Don't bail out on test errors by default

[ihnorton]

I don't always run tests
But when I do, I ignore the failures.

Seriously though: when I run all tests locally, it is because I want to know the status of all of them for individual investigation later (e.g. #10394). Tests are slow in general, really slow on Windows, and painfully slow with MCJIT on Windows. Stopping the run due to a (possibly random) failure in the middle of the list is unhelpful because it only tells me about one test failure at a time. (I usually go away when running tests because my 2-core 3320 gets really sluggish).

This PR makes the test suite continue through errors by default, and adds a check for the environment variable JL_TESTFAILURE_STOP for use on the CI systems.

[vtjnash]

fwiw, i think it's good to run to completion on CI also. just so long as any failure poisons the final return value from runtests to cause exit(1)

[ihnorton]

Sounds fine to me. That is easy to do for the make all case, but slightly harder for the make some selected subset case because there we call runtests individually for each specified target. GMake has a MAKECMDGOALS magic variable that might be useful but I'm not sure if it is portable.

cc @tkelman @staticfloat

[staticfloat]

I don't see why running make testfoo should act any differently from make testall; we're just changing the set of tests that get returned from choosetests(); whether there are 100 entries in the tests list, or just 1, the reduce() call will still run across all of them, and your all(x -> x == false, results) test should still work.

This looks like a very positive change to me, and I'm all in favor of it.

[ihnorton]

The case I am talking about is make testfoo1 testfoo2 testfoo3. There,
the Makefile invokes runtests.jl testfoo1 runtests.jl testfoo2 etc.
separately (AFAICT) for each target. If any runtests invocation calls
exit(1) due to a failure, make will give up on the remainder.

Maybe that is ok? I don't care about the limited multi-target case very
much because I'm never going to run more than a handful of tests that way.
But I was hoping there would be a simple way to handle it nicely.

On Wed, Apr 15, 2015 at 3:01 PM, Elliot Saba notifications@github.com
wrote:

I don't see why running make testfoo should act any differently from make
testall; we're just changing the set of tests that get returned from
choosetests(); whether there are 100 entries in the tests list, or just
1, the reduce() call will still run across all of them, and your all(x ->
x == false, results) test should still work.

—
Reply to this email directly or view it on GitHub
#10835 (comment).

[vtjnash]

i'm not sure there's much you can do about that case. this pull request (or my suggestion) shouldn't change the current behavior of that case anyways, since we should always be calling exit(1) if any test failed

[amitmurthy]

Just so that we are aware - for a single worker case, all tests will not be run in case of an error, even with err_stop=false, since pmap removes a worker from the scheduling pool in case of any error on that worker.

[ViralBShah]

I really would like a status at the end that says 5 out of 5000 tests failed or something like that.

[amitmurthy]

I think you are confusing @test lines with the .jl test files itself. The granularity here is at the test file level.

[ViralBShah]

But I wish it were at the test level.

[tkelman]

@ViralBShah - ref #9398

[ihnorton]

Fixed by #13062