Re: [PATCH v4 07/20] test-read-cache: print cache entries with --table

Hi,

On Wed, Mar 24, 2021 at 8:41 PM Ævar Arnfjörð Bjarmason
[off-list ref] wrote:

On Wed, Mar 24 2021, Derrick Stolee wrote:

quoted

On 3/23/21 9:24 PM, Ævar Arnfjörð Bjarmason wrote:

quoted

On Tue, Mar 23 2021, Derrick Stolee via GitGitGadget wrote:

quoted

From: Derrick Stolee <redacted>

...

quoted

quoted

It's not so much that I still disagree as I feel like I'm missing
something. I haven't gone through this topic with a fine toothed comb,
so ...

If and when these patches land and I'm using this nascent sparse
checkout support why wouldn't I want ls-files or another not-a-test-tool
to support extracting this new information that's in the index?

That's why I sent the RFC patches at
https://lore.kernel.org/git/20210317132814.30175-2-avarab@gmail.com/ (local) to
roll this functionality into ls-files.

And I recommend that you continue to pursue them as an independent
series, but I'm not going to incorporate them into this one. I'm
not going to distract from this internal data structure with changes
to user-facing commands until I think it's ready to use. As the design
document describes the plan, I don't expect this to be something I
will recommend to users until most of "Phase 3" is complete, making
the most common Git commands aware of a sparse index. (I expect to
fast-track a prototype to willing users that covers that functionality
while review continues on the mailing list.)

This series is 20 patches. Your current derrickstolee/sparse-index/wip
is another 36, and from skimming those patches & your design doc those
56 seem to be partway into Phase I of IV.

So at the rate things tend to get reviewed / re-rolled & land in git.git
it seems exceedingly likely that we'll have some part-way implementation
of this for at least a major release or two. No?

Which is why I'm suggesting/asking if we shouldn't have something like
this debugging helper as part of installed tooling, because people are
going to try it, it's probably going to have bugs and do other weird
things, and I'd rather not have to manually build some test-tool to
debug some local sparse checkout somewhere.

I'm curious why you feel it's critical that this particular piece of
debugging machinery needs to be prioritized early and exposed; in
particular, I'm not sure I follow the "people are going to try it"
assertion.  Are you the one who is going to try it or are you going to
give it to your users?  If so, what do you need out of the debugging
tool?

You are correct that this will span multiple releases; Stolee already
said he was planning to be working on this for most of 2021.  But just
because pieces of the code exist and are shipped doesn't mean it'll be
announced or supported.  For example, the git-2.30 and git-2.31
release notes were completely silent about merge-ort.  It existed in
both releases; in fact, the version that ships in git-2.31, could
theoretically be used successfully by the vast majority of users for
their daily workflow.  (But it does have known shortcomings and test
failures so I definitely did *not* want it to be announced at that
time.)

quoted

Making a change to a builtin is _forever_, and since the only
purpose right now is to expose the data in a test environment, I
don't want to adjust the builtin until either there is a real user
need or the feature has otherwise stabilized. If you want to take on
that responsibility, then please do.

That's just not the case, we have plenty of unstable debug-esque options
in various built-in commands, in fact ls-files already has a --debug
option whose docs say:

    This is intended to show as much information as possible for manual
    inspection; the exact format may change at any time.

It was added in 84974217151 (ls-files: learn a debugging dump format,
2010-07-31) and "just tacks all available data from the cache onto each
file's line" so in a way not adjusting it and using it would be a
regression, after all this is new data in the cache, so it should print
it :)

There's also PARSE_OPT_HIDDEN for other such in-tree use. Whatever the
sanity/merits of me suggesting that this specific thing be in ls-files
instead of a test-helper, it seems far fetched that something like that
hidden behind a GIT_TEST_* env var (or hidden option, --debug etc.) is
something we'd need to worry about backwards compatibility for.

So, whatever you think about the merits of including this functionality
in ls-files I think your stance of this being a no-go for adding to the
builtin is based on a false premise. It's fine to have
unstable/transitory/debug output in the builtins. We just name &
document them as such.

