Fix tests for resque v1.26.0

[nevans]

This has been tested with all currently supported versions of ruby (2.3, 2.4, 2.5, and 2.6).

Major changes (as documented in 9815ab6):

• Tests use normal fork and communicate with test runner via IO.pipe.
• Tests uses a zero interval to exit ASAP without timeout when done.
• Children need to exit! to avoid running test/unit at_exit hooks.
• Gemspec depends on a specific minor version of resque.

We must depend on ~> v1.26.0, because these methods are considered
internal, are not a stable API, and can change with every version. I.e.
as written, the tests do not pass with v1.25.0 nor with v1.27.0. Until
Resque::Worker has a stable API we can target, this gem will need to be
tested with each minor release.

[nevans]

For what it's worth, now that I have working tests I'm going to fix it for v1.27.x and v2.0.x as well. Hopefully I'll send in PRs for those later today. I'd personally recommend three separate gem releases (e.g. v0.4.5 supports "~> v1.26.0", v.0.5.0 supports "~> v1.27.0", and v0.6.0 supports "~> 2.0.0"

[stulentsev]

@nevans do all tests pass for you? I get one failure in test_timeout_limit_sequence_of_events.

[jeremy]

Wow awesome @nevans. Thank you for tackling this!

[nevans]

@stulentsev yes, they all pass for me, using bundle exec rake test using Linux (Ubuntu), and ruby 2.3.8, 2.4.5, 2.5.3, and 2.6.0. Can you share your error output? I can try to reproduce in docker if that'd be better for repeatability.

[nevans]

@stulentsev I didn't see the issues in this branch, but I can get irregular failures for older rubies in the branches for #13 and #14. I think it must be due to race conditions, since the tests rely on sleep in different threads and processes. The tests are less likely to fail for me when run with VERBOSE=true.

One way to lazily avoid the race conditions would be to simply increase the sleep times. I'll try that first.

[nevans]

It does seem like there might be a bug in earlier ruby versions, where the TSTP signal is being missed or the child isn't shutting down after TermException + TSTP. That results in a SIGKILL (after term_timeout). I'll try to figure it out ASAP.

[nevans]

I'm still doing some more testing and trying to find some documentation for why this would be the case, but it currently seems to me that this can be fixed by simply changing the child shutdown signal from TSTP (terminal stop, e.g. what usually happens when you press ^Z) to QUIT (which does the same thing in normal resque: a graceful shutdown). TSTP is sometimes simply not seen by the child (and I don't know why). But QUIT always is.