On 18.02.23 21:26, Andres Freund wrote:
When building with meson, TEMP_CONFIG is supported for TAP tests, but doesn't
do anything for regress/isolation.
The reason for that is that meson's (and ninja's) architecture is to separate
"build setup" from the "build/test/whatever" stage, moving dynamism (and more
costly operations) to the "setup" phase.
In this case the implication is that the command line for the test isn't
re-computed dynamically. But pg_regress doesn't look at TEMP_CONFIG, it just
has a --temp-config=... parameter, that src/Makefile.global.in dynamically
adds if TEMP_CONFIG is set.
In contrast to that, TEMP_CONFIG support for tap tests is implemented in
Cluster.pm, and thus works transparently.
My inclination is to move TEMP_CONFIG support from the Makefile to
pg_regress.c. That way it's consistent across the build tools and isn't
duplicated. pg_regress already looks at a bunch of temporary variables
(e.g. PG_REGRESS_SOCK_DIR, PG_TEST_USE_UNIX_SOCKETS), so this isn't really
breaking new ground.
I'm having a hard time understanding what TEMP_CONFIG is for. It
appears that the intention is to allow injecting arbitrary configuration
into the tests? In that case, I think your proposal makes sense. But I
don't see this documented, so who knows what it is actually used for.
