Re: [PATCH] Remove StGit patch detection from vcs_info

["Peter Grayson" <pete@xxxxxxxxxxxxx>]

On Sat, Nov 12, 2022, at 11:30 PM, Daniel Shahaf wrote:
> Peter Grayson wrote on Sat, Nov 12, 2022 at 09:46:02 -0500:
>> On Fri, Nov 11, 2022, at 6:49 AM, Daniel Shahaf wrote:
>> > [re #2]: It sounds like StGit 2.x support can be implemented at the cost
>> > of one fork(2) for those who don't use StGit and under a microsecond for
>> > those who do.  That doesn't sound like a deal breaker at all.
>> 
>> Running `stg series` with StGit 2.0 takes about 12ms in my environment.
>> StGit 1.5 it is about 32ms. Not a microsecond, but perhaps acceptable
>> nonetheless.
>> 
>
> To be clear, are these figures the duration of the «stg series
> --noprefix --applied» invocation?

Yes.

> What's the impact on people who don't have stg(1) installed, or who have
> stg(1) installed but are currently in a worktree that doesn't use StGit?
> I.e., are those figures immediately after `git init`, or in a worktree
> that has a StGit patch stack, or?

Without stg(1) installed, the cost would be however long it takes zsh to
determine that the executable is not available in $path, which is
ostensibly very fast (microseconds?).

If stg(1) is installed, but run in a repo with a branch that has not been
initialized with `stg init`, it's still about 12ms. Almost all that time
is taken just to initialize a libgit2 Repository structure, which is
used to interrogate the object database to determine whether a StGit
stack is initialized.

> In general, 32ms for everyone might be too much (what if another
> third-party tool wanted to do its own elif branch that also spent 32ms
> for everyone?  That'd be 64ms in total and counting…).

Well, that was my original thought as well and why my first patch simply
removed StGit support from vcs_info. Note that 32ms is on a particularly
fast machine and I stand by my assertion that Python startup time can be
on the order of hundreds of milliseconds on more modest machines and/or
with older versions of Python.

> However, if the 32ms would only be seen by users of an EOL'd stg(1),
> we can relax the threshold a bit.  On the other hand, if it's 32ms
> just in cases where stg(1) is used… well, there doesn't seem to be
> much we /can/ do there.

The relevant scenarios:

* stg(1) not installed => ~free
* stg(1) < 2.0 installed => 30+ ms
* stg(1) >= 2.0 installed => ~10 ms

The worst case scenario would be for someone where stg(1) <2.0 is
installed, but not used. This would create overhead in vcs_info without
any value for the user. All other scenarios seem to end up with the
user getting what they paid for.

> Perhaps hide some or all of the work behind an opt-in switch.  (For
> instance, the default settings show only the topmost applied patch, so
> if it's possible to tell stg(1) not to emit any other patches, that
> might help the code finish more quickly.)

Limiting the output to only the topmost patch will not have a material
effect on the runtime because of how the overhead is dominated by
libgit2 initialization. So an opt-in switch would have be all or
nothing; either try to run stg(1) or not.

Also worth noting is that changing StGit support in vcs_info opt-in
would be a bit of a regression and would add more configuration
surface area. Not sure if/how the zsh project does versioning for
something like this. I did not include such a switch in my latest
patch because of these issues.

> From the "Is your computer plugged in?" department: Is that 32ms figure
> on hot disk cache with .pyc files already having been generated?

Yes. I used hyperfine(1) to do these measurements.

Pete