Re: [PATCH 1/3] revision: complicated pathspecs disable filters

[Derrick Stolee <stolee@gmail.com>]

On 4/15/2020 8:52 PM, Taylor Blau wrote:
> On Thu, Apr 16, 2020 at 12:18:33AM +0200, Jakub Narębski wrote:
>> On Wed, 15 Apr 2020 at 20:37, Derrick Stolee <stolee@gmail.com> wrote:
>> [...]
>>> -->8--
>>> From 89beb9598daabb19e3c896bbceeb0fc1b9ccc6ca Mon Sep 17 00:00:00 2001
>>> From: Derrick Stolee <dstolee@microsoft.com>
>>> Date: Wed, 15 Apr 2020 18:04:25 +0000
>>> Subject: [PATCH] bloom: compute all Bloom hashes from lowercase
>>>
>>> The changed-path Bloom filters currently hash path strings using
>>> the exact string for the path. This makes it difficult* to use the
>>> filters when restricting to case-insensitive pathspecs.
>>>
>>> * I say "difficult" because it is possible to generate all 2^n
>>>   options for the case of a path and test them all, but this is
>>>   a bad idea and should not be done. "Impossible" is an appropriate
>>>   alternative.
>>>
>>> THIS IS A BREAKING CHANGE. Commit-graph files with changed-path
>>> Bloom filters computed by a previous commit will not be compatible
>>> with the filters computed in this commit, nor will we get correct
>>> results when testing across these incompatible versions. Normally,
>>> this would be a completely unacceptable change, but the filters
>>> have not been released and hence are still possible to update
>>> before release.
>>>
>>> TODO: If we decide to move in this direction, then the following
>>> steps should be done (and some of them should be done anyway):
>>>
>>> * We need to document the Bloom filter format to specify exactly
>>>   how we compute the filter data. The details should be careful
>>>   enough that someone can reproduce the exact file format without
>>>   looking at the C code.
>>>
>>> * That document would include the tolower() transformation that is
>>>   being done here.
>>
>> Why not modify the BDAT chunk to include version of
>> case folding transformation or other collation algorithm
>> (other transformation).that is done prior to computing
>> the Bloom filter key? Though that might be unnecessary
>> flexibility...
> 
> If this ends up being something that we want to do, I agree with
> Stolee's reasoning that this should be a breaking change. If we were,
> say, several months into having Bloom filters in a release and decided
> at that point to make the change, then: sure, supporting both by writing
> a bit in the BDAT chunk makes sense.
> 
> But, we're many months away from that state yet, and so I don't think
> the cost of rebuilding what few commit-graphs exist with bloom filters
> in them today to support both ordinary and lower-cased paths in the
> filter.
> 
> Anyway, I'm still not sold on this idea in general (nor do I understand
> it that others are), so I'll respond in more detail in another part of
> the thread...

I agree that this is not a good direction to go. I created the patch
because I was curious how difficult it would be, and it is good to have
a record of the possible direction. However, it complicates the file
format and will have unpredictable effects on the entropy (or on the
performance of history for case-colliding paths).

It is good that we have the capability to extend the filter data in
the future if we really need to.

I'll make a TODO item for myself to try writing that detailed Bloom
filter format documentation as a follow-up. In the meantime, I'll try
to close this out by responding to the feedback we have so far.

Thanks,
-Stolee