the new
> set of PGXS options.


I think you should just proceed with the changes above. I just had a 
quick look at the patch you posted before, and it looks sane enough.


I don't see a need to backpatch.


cheers


andrew

-- 
Andrew Dunstan                https://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

  7. Re: Handling of REGRESS_OPTS in MSVC for regression tests

    Michael Paquier <michael@paquier.xyz> — 2018-11-28T01:00:00Z

    On Tue, Nov 27, 2018 at 05:59:34PM -0500, Andrew Dunstan wrote:
> I think you should just proceed with the changes above. I just had a quick
> look at the patch you posted before, and it looks sane enough.

Thanks for the feedback, Andrew.  Let's wait a couple of days and see if
anybody has any objection about the patch set.

> I don't see a need to backpatch.

Okay, no problem with that.
--
Michael

  8. Re: Handling of REGRESS_OPTS in MSVC for regression tests

    Michael Paquier <michael@paquier.xyz> — 2018-11-29T01:33:18Z

    On Wed, Nov 28, 2018 at 10:00:00AM +0900, Michael Paquier wrote:
> Okay, no problem with that.

Both things have been committed, but not back-patched.
--
Michael