Intermittent failure in math test

[martinholters]

I just had a make test (against 0.6.0-dev.711) spitting out

    From worker 7:       * math                 Test Failed
    From worker 7:    Expression: lgamma(-300im) ≅ -473.1718507425924 - 1410.3490664555823im
    From worker 7:     Evaluated: -473.1718507425924 - 1410.349066455582im isequal -473.1718507425924 - 1410.3490664555823im
error in running finalizer: Base.MethodError(f=Base.#real(), args=(Base.RemoteChannel{Base.Channel{Int64}}(where=1, whence=1, id=2),))
error in running finalizer: Base.MethodError(f=Base.#real(), args=(Base.RemoteChannel{Base.Channel{Int64}}(where=1, whence=1, id=2),))
jl_uv_writecb() ERROR: bad file descriptor EBADF
jl_uv_writecb() ERROR: bad file descriptor EBADF
jl_uv_writecb() ERROR: bad file descriptor EBADF
jl_uv_writecb() ERROR: bad file descriptor EBADF
jl_uv_writecb() ERROR: bad file descriptor EBADF

some more pages of that line repeating, ending in

Exception running test math :
On worker 7:
LoadError: There was an error during testing
 in record at ./test.jl:397
 in do_test at ./test.jl:281
 in include_string at ./loading.jl:478
 in include_from_node1 at ./loading.jl:544
 in macro expansion at ./util.jl:232 [inlined]
 in runtests at /home/hol/Projects/julia/test/testdefs.jl:7
 in #619 at ./multi.jl:1425
 in run_work_thunk at ./multi.jl:1005
 in macro expansion at ./multi.jl:1425 [inlined]
 in #618 at ./event.jl:68
while loading /home/hol/Projects/julia/test/math.jl, in expression starting on line 658
ERROR: LoadError: Some tests exited with errors.
 in (::##11#18)() at /home/hol/Projects/julia/test/runtests.jl:72
 in cd(::##11#18, ::String) at ./file.jl:69
 in include_from_node1(::String) at ./loading.jl:541
 in process_options(::Base.JLOptions) at ./client.jl:262
 in _start() at ./client.jl:326
while loading /home/hol/Projects/julia/test/runtests.jl, in expression starting on line 26
Makefile:12: recipe for target 'all' failed
make[1]: *** [all] Error 1
Makefile:569: recipe for target 'test' failed
make: *** [test] Error 2

(All other tests completed successfully, output omitted.)

Re-running the tests, everything looks fine. Can anyone make sense of that?

[tkelman]

I've seen this too. It depends which tests you run in the same worker, and it started when #12563 was merged (cc @stevengj). It's reproducible 100% of the time if you do make testall1 and I'm currently running a little reducer loop to narrow that down to a specific subset.

edit: yeah, numbers and math in the same worker is enough

[martinholters]

If that is any help, I had

    From worker 7:       * linalg/special        in  44.59 seconds, maxrss  240.78 MB
    From worker 7:       * linalg/lapack         in  57.78 seconds, maxrss  295.49 MB
    From worker 7:       * linalg/symmetric      in  26.01 seconds, maxrss  327.53 MB
    From worker 7:       * linalg/uniformscaling in  38.01 seconds, maxrss  363.84 MB
    From worker 7:       * inference             in   1.06 seconds, maxrss  364.09 MB
    From worker 7:       * keywordargs           in   1.88 seconds, maxrss  364.71 MB
    From worker 7:       * numbers               in  99.19 seconds, maxrss  475.17 MB
    From worker 7:       * copy                  in   1.98 seconds, maxrss  479.64 MB

on the same worker before the failure happened.

[pabloferz]

Shouldn't lgamma(-300im) ≅ -473.1718507425924 - 1410.3490664555823im be lgamma(-300im) ≈ -473.1718507425924 - 1410.3490664555823im?

Note the symbol for isapprox vs. the used, defined as isequal for the tests.

[pabloferz]

After #12563, ≅ is redefined in test/numbers.jl, whereas for test/math.jl it is ≅(a,b) = relerrc(a,b) ≤ 1e-13. So I'd guess when numbers is tested before math on the same worker it does not changes ≅ definition.

[pabloferz]

Putting a big let block around those test files or making them modules should be two ways of fixing this.

[tkelman]

@pabloferz yep that looks about right. I grepped for other uses of ≅ but must've missed that it was defined in both those test files. We could switch one of them to a different symbol, or not use const, or use a let block for one or both.

[tkelman]

also x-ref #17435

[stevengj]

Yeah, I probably should have used ≣ in numbers.jl anyway.

(The reason I don't use ≈ in math.jl is that the relerrc test is stronger: not only is the tolerance lower, but it checks that the relative error in both the real and imaginary parts is small independently.)