All my slaves' /tmp's are polluted with regrtest fluff.  I haven't checked yet, but I presume no cleanup is done if a test/run fails.

nitrogen% find /tmp -user cpython 2> /dev/null | wc -l
197
netbsd51-x64-1$ find /tmp -user cpython 2> /dev/null | wc -l
     142
oxygen% find /tmp -user cpython 2> /dev/null | wc -l                
456

Placeholder issue until it annoys me enough to patch.

Cleanup on test failure is supposed to be done.  Cleanup on crash or buildbot timeout isn't done as far as I know (and that was a concern I had with the changes made to support.TESTFN and the cwd, but I didn't articulate it very well).  

If you find tests that do no clean up on error, that's definitely a bug and we should fix them.

To follow up on David's comment, the unit tests in the test suite aren't consistent in their treatment of temp directories (e.g. they don't use a common API or code path).  So it may be hard to address this globally short of wiping the entire temp directory (though I could be wrong).

I have a patch in issue 15415 to add a temp_dir() context manager to test.support (and consolidate with script_helper's) that would do such clean-up.  The module already has a temp_cwd(), but that context manager serves two purposes.

> All my slaves' /tmp's are polluted with regrtest fluff.

Which "regrtest fluff" exactly?

> > All my slaves' /tmp's are polluted with regrtest fluff.

> Which "regrtest fluff" exactly?

I was being lazy when I wrote that.  By "regrtest fluff" I was referring to lingering files/directories in /tmp, owned by cpython, starting with tmp*.  Few examples below.

openbsd51-x64-1% ls -l /tmp
total 32
drwxr-xr-x  2 root     wheel  512 Aug 18 05:40 aucat
drwx------  4 cpython  wheel  512 Sep 19 02:26 tmp03dyjv
-rw-------  1 cpython  wheel    3 Sep 19 02:03 tmp4t3_bk
-rw-------  1 cpython  wheel    3 Sep 18 05:02 tmp5ljha3
-rw-------  1 cpython  wheel    3 Sep 18 05:02 tmp7p3sht
-rw-------  1 cpython  wheel    0 Sep 19 05:34 tmpCFMYVJ
-rw-------  1 cpython  wheel    3 Sep 19 02:34 tmpcnwqcn
-rw-------  1 cpython  wheel    3 Sep 19 02:34 tmpk0gthu
-rw-------  1 cpython  wheel    3 Sep 19 02:03 tmpmf973r
-rw-------  1 cpython  wheel    0 Sep 19 05:34 tmpqO20FC

