CI: Run latest JRuby release

[olleolleolle]

This PR updates the CI matrix to use latest JRuby, 9.2.11.1.

JRuby 9.2.11.1 release blog post

Update: I changed the configuration not to point to a release, but to an rvm alias.

rvm responded with this:

$ rvm use jruby --install --binary --fuzzy
curl: (22) The requested URL returned error: 404 Not Found
Required jruby-head is not installed - installing.

[mperham]

I'd like to know if there is a way to ask for the latest stable, so we don't have to bump this constantly.

[olleolleolle]

If rvm has an alias in its "known" file, we can use that. I'll update this.

[olleolleolle]

https://github.com/rvm/rvm/blob/master/config/known

[olleolleolle]

@mperham There.

[mperham]

The jruby build is broken and I have no idea how to fix it. Are we just going to ignore it? If so, what's the point of including it?

[headius]

For reference the JRuby failure is here: https://travis-ci.org/github/mperham/connection_pool/jobs/685746458

I also noticed Ruby 2.6 failed one spec, but it seems to be a different spec.

There's only two failures on JRuby and they both appear to be related to concurrent-ruby executor logic. Perhaps the bug is there? I don't see anything that seems likely to be a JRuby bug.

[olleolleolle]

Tests do not fail on JRuby 9.2.11.1 locally

$ uname -a
Darwin Olles-MBP-2.lan 19.4.0 Darwin Kernel Version 19.4.0: Wed Mar  4 22:28:40 PST 2020; root:xnu-6153.101.6~15/RELEASE_X86_64 x86_64
$ ruby -v
jruby 9.2.11.1 (2.5.7) 2020-03-25 b1f55b1a40 OpenJDK 64-Bit Server VM 14.0.1+7 on 14.0.1+7 +jit [darwin-x86_64]

I seem to run a newer JDK.

Update Well, what unites the two tests that fail is that they use Timeout.timeout(n) { }.

[headius]

Nevermind... I saw "concurrent" in the trace and thought "concurrent-ruby". This appears to be happening in JRuby's timeout implementation.

@olleolleolle That's helpful information, thanks. I just ran it on 8 with JRuby master and it does fail locally there.

[headius]

This was introduced by jruby/jruby@37325d6

In order to avoid leaving the Timeout executor running (e.g. for embedded JRuby use cases where the JVM continues running) we introduced code to terminate it at shutdown. Unfortunately we introduced that shutdown using the same mechanism as at_exit, which means the executor might be shut down before all Ruby code registered with at_exit has run.

This is the cause of the failure, because Minitest uses an at_exit hook to run all its tests. That hook gets registered later than Timeout's executor-shutdown hook, so there's no executor by the time the tests run.

This is a trivial fix in JRuby. The only possible workaround would be to forcibly remove the poison pill task in at_exit, which can't be done without digging around a lot of JRuby internals.

[headius]

The introduced bug only happens on JRuby master, so 9.2.11.1 is not affected.

I have filed jruby/jruby#6212 and will land a fix shortly.

[headius]

I have merged jruby/jruby#6212 into JRuby master and it should go out in the next snapshot and nightly builds.

[headius]

Oops, jruby/jruby#6213 is the PR.