On Sat, Oct 20, 2018 at 12:25:19AM -0400, Stephen Frost wrote:
> On Fri, Oct 19, 2018 at 23:40 Michael Paquier <michael@paquier.xyz> wrote:
>> - I agree with not doing a simple revert to not turn the buildfarm red
>> again. This is annoying for animal maintainers. Andrew has done a very
>> nice work in disabling manually those tests temporarily.
>
> This is a red herring, and always was, so I’m rather unimpressed at how it
> keeps coming up- no, I’m not advocating that we should just make the build
> farm red and just leave it that way. Yes, we should fix this case, and fix
> pg_basebackup, and maybe even try to add some regression tests which test
> this exact same case in pg_basebackup, but making the build farm green is
> *not* the only thing we should care about.
Well, the root of the problem was that pg_verify_checksums has been
committed without any tests on its own. If we had those tests from the
start, then we would not be having this discussion post-release time,
still trying to figure out if whitelisting or blacklisting is
appropriate.
The validation checksums in base backups has been added in 4eb77d50, a
couple of days before pg_verify_checksums has been introduced. This has
added corruption-related tests in src/bin/pg_basebackup, which is a good
thing. However the feature has been designed so as checksum mismatches
are ignored after 5 failures, which actually *masked* the fact that
EXEC_BACKEND files like CONFIG_EXEC_PARAMS should have been skipped
instead of getting checksum failures. And that's a bad thing. So this
gives in my opinion a good point about using a whitelist.
--
Michael