dragonfly30-x86-1% ls -l /tmp | grep cpython
drwx------  1 cpython  wheel     0 Aug 22 03:46 ssh-v3WDeO7FhO
-rw-------  1 cpython  wheel     3 Sep 14 16:05 tmp31idrc
-rw-------  1 cpython  wheel     3 Sep 11 15:58 tmp34z8iv
-rw-------  1 cpython  wheel     3 Sep 18 05:06 tmp43uhjt
-rw-------  1 cpython  wheel     3 Sep 19 02:03 tmp4jypkx
-rw-------  1 cpython  wheel     3 Sep 12 04:31 tmp4odqso
-rw-------  1 cpython  wheel     3 Sep 12 20:36 tmp6acej6
-rw-------  1 cpython  wheel     3 Sep 12 16:22 tmp9m5vie
-rw-------  1 cpython  wheel     3 Sep 11 16:15 tmp9oj443
-rw-------  1 cpython  wheel     3 Sep 12 20:10 tmpb086ow
-rw-------  1 cpython  wheel     3 Sep 12 14:02 tmpcm0s2g
-rw-------  1 cpython  wheel     3 Sep 11 16:15 tmpd81jie
-rw-------  1 cpython  wheel     3 Sep 12 17:05 tmpdm8vv9
-rw-------  1 cpython  wheel     3 Sep 15 23:01 tmpef4d2m
-rw-------  1 cpython  wheel     3 Sep 18 05:06 tmpel_xi_
-rw-------  1 cpython  wheel     3 Sep 12 04:31 tmpf41fj3
-rw-------  1 cpython  wheel     3 Sep 11 12:08 tmpi94ajz
-rw-------  1 cpython  wheel     3 Sep 19 02:25 tmpilpy7y
-rw-------  1 cpython  wheel     3 Sep 15 22:27 tmpiq1yo9
-rw-------  1 cpython  wheel     3 Sep 13 16:36 tmpkq6d8i
-rw-------  1 cpython  wheel     3 Sep 11 12:28 tmpkxqa4a
-rw-------  1 cpython  wheel     3 Sep 19 02:25 tmpl35290
-rw-------  1 cpython  wheel     3 Sep 11 12:08 tmpn1dq2c
-rw-------  1 cpython  wheel     3 Sep 12 17:05 tmpnxrnxp
-rw-------  1 cpython  wheel     3 Sep 19 02:03 tmpocz5ux
-rw-------  1 cpython  wheel     3 Sep 14 16:05 tmppsebwf
-rw-------  1 cpython  wheel     3 Sep 11 13:50 tmpptpcep
-rw-------  1 cpython  wheel     3 Sep 15 22:27 tmpqk1s3k
-rw-------  1 cpython  wheel     3 Sep 12 14:02 tmprk8uxy
-rw-------  1 cpython  wheel     3 Sep 11 13:50 tmprqmmfw
-rw-------  1 cpython  wheel     3 Sep 13 16:36 tmps99aky
-rw-------  1 cpython  wheel     3 Sep 11 15:58 tmpsfsmoc
-rw-------  1 cpython  wheel     3 Sep 12 20:36 tmpt8_0vu
-rw-------  1 cpython  wheel     3 Sep 15 23:01 tmpthosfe
-rw-------  1 cpython  wheel     3 Sep 11 12:28 tmpwhsp9p
-rw-------  1 cpython  wheel     3 Sep 12 20:10 tmpy5774k
-rw-------  1 cpython  wheel     3 Sep 12 16:22 tmpzgafbz


oxygen% ls -l /tmp | grep cpython
drwxr-xr-x 3 root    cpython   96 Aug 31 15:41 screens
-rw------- 1 cpython cpython    0 Sep 19 05:30 tmp249ZUA
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpAvStO1
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpFLfr2x
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpOYB_X6
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpP1Hsdf
-rw------- 1 cpython cpython    0 Sep 19 05:30 tmpTqctqQ
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpUW1Gov
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpYU0lzE
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpazia2w
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpgemyJY
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpkfg0aZ
drwx------ 2 cpython cpython   96 Sep 19 05:31 tmpt1e0Jc

> Cleanup on test failure is supposed to be done.  Cleanup on crash or 
> buildbot timeout isn't done as far as I know (and that was a concern I
> had with the changes made to support.TESTFN and the cwd, but I didn't
> articulate it very well).  

Ah, yeah, this is almost certainly to do with crash/timeout.  I remember running into this way back in 2008 when I was running a few slaves.

Personally I think the best solution is to have the test framework allocate a single test directory, inform the parent* as to what it is, then make sure all temp files are rooted in it.  The parent* should then rm -rf the tempdir at the end of the run as a final 'catch all'.

parent*: this will mean different things in different contexts... could be the Tools/run_tests.py wrapper, buildbot slave wrapper, etc.

> Personally I think the best solution is to have the test framework allocate a single test directory

This is partially done.  See here:

http://hg.python.org/cpython/file/19c74cadea95/Lib/test/regrtest.py#l1810

# Run the tests in a context manager that temporary changes the CWD to a
# temporary and writable directory.

regrtest sets things up such that the current working directory is supposed to be this temp directory (constructed from _make_temp_dir_for_build()).  But the tests adhere to this only weakly.  Many or most tests create their own temp directory rather than relying on the caller having set the current working directory to a temp directory.  If the tests used a common API, we could control this behavior globally.