Hi,
We have some issues with CI on macos and windows being too expensive (more on
that soon in a separate email), which reminded me of this thread (with
original title: [1])
I've attached a somewhat cleaned up version of the patch to cache initdb
across runs. The results are still fairly impressive in my opinion.
One thing I do not like, but don't have a good idea for how to improve, is
that there's a bunch of duplicated logic in pg_regress.c and Cluster.pm. I've
tried to move that into initdb.c itself, but that ends up pretty ugly, because
we need to be a lot more careful about checking whether options are compatible
etc. I've also thought about just putting this into a separate perl script,
but right now we still allow basic regression tests without perl being
available. So I concluded that for now just having the copies is the best
answer.
Times for running all tests under meson, on my workstation (20 cores / 40
threads):
cassert build -O2:
Before:
real 0m44.638s
user 7m58.780s
sys 2m48.773s
After:
real 0m38.938s
user 2m37.615s
sys 2m0.570s
cassert build -O0:
Before:
real 1m11.290s
user 13m9.817s
sys 2m54.946s
After:
real 1m2.959s
user 3m5.835s
sys 1m59.887s
non-cassert build:
Before:
real 0m34.579s
user 5m30.418s
sys 2m40.507s
After:
real 0m27.710s
user 2m20.644s
sys 1m55.770s
On CI this reduces the test times substantially:
Freebsd 8:51 -> 5:35
Debian w/ asan, autoconf 6:43 -> 4:55
Debian w/ alignmentsan, ubsan 4:02 -> 2:33
macos 5:07 -> 4:29
windows 10:21 -> 9:49
This is ignoring a bit of run-to-run variance, but the trend is obvious enough
that it's not worth worrying about that.
Greetings,
Andres Freund
[1] https://postgr.es/m/20220120021859.3zpsfqn4z7ob7afz%40alap3.anarazel.de