Re: Article about "git bisect run" on LWN

Le jeudi 5 février 2009, Ingo Molnar a écrit :

* Christian Couder [off-list ref] wrote:

quoted

Hi,

For information, an article from me, 'Fully automated bisecting with
"git bisect run"' has been published in today's edition of LWN on the
development page:

http://lwn.net/Articles/317154/

Nice article!

In terms of possible future enhancements of git bisect, here's a couple
of random ideas that would help my auto-bisection efforts:

 - Feature: support "Bisection Redundancy"

   This feature helps developers realize if a bug is sporadic. This
happens quite often in the kernel space: a bug looks deterministic, but
down the line it becomes sporadic. Sometimes a boot crash only occurs
with a 75% probability - and if one is unlucky it can cause a _lot_ of
wasted bisection time. The wrong commit gets blamed and the wrong set of
developers start scratching their heads. It's a reoccuring theme on lkml.

   What git could do here is to allow testers to inject a bit of extra
   "redundancy" automatically, and use the redundant test-points to
detect conflicts in good/bad constraints.

   It would work like this:

      git bisect start --redundancy=33%

   It would mean that for every third bisection points, Git would
   _not_ chose the ideal (estimated) 'middle point' from the set of
"unknown quality" changes that are still outstanding - but would
intentionally "weer outside" and select one commit from the _known_ set
of commits.

   If such a redundant re-test of the known-good or known-bad set yields
a nonsensical result then Git aborts the bisection with a "logic
inconsistency detected" kind of message - and people could at this point
realize the non-determinism of the test.

   ( Git can do this when a "redundant" test point is marked as 'bad' -
     despite an earlier bisection already categorizing that test point as
     'good' - or if it's the other way around. Git will only continue
with the bisection if the test point has the expected quality. )

   This essentially means an automated re-test - but it's much better
than just a repeated bisection - i've often met non-deterministic bugs
that yield the _exact same_ nonsensical commit even on repeat bisections.
That happens when a timing bug depends on the exact kernel layout, or a
miscompilation or linker bug depends on the exact kernel layout, etc.

   It's also faster than a re-done bisection: 33% more testpoints is
better than twice as many test-points. Also, auto-bisection can deal with
redundancy just fine - it does not really matter whether i have to wait
20 or 30 minutes for a test result since there's no manual intervention
needed - but it _very_ much matters whether i can trust the validity of
the bisection result.

I see. With the current code it seems difficult to have a "weer outside" 
algorithm, but if the "git bisect skip" code is ported (from shell 
in "git-bisect.sh") to C in "builtin-rev-list.c", that might be doable. And 
anyway moving the bisect skip code to C is needed for other improvements.

- Feature: better "git bisect next" support.

You probably mean "git bisect skip" here.

  Sometimes a commit wont build. In that case we have "git bisect next",
but last i checked that only jumps a single commit - and build breakages 
often have a large scope - full trees that got merged upstream, etc. Most
of the time those build breakages are uninteresting and the build-broken
window does not contain the bad commit.

  So it would be nice to have a "git bisect next --left=20%" type of
  feature. This would jump 20% commits to the "left" from the bisection
  point, towards the 'known bad' set of commits, but still within the
  bisection window.

  Similarly, "git bisect next --right=20%" would jump towards the
known-good edge of the bisection window (but still within the bisection
window).

In the following thread, H. Peter Anvin suggested an algorithm to deal with 
this kind of problem:

http://thread.gmane.org/gmane.comp.version-control.git/98164/

And I suggested a simpler one, that might be implemented without having to 
port "git bisect skip" code to C first, but I did not work on it yet.

  Currently when i hit a build error during auto-bisection, it aborts and
i have to intervene manually. But with a bigger jump distance i could use
git-bisect-next reliably in scripts too.

  Likewise, users too hit build breakages often, and find it hard to get
out of the window of breakage. With the high-order tree structure of the
kernel repository that is rather non-intuitive to do as well, and often
people make mistakes and test the wrong commit.

I am working slowly on "git replace" these days and, if everything goes 
well, it should make it possible to use "replace" refs when bisecting, so 
that people could bisect on commit trees where many breakages have been 
removed. And as refs can be shared, this means that users and developers 
should be able to easily share these improved trees.

Another way to work around breakages could be to have a list of commits and 
ranges of commits that should always be skipped and always pass them 
to "git bisect skip" before using "git bisect run". Something like that 
perhaps:

$ git bisect start <bad> <good>
$ git bisect skip $(cat always_skipped.txt)
$ git bisect run ./my_test_script.sh

But I agree that it would also be a good thing to have an improved "git 
bisect skip" that could jump out of breakage windows.

- Feature: detect "redundant" and "inconsistent" test points

  This is a variation of the redunant testing theme, but from a different
  angle: often newbies when they bisect the kernel weer outside of the
  bisection window without realizing it. It would be nice if Git printed
a friendly notifier that:

     git bisect good 12341234
     info: bisection point 12341234 was already in the 'good' range

I agree that it would be nice. And it might be easy to implement.

  Or, if the redunant test point is conflicting, print:

     git bisect good 12341234
     fatal: bisection point 12341234 was already in the 'bad' range!

  And give an error return as well, so that scripts can abort.

  Currently Git seems to be very forgiving and accepts all bisection
points that we feed it, without checking them for consistency. (this
might have changed in current development versions, i dont know.)

I don't think conflicting test points are accepted. An error should be 
reported (with exit code 1), like this:

$ git bisect start HEAD~5 HEAD
Some good revs are not ancestor of the bad rev.
git bisect cannot work properly in this case.
Maybe you mistake good and bad revs?

- User friendliness: give an estimation about how many steps are
remaining

  Right now git prints this when a bisection session begins:

     aldebaran:~/tip> git bisect start
     aldebaran:~/tip> git bisect bad linus
     aldebaran:~/tip> git bisect good v2.6.28
     Bisecting: 5449 revisions left to test after this
     [e0b685d39a0404e7f87fb7b7808c3b37a115fe11] Updated contact info for
CREDITS file

  It would be nice if Git estimated the expected number of bisection
points. Something like this would be helpful:

     aldebaran:~/tip> git bisect good v2.6.28
     Bisecting: 5449 revisions left to test after this
                About ~16 test steps left [approximated]
     [e0b685d39a0404e7f87fb7b7808c3b37a115fe11] Updated contact info for
CREDITS file

  The real number of test points might be higher than this, if the tree
  layout is unlucky, or it might be less than this if the user manually
  narrows the bisection window to a suspected set of commits - but that's
  OK - most kernel testers use the default variant and the message is
clear enough that it's only an estimation.

I will have a look at that. It might be easy to do.

Thanks,
Christian.