Re: Train station analogy, was Re: [PATCH v3 00/15] [RFC] Upstreaming the Scalar command

On Tue, Sep 14 2021, Derrick Stolee wrote:

On 9/14/2021 2:09 PM, Ævar Arnfjörð Bjarmason wrote:

quoted

On Tue, Sep 14 2021, Junio C Hamano wrote:

quoted

Johannes Schindelin [off-list ref] writes:

At least to me, how this Makefile for Scalar should interact with
the overall build process does not mesh well with the story about
hwo direction to and history of the station are unrelated.  If we
plan to start from contrib/ and eventually want to make it a part
of the core Git (i.e. "git scalar <subcmd> ..." becomes just like
"git bisect <subcmd> ..."), we would eventually need to see the
recipe needed for including "bisect" and "scalar" work the same
way, no?

We should definitely work to find a better way to describe our
vision for how _the ideas in Scalar_ can be adopted into Git proper.

Before this series, we were adding functionality to Git that allowed
Scalar to simplify to just a CLI that configures Git features. This
submission allows that CLI to be available via an opt-in compile flag.
This should allow more users to try out the ideas and perhaps we find
the things that really work for people (and almost more importantly,
the ideas that are _too_ opinionated).

Yeah that makes sense. I think it looks like a useful command & it's
already useful to users.

I haven't been suggesting any changes to what gets installed here, FWIW
I think we could be even more aggressive on that front, e.g. shipping it
unconditionally in libexec, maybe with an optional switch for
/usr/bin/scalar, or to ship "scalar" symlinked to "git" and have it
route to the top-level scalar command depending on argv.

I dabbled in that a bit locally, FWIW it seems if anything even easier
to do than the approaches we've discussed so far, but I wanted to focus
on providing the same behavior in terms of build system maintenance.

But the way I see Scalar being fully incorporated into Git is not as
a "git scalar <foo>" command or even having "scalar" be included by
default. Instead, perhaps a new builtin would need to be created and
its CLI would need to be presented and reviewed with significant
attention to long-term support in the Git project. Having Scalar as
a testing ground for these ideas seems like a positive way forward.

*Nod*

This is a big reason why we think that contrib/ is a good place for
it to exist.

Here's where you and Johannes lose me. There's some rationale in your
minds for why sticking it in contrib is the obvious way to go. So far
you've been describing how it'll look to users etc, how it's arranged in
our source tree only matters to git.git developers.

I think the actual reason is to carve in advance some subjective
ownership/apartness or whatever for this thing, if that's the case I
think documentation/commit messages would also work.

I really don't care much if something lives in contrib or not in the
abstract, but various integration around builds in Makefile makes that
much easier in practice, and if "make install" looks the same...

quoted

quoted

I am getting the impression that such a unified build process is
Ævar wants to see at the end, I am not even sure if you do from
the above "analogy".  Cool down a bit, perhaps?

I agree that the temperature of this thread has gotten a bit
heated. I think there is something valuable to be gained from
each perspective, but not in a way that either has presented it.

Thanks, hopefully we can keep it jovial going forward. If you've got any
(either on-list or off-list) feedback about how I can improve my side of
that it would be most welcome.

quoted

quoted

The following assumes that you share the goal of making "git
scalar" just like "git bisect"---another first class citizen of
Git toolbox, the user can choose to use it or the user may not
have a need to interact with it, but it exists there by default
and is not an opt-in add-on component.

I would understand it if your plan is to convert to a unified
build procedure at the very end of the upstreaming process, and
not while you populate contrib/ with more and more scalar stuff,
because the Makefile bits for the entire scalar, while not yet
upstreamed, has already been written as a separate procedure and
having to convert the whole thing upfront before you can start
trickle parts would mean you need to (re)start the process.  And
I would even be sympathetic if you felt it like a distraction.

But at least I view it as a step that needs to happen sometime
between now and at the end.  I do not yet have an opinion on
which one is more pleasant, between (1) having to deal with a
single Makefile that needs to be aware of two different locations
*.[ch] lives in, and (2) having to deal with two Makefiles that
duplicates definitions and risks them needlessly diverging.

Since we already need to modify the root Makefile, I think having
the root Makefile add the files from contrib/scalar from an
optional flag is a great way to reduce duplication across multiple
Makefiles while also maintaining the Scalar is compiled optionally.

One big goal is to minimize how often we need to update Scalar. I
can see things like adjusting the recommended config once per
release cycle based on which new features are available. I don't
really want to be spending time updating the Makefile to match a
contribution that was already carefully reviewed and tested. I
also don't want to put the burden of updating contrib/scalar upon
those contributors.

