[RFC PATCH] tests: print failed test numbers at the end of the test run

[RFC PATCH] tests: print failed test numbers at the end of the test run

[Jens Lehmann]

On modern multi-core processors "make test" is often run in multiple jobs.
If one of them fails the test run does stop, but the concurrently running
tests finish their run. Finding out what test is broken involves a lot of
scrolling. That gets even worse when the -i option is used.

If one or more tests failed, print a list of them before the test summary:

failed test(s): t1000 t6500

fixed   0
success 7638
failed  3
broken  49
total   7723

This makes it possible to just run the test suite with -i and collect all
failed test scripts at the end for further examination.

Signed-off-by: Jens Lehmann <Jens.Lehmann@web.de>
---

Maybe I'm missing something completely obvious, but I always have a hard
time finding out which test scripts did fail in a test run with -j30.

 t/aggregate-results.sh |   12 +++++++++++-
 1 files changed, 11 insertions(+), 1 deletions(-)

diff --git a/t/aggregate-results.sh b/t/aggregate-results.sh
index d206b7c..b8e929a 100755
--- a/t/aggregate-results.sh
+++ b/t/aggregate-results.sh
@@ -1,5 +1,6 @@
 #!/bin/sh

+failed_tests=
 fixed=0
 success=0
 failed=0
@@ -18,7 +19,12 @@ do
 		success)
 			success=$(($success + $value)) ;;
 		failed)
-			failed=$(($failed + $value)) ;;
+			failed=$(($failed + $value))
+			if test $value != 0; then
+				testnum=$(echo $file | cut -b 14-18)
+				failed_tests="$failed_tests $testnum"
+			fi
+			;;
 		broken)
 			broken=$(($broken + $value)) ;;
 		total)
@@ -27,6 +33,10 @@ do
 	done <"$file"
 done

+if [ -n "$failed_tests" ]; then
+	printf "\nfailed test(s):$failed_tests\n\n"
+fi
+
 printf "%-8s%d\n" fixed $fixed
 printf "%-8s%d\n" success $success
 printf "%-8s%d\n" failed $failed
-- 
1.7.6.346.g750efc

Re: [RFC PATCH] tests: print failed test numbers at the end of the test run

[Junio C Hamano]

Jens Lehmann <Jens.Lehmann@web.de> writes:

> Maybe I'm missing something completely obvious, but I always have a hard
> time finding out which test scripts did fail in a test run with -j30.

I run "ls -d t/trash*" for that ;-)

> +			if test $value != 0; then
> +				testnum=$(echo $file | cut -b 14-18)
> +				failed_tests="$failed_tests $testnum"
> +			fi

Somehow "cut" here feels dirty, as it hardcodes a mysterious offset that
cannot be guessed without knowing what string it is cutting from by
looking at t/Makefile.

	testnum=$(expr "$file" : 'test-results/\(t[0-9]*\)-')

[PATCH v2] tests: print failed test numbers at the end of the test run

[Jens Lehmann]

On modern multi-core processors "make test" is often run in multiple jobs.
If one of them fails the test run does stop, but the concurrently running
tests finish their run. It is rather easy to find out which test failed by
doing a "ls -d t/trash*". But that only works when you don't use the "-i"
option to "make test" because you want to get an overview of all failing
tests. In that case all thrash directories are deleted end and the
information which tests failed is lost.

If one or more tests failed, print a list of them before the test summary:

failed test(s): t1000 t6500

fixed   0
success 7638
failed  3
broken  49
total   7723

This makes it possible to just run the test suite with -i and collect all
failed test scripts at the end for further examination.

Signed-off-by: Jens Lehmann <Jens.Lehmann@web.de>
---

Am 24.07.2011 09:41, schrieb Junio C Hamano:
> Jens Lehmann <Jens.Lehmann@web.de> writes:
> 
>> Maybe I'm missing something completely obvious, but I always have a hard
>> time finding out which test scripts did fail in a test run with -j30.
> 
> I run "ls -d t/trash*" for that ;-)

Yup. But that won't work when you use "-i" to continue to run all tests
even when some fail, as all trash directories get deleted in the end. Or
did I manage to overlook an option which stops that? ;-)

I rephrased the commit message to make it clear my patch only changes the
behavior when "-i" is used and one or more tests fail, as the code in
question isn't executed otherwise.

>> +			if test $value != 0; then
>> +				testnum=$(echo $file | cut -b 14-18)
>> +				failed_tests="$failed_tests $testnum"
>> +			fi
> 
> Somehow "cut" here feels dirty, as it hardcodes a mysterious offset that
> cannot be guessed without knowing what string it is cutting from by
> looking at t/Makefile.
> 
> 	testnum=$(expr "$file" : 'test-results/\(t[0-9]*\)-')

Thanks, this version is using that expression.


 t/aggregate-results.sh |   12 +++++++++++-
 1 files changed, 11 insertions(+), 1 deletions(-)

diff --git a/t/aggregate-results.sh b/t/aggregate-results.sh
index d206b7c..60465fa 100755
--- a/t/aggregate-results.sh
+++ b/t/aggregate-results.sh
@@ -1,5 +1,6 @@
 #!/bin/sh

+failed_tests=
 fixed=0
 success=0
 failed=0
@@ -18,7 +19,12 @@ do
 		success)
 			success=$(($success + $value)) ;;
 		failed)
-			failed=$(($failed + $value)) ;;
+			failed=$(($failed + $value))
+			if test $value != 0; then
+				testnum=$(expr "$file" : 'test-results/\(t[0-9]*\)-')
+				failed_tests="$failed_tests $testnum"
+			fi
+			;;
 		broken)
 			broken=$(($broken + $value)) ;;
 		total)
@@ -27,6 +33,10 @@ do
 	done <"$file"
 done

