Re: [PATCH 0/4] make it possible to skip away from broken commits

From: H. Peter Anvin <hpa@zytor.com>
Date: Jun 3, 2009, 03:14 UTC
Message-ID: <4A25EA85.5090208@zytor.com>
In-Reply-To: <7vmy8qe4ru.fsf@alter.siamese.dyndns.org>

Junio C Hamano wrote:

Show 19 quoted lines

> Christian Couder <chriscool@tuxfamily.org> writes:
> 
>> This patch series adds a "--ratio=x/y" option to "git bisect skip" so
>> that it is possible to skip away from an area were the commits cannot
>> be tested.
>>
>> Note that in this series "--ratio=4" means the same as "--ratio=1/4".
>> But I am not sure if this shortcut is worth it.
> 
> Actually my gut feeling is that a tweakable knob itself is worth it,
> because the user can never tell what the right value should be.
> 
> Especially without any documentation updates that explains what this ratio
> refers to ;-), but I suspect, unless the user is very familiar with how
> the revision graph bisection internally works, such an explanation would
> not help him find a skip ratio that is closer to the optimum than a random
> guess.  Why not use a constant ratio (or perhaps a pair of alternating
> ratios) on "bisect skip" without any new options?
>

I would agree with this assessment. It's hard enough to teach a user how to use "git bisect" as it is... and being able to have a *user* bisect a problem is worth its weight in gold.

If the algorithm I proposed earlier is too complex, here is a very simple approximation:

start:
	num = 1
	den = 2

again:
	run test (num/den)
	if (!skip)
		goto start

	num = num + 2
	if (num > den)	
		num = 1
		den = den * 2

	goto again

This creates test ratios in the following sequence:

1/2 1/4 3/4 1/8 3/8 5/8 7/8 1/16 3/16 ...

When one gets down to a small number of points this could get weird, but
as long as skip points are filtered (which looks like it's already being
 done) it should converge.

This is almost certainly suboptimal, because there are at least two possibilities as to which is better, and this isn't either:

a) one should seek points closer to the periphery, because the likelihood of a contiguous skip region goes down.

1/2 1/4 3/4 1/8 7/8 1/16 15/16 ...

b) one should seek points closer to the center in the hope of getting more information; this is better on the assumption that skip points are generally scattered:

1/2 1/4 3/4 3/8 5/8 1/8 7/8 7/16 9/16 5/16 11/16 ...

However, as before, I think there is every reason to believe something like that should be plenty good enough.

	-hpa

-- 
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.