Thread

Getting rid of intermittent PPC64 buildfarm failures

Tom Lane <tgl@sss.pgh.pa.us> — 2020-10-13T16:50:48Z

I grow quite weary of the number of buildfarm failures we see as a
consequence of the Linux PPC64 bug discussed in [1].  Although we can
anticipate that the fix will roll out into new kernel builds before much
longer, that will have very little effect on the buildfarm situation,
given that a lot of Mark's PPC64 armada is running hoary "stable" kernels.
It might be many years before there are no unpatched systems to worry about.

I think it's time to give up and disable the infinite_recurse test on such
platforms.  It's teaching us nothing and we waste valuable developer time
eyeballing failures to make sure they're just the same old same old.
Testing the case on not-Linux-PPC64 is enough to verify that our own code
works.

We can use the same technique used in collate.linux.utf8.sql,
namely check the output of version() and abandon the test if it matches.
To minimize the maintenance pain from needing two expected-files, it seems
prudent to split infinite_recurse into its own test script, which leads
to the attached proposed patch.

Any objections?

			regards, tom lane

[1] https://www.postgresql.org/message-id/flat/20190723162703.GM22387%40telsasoft.com