919402 - automatic machine loaning on an intermittently crashing test

Status: RESOLVED DUPLICATE of bug 965691

(Release Engineering :: General, defect)

Description

[Cameron McCormack (:heycam)]

For intermittent test failures that cause a crash, it would be convenient to be able to tell the build infrastructure to attach a debugger to the process that just crashed and then make the machine available to be loaned to the person who requested this.  For example say I want to investigate bug 916704, which we're getting random crashes in currently but which I haven't managed to reproduce locally yet.  It would be great if I could say either update the reftest manifest entry from

  == svg-glyph-paint-server.svg svg-glyph-paint-server-ref.svg

to

  debug-on-crash == svg-glyph-paint-server.svg svg-glyph-paint-server-ref.svg

or maybe have a script function to call from within the test file itself.  This would then let the build infrastructure know that if the firefox process crashes while running this test, it will end the reftest run (as it does), but rather than returning the slave to the pool, it keeps the process alive, maybe opens a terminal and attaches a debugger to it or displays some instructions on how someone can do that, and then notifies someone (email?) that the crash has been caught.  Or maybe it lets the process die but writes out a core dump that can be opened with gdb.

Maybe this could be extended to non-crashing intermittent failures too.  In that case stopping the process (but not killing it) so that a debugger can be attached to it, rather than writing out a core dump, would definitely be preferable.