I also had some feedback in that series and on the earlier iteration
that I think is appropriate to be incorporated into a re-roll of this
one, which doesn't have anything to do with the question of whether we
use ls-files or the helper in the tests. Such as us showing more stuff
into the read-cache.c test-tool v.s. splitting it up making that code
needlessly convoluted.

Well:
  * you seem to be strongly opposed to test-read-cache.c containing
this code (though I don't quite follow why)
  * Stolee seems to be strongly opposed to modifying
builtin/ls-files.c until he has time to think through how builtins
should work.

So putting it in another test file that looks slightly duplicative of
test-read-cache.c might indeed be a good way out of this conundrum.
:-)

(I'm not opposed to any of the three solutions, I'm mostly chiming in
here because I'm worried about possible bubbling frustration; see
below.)

I don't see how recommending that I pursue that as an independent series
is productive for anyone. So as you re-roll this I should submit another
series on top to refactor your in-flight code & tests?

Your tone suggests some frustration; I have a suspicion there's some
lack of understanding or misreading that has occurred (perhaps on my
part too), and before that misunderstanding morphs into motive
questioning, let me see if I might be able to help...

So far, you have advocated for:
  A) Moving the checks to ls-files with a permanent new flag (--sparse)
  B) Duplicating test-read-cache.c (which is admittedly pretty small)
and then modifying the duplicate to have the new behavior, or
alternatively:
  C) Just stating files to get the information
  D) Creating new debug option(s) to ls-files so that end users can
use this in the next few releases before the feature is ready for
prime time
You also mentioned you had read just part of the series.

Option D comes with the problem that it's not at all clear who these
end-users are, why they want the option, or how we should design it.
Personally, I'm totally onboard that ls-files should generally have
the ability to show information in the index (e.g. if there are tree
entries in addition to blob entries, it should be able to show both),
but I'm not following the reasoning for why it needs to be there as
part of the early stages of development of the sparse-index feature
and who it's supposed to be helping in these next few releases.

The progression also suggests that Option B might have just been a
step along the way and that you were advocating for Option D now.  I
think it'd be easy to miss that you still had option B open and
considered it equivalently good to option D (or am I misreading?),
much like you missed how option C wasn't even relevant to the problem
at hand or option A would have introduced perpetual confusion as a
mere duplicate of --stage (in the best case scenario, anyway).
They're all easy misunderstandings.

Either my suggestions are just bad, and we shouldn't do them at all, or
it makes sense to incorporate relevant feedback in re-rolls. I'll let
other reviewers draw their own conclusions on that.

I think that's a bit unfair; Stolee has been incorporating feedback.
He even called out fixing up things at your suggestion in v4 of his
re-roll.

That's not a snarky "I'm right" b.t.w., I may honestly be full of it on
this particular topic.

But if those suggested changes are worth doing at all, then doing them
in that way seems like a massive waste of time for everyone involved, or
maybe I'm not getting what you're suggesting by pursuing them as an
independent series.

I think you should instead read it as he has no idea why this needs to
be exposed in ls-files, who these users are you are asserting will be
using it, or how to cater for their needs.  Shouldn't the person who
implements this understand those pieces to avoid a massive waste of
time?

quoted

Otherwise, I will need to eventually handle "git ls-files" being
sparse-aware when eventually removing 'command_requires_full_index',
(Phase 4) so that would be a good opportunity to adjust the
expectations.

At which point you'd be adjusting your tests that expect ls-tree format
output to using ls-files output, instead of using ls-files-like output
from the beginning?

I don't understand what you're getting at here.  I was the one who
requested Stolee make the output look like ls-trees in his original
RFC series, so if there's a problem with this style of output, I'm to
blame.  But, what is exactly the problem?  Old-style ls-files output
just isn't relevant anymore.  ls-tree prints four things: mode, type,
hash, and filename.  ls-files prints all of those except "type".  The
reason ls-files never included type before was because it was always
"blob".  This series changes that, and adds "tree" to the mix.  Once
you have different types included in the index, then ls-files has to
print all the same fields that ls-tree does...so why not make it look
similar?