+if [ -n "$failed_tests" ]; then
+	printf "\nfailed test(s):$failed_tests\n\n"
+fi
+
 printf "%-8s%d\n" fixed $fixed
 printf "%-8s%d\n" success $success
 printf "%-8s%d\n" failed $failed
-- 
1.7.6.346.g940def.dirty

Re: [PATCH v2] tests: print failed test numbers at the end of the test run

[Jeff King]

On Sun, Jul 24, 2011 at 03:35:54PM +0200, Jens Lehmann wrote:

> Am 24.07.2011 09:41, schrieb Junio C Hamano:
> > Jens Lehmann <Jens.Lehmann@web.de> writes:
> > 
> >> Maybe I'm missing something completely obvious, but I always have a hard
> >> time finding out which test scripts did fail in a test run with -j30.
> > 
> > I run "ls -d t/trash*" for that ;-)
> 
> Yup. But that won't work when you use "-i" to continue to run all tests
> even when some fail, as all trash directories get deleted in the end. Or
> did I manage to overlook an option which stops that? ;-)

I used to use:

  make -k -j30
  grep 'failed [^0]' test-results/*

but these days I use the "prove" tool, which has much nicer output. You
can get Makefile support for it with:

  GIT_PROVE_OPTS = -j30
  DEFAULT_TEST_TARGET = prove

in your config.mak.

-Peff

Re: [PATCH v2] tests: print failed test numbers at the end of the test run

[Jens Lehmann]

Am 24.07.2011 17:39, schrieb Jeff King:
> On Sun, Jul 24, 2011 at 03:35:54PM +0200, Jens Lehmann wrote:
> 
>> Am 24.07.2011 09:41, schrieb Junio C Hamano:
>>> Jens Lehmann <Jens.Lehmann@web.de> writes:
>>>
>>>> Maybe I'm missing something completely obvious, but I always have a hard
>>>> time finding out which test scripts did fail in a test run with -j30.
>>>
>>> I run "ls -d t/trash*" for that ;-)
>>
>> Yup. But that won't work when you use "-i" to continue to run all tests
>> even when some fail, as all trash directories get deleted in the end. Or
>> did I manage to overlook an option which stops that? ;-)
> 
> I used to use:
> 
>   make -k -j30
>   grep 'failed [^0]' test-results/*

Ah, now I understand how I managed to shoot myself in the foot when I tried
that: I used "-i" instead of "-k" for make, so the test-results directory
always was deleted after the tests ran.

> but these days I use the "prove" tool, which has much nicer output. You
> can get Makefile support for it with:
> 
>   GIT_PROVE_OPTS = -j30
>   DEFAULT_TEST_TARGET = prove
> 
> in your config.mak.

Thanks, the output really is nicer and it even runs a bit faster! :-)

Re: [PATCH v2] tests: print failed test numbers at the end of the test run

[Sverre Rabbelier]

Heya,

On Sun, Jul 24, 2011 at 21:21, Jens Lehmann <Jens.Lehmann@web.de> wrote:
> Ah, now I understand how I managed to shoot myself in the foot when I tried
> that: I used "-i" instead of "-k" for make, so the test-results directory
> always was deleted after the tests ran.

Regardless, I like your patch. I don't use prove, and I'm always
afraid that when I run the test suite I'll find a test that failed
with no way to find out which :).

-- 
Cheers,

Sverre Rabbelier

Re: [RFC PATCH] tests: print failed test numbers at the end of the test run

[Nguyen Thai Ngoc Duy]

On Sun, Jul 24, 2011 at 2:16 AM, Jens Lehmann <Jens.Lehmann@web.de> wrote:
> On modern multi-core processors "make test" is often run in multiple jobs.
> If one of them fails the test run does stop, but the concurrently running
> tests finish their run.

Somewhat related (or not). I change something. I know it breaks things
and want to know _all_ tests it breaks, but "make test" would stop
early. Is there anyway to make it keep going through all tests even if
some fails? "make -j<big number>" improves the situation but does not
really solve it.
-- 
Duy

Re: [RFC PATCH] tests: print failed test numbers at the end of the test run

[Jeff King]

On Tue, Jul 26, 2011 at 11:25:21AM +0700, Nguyen Thai Ngoc Duy wrote:

> On Sun, Jul 24, 2011 at 2:16 AM, Jens Lehmann <Jens.Lehmann@web.de> wrote:
> > On modern multi-core processors "make test" is often run in multiple jobs.
> > If one of them fails the test run does stop, but the concurrently running
> > tests finish their run.
> 
> Somewhat related (or not). I change something. I know it breaks things
> and want to know _all_ tests it breaks, but "make test" would stop
> early. Is there anyway to make it keep going through all tests even if
> some fails? "make -j<big number>" improves the situation but does not
> really solve it.

Try "make -k", which will keep running rules that don't have failed
dependencies (in this case, all of the tests are independent, so it will
run all of them).

Or use "prove", which can do fancy things like putting the last-failed
(so you can see the likely failures immediately and start to probe them
while the rest of the suite runs).

-Peff

Re: [RFC PATCH] tests: print failed test numbers at the end of the test run

[Jeff King]

On Mon, Jul 25, 2011 at 11:42:36PM -0600, Jeff King wrote:

> Or use "prove", which can do fancy things like putting the last-failed
> (so you can see the likely failures immediately and start to probe them
> while the rest of the suite runs).

Oops, that should read "...putting the last-failed tests at the
beginning".

-Peff