MDBF-1076: Create MSAN Debug builder

[grooverdan]

Container prep work for new builder (and a general aarch64 rr capability)

On aarch64 g++-multilib isn't a package. The requirement for this package was to produce 32bit RR replay capability.

I don't think we've needed this capability so we can just remove this.

[grooverdan]

builds successful on native arm64.

[RazvanLiviuVarzaru]

Will replace amd64-msan-clang-20-debug ?

[grooverdan]

Will replace amd64-msan-clang-20-debug ?

That was my thought provided there's infra for it. Getting an auto-built working container is useful regardless.

[RazvanLiviuVarzaru]

Will replace amd64-msan-clang-20-debug ?

That was my thought provided there's infra for it. Getting an auto-built working container is useful regardless.

At first sight, there's room on our arm servers from Hetzner.
How much room, depends on you; if you want this builder to be part of Branch Protection, in which case, build speed and stability become critical and I need to look closer on resource consumption.

[grooverdan]

After this is merged can #823 be merged which will rebuild the MSAN again with MOTD updates.

[RazvanLiviuVarzaru]

10.11 - MTR nm 2h - not acceptable - https://buildbot.dev.mariadb.org/#/builders/535/builds/7
10.11 - after a3f2587 - https://buildbot.dev.mariadb.org/#/builders/535/builds/8
11.8 - MariaDB/server@ab6bcd8 not in 11.8 yet https://buildbot.dev.mariadb.org/#/builders/535/builds/6

@grooverdan Running with big-test increases the coverage in a debug , msan setup as opposed to the already defined big-test builder https://buildbot.mariadb.org/#/builders/914 or the big-test setup on the non-debug clang-20 -msan https://buildbot.mariadb.org/#/builders/867/builds/3842/steps/6/logs/stdio ?

Or what are the most time consuming tests that bring the smallest value in this setup?

For now I can't increase parallel over 20, which is a lot already and the running time is still over 45 mins only for the default MTR test suites.

[grooverdan]

From the run 7:

https://docs.google.com/spreadsheets/d/1ew284jTiPIUhN2DyXLkv5VfAOpc-jEwl61bfOo-zf8s

Taking the tests < 3 minutes for the 12 builders the idealistic time is 53 minutes - (row 6910).

The > 3 minute tests (57 of them) for the 12 builders ideally would be an extra 25 minutes (row 6968) but with the larger test ~10 minutes there would probably more.

Spreadsheet ideal split per worker (E6969) of 79 minutes scales up to 2h11min so a realty scaleup is 65% more that idealistic.

The run number 8 of 20 parallel looks about the same time run as the valgrind builder, but has the cover of bigtest + everything else that MSAN offers.

So options:

option	why	pro	con
get more hardware	with MDBF-143: Add infer #835, and MDBF-1115 create UBASAN Debug builder #839, and gaps of ps-protocol/cursor/view maybe we'll need it.	capacity building, lowish implementation	cost/sponsorship
disable perfschema in build	not so much the ~400 tests, the overhead in execution throughout the server	simple enough - exact gain unknown	drops in coverage
push a c,cxx_flags Og, O1, O2 or O3 into the build	its about testing the results rather than the most resolvable stack	simple unknown benefit	debugs from artifacts little harder, forks another custom build
disable bigtest	because its easy	easy	drops large coverage

Masking the ~20 5 minute + tests under Debug+MSAN isn't likely to fully push low enough on its own.