Re: [PATCH] Reduce cost of deletion in levenstein distance (4 -> 3)

From: Junio C Hamano <gitster@pobox.com>
Date: May 29, 2012, 18:25 UTC
Message-ID: <7v4nqy6e0k.fsf@alter.siamese.dyndns.org>
In-Reply-To: <1338134578-29011-1-git-send-email-Matthieu.Moy@imag.fr>

Matthieu Moy <Matthieu.Moy@imag.fr> writes:

Show 15 quoted lines

> Before this patch, a character deletion has the same cost as 2 swaps, or
> 4 additions, so Git prefers suggesting a completely scrambled command
> name to removing a character. For example, "git tags" suggests "stage",
> but not "tag".
>
> By setting the deletion cost to 3, we keep it higher than swaps or
> additions, but prefer 1 deletion to 2 swaps. "git tags" now suggests
> "tag" in addition to staged.
>
> Signed-off-by: Matthieu Moy <Matthieu.Moy@imag.fr>
> ---
> The RFC sent earlier [1] didn't receive negative comments, so I think this
> is a good change.
>
> http://thread.gmane.org/gmane.comp.version-control.git/196457

Lack of objections is never a good reason to assume it is a good change. In this particular case, I think the reason why you saw no comments, either positive or negative, was because nobody came up with a more "scientific" way to judge the weighting. Choosing between "tag" and "stage" given "tags" is just one datapoint, but does not convince anybody there are not surprising combinations where this change affects negatively.

I wonder if we can mechanically find a set of parameters that optimally separates the built-in commands by computing N^2 distances (e.g. compute "tag" and all other command names and record minimum. Do so for all other commands. Now fudge the parameters and repeat to see if it results in better minimum separation. Something like that).

Having said all that, until somebody comes up with a better method of judging, I'd say that the best thing we could do is to apply this patch and see if anybody finds a "surprising" case where this leads to a regression.

Thanks.