Re: RFE: "git bisect reverse"

[Ealdwulf Wuffinga <ealdwulf@googlemail.com>]

On Thu, May 28, 2009 at 10:54 PM, H. Peter Anvin <hpa@zytor.com> wrote:

> How about simply modelling it linearly, with 100% probability for known
> skip point, 0% for a known good/bad point, and a linear gradient in
> between?  It's probably a good enough model.  In practice, it will
> vastly overestimate the probability of a skip, so if a linear model
> turns out to be too conservative, I would probably just try to model it
> as a higher-order power function.

Sounds plausible. It's not obvious how to generalise it to a DAG, though.

What's easier to implement is simple geometric decay (from a
probability of 1 at the commit of an actual skip).  It even has a
plausible rationale - the probability that someone notices the
brokenness
and fixes it is probably a constant, which would lead to geometric
decay. That probably doesn't reflect what happens when someone breaks
a whole swath of stuff retrospectively somehow, though.

I've implemented geometric (and arithmetic) decay in bbchop, with a
configurable decay factor. With a factor sufficiently close to one
(eg, 0.99) it can be persuaded to hop a reasonable distance, which
seems to scale to a certain extent with the number of commits left, so
hopefully it won't be necessary to fiddle with the factor a lot.

http://github.com/Ealdwulf/bbchop/tree/master

Ealdwulf