Re: cirrus-ci cross-build interactions?

Hi,

On 2022-09-26 22:36:24 -0400, James Coleman wrote:
> I had a build on Cirrus CI fail tonight in what I have to assume was
> either a problem with caching across builds or some such similar
> flakiness. In the Debian task [1] I received this error:
> 
> su postgres -c "make -s -j${BUILD_JOBS} world-bin"
> In file included from parser.c:25:
> ./gramparse.h:29:10: fatal error: 'gram.h' file not found
> #include "gram.h"
> ^~~~~~~~
> 1 error generated.
> make[3]: *** [../../../src/Makefile.global:1078: parser.bc] Error 1
> make[3]: *** Waiting for unfinished jobs....
> make[2]: *** [common.mk:36: parser-recursive] Error 2
> make[1]: *** [Makefile:42: all-backend-recurse] Error 2
> make: *** [GNUmakefile:21: world-bin-src-recurse] Error 2
> 
> There were no changes in the commits I'd made to either parser.c or
> gramparse.h or gram.h. After running "git commit --amend --no-edit"
> (with zero changes) to rewrite the commit and forcing pushing the
> build [2] seems to be fine. I've double-checked there are no
> differences between the commits on the two builds (git diff shows no
> output).
> 
> Is it possible we're missing some kind of necessary build isolation in
> the Cirrus CI scripting?

Very unlikely - most of the tasks, including debian, use VMs that are thrown
away after a single use.

The explanation is likely that you're missing

commit 16492df70bb25bc99ca3c340a75ba84ca64171b8
Author: John Naylor <john.naylor@postgresql.org>
Date:   2022-09-15 10:24:55 +0700
 
    Blind attempt to fix LLVM dependency in the backend

and that the reason you noticed this in one build but not another is purely
due to scheduling variances.

Greetings,

Andres Freund