patcht/Makefile: retain cache t/.prove across prove runs

From: Michael Haggerty <mhagger@alum.mit.edu>

prove(1) can write a summary of its test results and timings into a cache file, t/.prove, then use this information during later runs for various purposes. But deleting t/.prove after every test run defeats this purpose. So do not delete t/.prove as part of "make DEFAILT_TEST_TARGET=prove test". (Continue to delete the file on "make clean".)

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
---
prove(1) can be told to retain information about test runs, and to use
it for interesting purposes during future invocations.  For example,
it can be told to run tests that failed during the last run:

    GIT_PROVE_OPTS = --state=failed,save

or that have failed recently:

    GIT_PROVE_OPTS = --state=hot,save

or that have been edited since the last run of prove:

    GIT_PROVE_OPTS = --state=fresh,save

It can also run tests in parallel, and start them in order from slowest to fastest, which, combined with running tests in parallel, can improve CPU utilization:

    GIT_PROVE_OPTS = --timer --jobs 5 --state=slow,save

(On my 4-core notebook, the latter speeds up the tests by about 10% compared to running them in parallel in numerical order.)

 t/Makefile |    6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

On Wed, May 02, 2012 at 05:31:52PM +0200, mhagger@alum.mit.edu wrote:

Thanks. I had found some odd behaviors with putting "--state" into GIT_PROVE_OPTS when I first switched my setup to use prove, but I didn't investigate. I think this issue explained some of what I saw.

Do note, though, that we only clean ".prove" on a successful run. So something like "--state=hot,save" will still see the cache from the last run if there were failures. However, the prove cache is meant to be kept across many runs, even successful ones, so I think your patch makes sense.

I don't think this actually works, because we also feed all of the test scripts to prove on the command line. So it will run them all, both failed and successful.

I'm not sure if it is worth solving for "make test". I suspect having that not test everything would be error prone (i.e., it's too easy to accidentally not run the full suite when you meant to, and unexpected regressions in other tests will go unnoticed). You can always run "prove" yourself from the command line, or we can add a new target to do a partial test (with other options set up properly).

-Peff

On 05/02/2012 06:07 PM, Jeff King wrote:

I totally believe you that some of the variations that I listed in my commentary don't work in the git context. I'm not a prove expert; I just noticed that removing the .prove file is counterproductive and breaks some other prove features. I also agree with you that it would be dangerous to encourage partial testing and that it is therefore not a priority to make the use case that you mentioned work in the git context.

I still think my patch makes sense. The error that Peff pointed out was in my commentary, not in the patch itself or in the log message. Junio, is there something else keeping you from applying this patch?

Michael

On Sat, May 19, 2012 at 04:54:33PM +0200, Michael Haggerty wrote:

Yeah, I hope my comments weren't interpreted as "don't apply this". Keeping the .prove file around is a prerequisite for lots of clever things, including some useful (--state=slow) and some less so (--state=failed). But if the latter case does not work (if we even want it to), it is because this patch is only the first building block. We should definitely apply it.

-Peff

Michael Haggerty <mhagger@alum.mit.edu> writes:

No.

I was waiting just in case if anything other than "Ah, that does indeed not work but it does not mean this patch is wrong" comes out (e.g. "With the patch it does not work as pointed out, but I made it also work by this reroll, which also is correct") to avoid the trouble of applying and replacing.

Thanks.