Re: DEVEL: Help with feature implementation

From: Antonio Russo <aerusso@aerusso.net>
Date: Jan 19, 2021, 14:52 UTC
Message-ID: <9fc9d752-06f4-e75d-1549-f257fd34e0c3@aerusso.net>
In-Reply-To: <xmqq7do9iuqj.fsf@gitster.c.googlers.com>

On 1/18/21 6:24 PM, Junio C Hamano wrote:

Show 9 quoted lines

> Antonio Russo <aerusso@aerusso.net> writes:
> 
>> As a side note, would this list be willing to look at patches that remove
>> the need to use revs->limited?  Adding new features would be much easier if
>> we could restrict git to use streaming algorithms for these simplifications.
> 
> Do you mean you will write new implementations of operations like
> "--simplify-merges", "--ancestory-path" and all its friends without
> the need for the "limited" bit?

Yes.

Show 11 quoted lines

> Willing to look at?  I do not know about others, but sure, it would
> make be extremely excited.
> 
> You may be able to rewrite some other operations that implicitly
> require the limited bit into an incremental implementation, but I am
> skeptical that you can do "simplify-merges" in a streaming fashion
> in such a way that we'd dig a bit but not all the way down before
> making decisions on which commits should be included in the output
> and how their parenthood relationship should appear etc.  I and
> Linus tried independently and we did not get that far in our
> attempts (note: there wasn't commit-graph back then).

Yeah, it's a big task (but, it'd be useful work, rather than doing another rewrite of my feature to make it work when revs->limited). Each of the flags (simplify-merges, ancestry-path, etc.) is its own little sub-project.

I haven't thought about any one flag specifically, but the fact that two complete code paths exist right now just seem like a nightmare. I'm not necessarily interested in making the implementations faster, but rather getting rid of the duplicated code path.

It's also a long-term goal, but it's nice to hear that people would be interested in it.

> If you are talking about precomputed stuff like commit-graph, that
> may change the equation, but we'd prefer to have the system work
> without requiring these auxiliary data (in other words, it would be
> fine to use them as optimization).

No, I understand that generation numbers can only be used as an optimization (i.e., they might all be GENERATION_NUMBER_INFINITY).