I'd think not having large parts of t/Makefile & Makefile should ease
that maintenance burden for you & others.

quoted

For what it's worth what I had on top of this is not (1) or (2), but a
(0): I.e. there isn't a contrib/scalar anymore, I moved:

    contrib/scalar/scalar.c -> scalar>     contrib/scalar/scalar.txt -> Documentation/scalar.txt
    contrib/scalar/t9099-scalar.sh -> t/t9099-scalar.sh

We build, test, and otherwise check (e.g. "make check-docs") it by
default, what we don't do is install it unless you ask. You need to run:

    # Or any other install* target
    make install install-doc INSTALL_SCALAR=YesPlease

It could be be kept in contrib/scalar/ even with that sort of approach,
and it would still be simpler than the two-Makefile approach.

I think keeping it in contrib/scalar is best for now. But I do
agree that a single Makefile has benefits.

I noted the "why contrib" above.

One early suggestion from a while back was to modify git.c to
handle the "scalar" executable as well as the "git" executable,
specifically to reduce duplication handling options such as

  -c config.key=value
  -C worktree
  --exec-path

and similar commands. While our duplication of the "-c" option
does add similar code in a second place, these other options
are less critical for Scalar, especially in its current version.
I think refactoring the code in git.c to cater to the "scalar"
executable is at least premature. If we want to pursue these
other options in the future, then that refactoring could happen
as a separate discussion after the rest of the build system and
CLI have been figured out.

As noted above that seems like a sensible way forward, I hadn't noticed
how much of git.c's setup was copied into scalar.c.

It seems to me that it wouldn't be that hard, on the order of the
existing setup code or less. I.e. just make "git.c" learn that it may be
running some arbitrary command name, and do some options parsing, but
and finally dispatch to a cmd_scalar(). IOW mostly like a built-in.

_Perhaps_ Johannes still had that level of integration in his
head when responding to the single-Makefile recommendations.

quoted

But just moving the code, tests and documentation where everything else
lives cuts down an all sorts of special cases, file globs in various
places (e.g. doc lints) will just work and won't need adjustment.

quoted

I also would understand it if the reason why you want to keep the
top-level Makefile as intact as possible because you sense a high
probability that scalar will stay in contrib/ and even turn out
to be a failure.  Keeping the build procedure separated certainly
will keep it easier to yank it out later.  But I do not think
such a case is quite likely.

For what it's worth the WIP patch(es) I have on top of it will probably
make such a thing even easier, not that removing it from the tree would
be much of a problem in either case. It's mostly a few lines added to
lists in various places in the Makfile.

Do you have a version of these patches available for adaptation
into this series? I'd like to take a look and see what it would
look like to squash them into this series. Forgive me if I just
missed the link. (I see the diff you posted earlier in this thread.)

I've got it at
https://github.com/avar/git/tree/avar-dscho/scalar-the-beginning-normalize-Makefile

Not very ML-ready, and soft-depends on some other Makefile cleanups I
thought I'd do & still haven't untangled and submitted. Soft-depends as
in this can easily be done on master, but some of the variable names
etc. are quite confusing there.

But you should be able to check it out, it'll build, test and install if
you run "install" with "INSTALL_SCALAR=Y".

You may run into on everly eager new (but unrelated to this, I just
merged it on top) Makefile assertion I'm experimenting with, just
comment out the relevant line in the Makefile if that happens,
i.e. something like this error:

    Makefile:3608: *** "please sort and de-duplicate BUILT_INS_EXTRA!".  Stop.

quoted

If I were to clean this up properly most of the changes would be
teaching the Makefile that it can build N number of named top-level
"special" commands that get dropped into bin/, not just the "git" we
hardcode now.

This is an interesting idea for revamping how adjacent tools are
compiled and shipped with Git from contrib/ (or possibly elsewhere
if we decided to start including more things as "blessed helpers".

As a complete aside: I'm interested in using the sparse-checkout
feature as I work on the Git codebase, just to make sure I hit
pain points before any other user.

This is the best that I could do for my purposes:

$ git sparse-checkout list
.github
Documentation
builtin
compat
contrib/scalar
ewah
git-gui
gitk-git
gitk-gui
gitweb
mergetools
negotiator
perl
po
refs
sha1dc
sha256
t
templates
trace2
xdiff

And 'git status' reports that this includes 97% of the tracked
files. Perhaps there are ways to make this be smaller by having
make skip building things like git-gui if the directory doesn't
exist. Another idea would be to skip any logic around translating
messages if the 'po' directory is missing.

For git-gui in particular NO_TCLTK=Y should do it.

The reason I bring this up is that I'm interested in finding
ways to make our build system be streamlined a bit using the
presence of directories as a way to opt in/out of certain build
outputs. Since Scalar is being added as a new component, this is
a good opportunity to establish a pattern that works for this
effort, too.

Sure, the hard part isn't that you can't grep out nonexisting files or
directories when building where we now use a glob.

It's that everything downstream of that, i.e. tests, installation
etc. is going to have to work properly in the face of arbitrary parts of
what the developer who tested the code expected going missing.

Which is why we've generally carved out very specific things, usually
along the boundaries of installed dependencies.