[Bioc-devel] using covr with the parallel package

[Leonard Goldstein]

Hi Jim,

You are right -- the problem is not related to mc.preschedule but
simply arises when there are child processes. Thanks for pointing this
out.

I will make some changes to my package to make sure all *apply calls
are run in the current process in situations where there is no
parallelization.

Thanks for the covr package! It is extremely useful, as many others
have pointed out.

Leonard


On Mon, Aug 17, 2015 at 1:29 PM, Jim Hester <james.f.hester at gmail.com> wrote:
> Leonard,
>
> Setting `mc.preschedule = TRUE` and `ncores < 2` results in `mclapply`
> (which `mcmlapply` calls) delegating to a simple `lapply` (see
> https://github.com/wch/r-source/blob/7578138223130d3c6197b95c3cbbe1e78a5cf230/src/library/parallel/R/unix/mclapply.R#L117).
> Since this `lapply` runs in the current process coverage works fine.
>
> `mc.preschedule = FALSE` calls `mcparallel` in all cases, which spawns child
> processes, and coverage is not tracked in child processes.
>
> I don't have any plans to support tracking coverage in child processes at
> the moment (although I suppose it is possible in the future).
>
> So the answer seems to be if you want test coverage to work with parallel
> code you need to run it with `ncores = 1` and `mc.preschedule = TRUE`.
>
> I will add something to that effect in the covr README.
>
> Thank you for reporting the issue!
>
> Jim
>
> (posting includes off-list email exchange when I neglected to include
> bioc-devel in the response)
>
> On Mon, Aug 17, 2015 at 2:34 PM, Leonard Goldstein
> <goldstein.leonard at gene.com> wrote:
>>
>> Hi Jim,
>>
>> Thanks for the quick reply!
>>
>> For example, predictTxFeaturesPerSample in
>>
>>
>> https://codecov.io/github/Bioconductor-mirror/SGSeq/R/features-prediction.R
>>
>> is called by predictTxFeatures in main.R (inside mcmapply with
>> mc.preschedule = FALSE)
>>
>> but it is considered untested.
>>
>> Thanks for looking into this.
>>
>> Leonard
>>
>>
>> On Mon, Aug 17, 2015 at 11:24 AM, Jim Hester <jhester at fredhutch.org>
>> wrote:
>> > Leonard,
>> >
>> > Can you link to an example where this behavior is occurring?  I can see
>> > evidence where the coverage is working as expected when using mclapply,
>> > e.g.
>> >
>> > (https://codecov.io/github/Bioconductor-mirror/SGSeq/R/graphs.R?ref=ff09041bd0fec09a6c138e3103bfee23f876f675#l-223),
>> > but not the reverse.
>> >
>> > Jim
>> >
>> > On Mon, Aug 17, 2015 at 2:13 PM, Leonard Goldstein
>> > <goldstein.leonard at gene.com> wrote:
>> >>
>> >> Hi Jim,
>> >>
>> >> I noticed that when covr calculates test coverage, functions called
>> >> inside mclapply or mcmapply with argument mc.preschedule = FALSE are
>> >> considered untested (even if unit tests exist)
>> >>
>> >> When running checks I only use one core. So an easy fix would be to
>> >> set mc.preschedule to TRUE if mc.cores = 1. But I was wondering if you
>> >> are aware of this behavior and whether there is a way to avoid it?
>> >>
>> >> Many thanks for your help.
>> >>
>> >> Leonard
>> >>
>> >> _______________________________________________
>> >> Bioc-devel at r-project.org mailing list
>> >> https://stat.ethz.ch/mailman/listinfo/bioc-devel
>> >
>> >