Re: [PATCH 00/25] progress.c: various fixes + SZEDER's RFC code

On Wed, Jun 23 2021, Randall S. Becker wrote:

On June 23, 2021 1:48 PM, Ævar Arnfjörð Bjarmason wrote:

quoted

quoted

On Mon, Jun 21, 2021 at 02:59:53AM +0200, Ævar Arnfjörð Bjarmason wrote:

quoted

On Sun, Jun 20 2021, SZEDER Gábor wrote:

quoted

Splitting off from:


https://public-inbox.org/git/cover-0.2-0000000000-20210607T144206Z-
avarab@gmail.com/T/#me5d3176914d4268fd9f2a96fc63f4e41beb26bd6

On Tue, Jun 08, 2021 at 06:14:42PM +0200, René Scharfe wrote:

quoted

I wonder (only in a semi-curious way, though) if we can detect
off-by-one errors by adding an assertion to display_progress()
that requires the first update to have the value 0, and in
stop_progress() one that requires the previous display_progress()
call to have a value equal to the total number of work items.  Not
sure it'd be worth the hassle..

I fixed and reported a number of bogus progress lines in the past,
the last one during v2.31.0-rc phase, so I've looked into whether
progress counters could be automatically validated in our tests,
and came up with these patches a few months ago.  It turned out
that progress counters can be checked easily and transparently in
case of progress lines that are shown in the tests, i.e. that are
shown even when stderr is not a terminal or are forced with
'--progress'.  (In other cases it's still fairly easy but not quite
transparent, as I think we need changes to the progress API; more
on that later in a separate
series.)

I've also been working on some progress.[ch] patches that are mostly
finished, and I'm some 20 patches in at the moment. I wasn't sure
about whether to send an alternate 20-patch "let's do this (mostly) instead?"
series, hence this message.

Much of what you're doing here becomes easier after that series, e.g.
your global process struct in 2/7 is something I ended up
implementing as part of a general feature to allow progress to be
driven by either display_progress() *or* the signal handler itself.

It's difficult to know who should rebase onto who without seeing one
half of the patches.

I was sort of hoping he'd take me word for it, but here it is. Don't say I didn't warn you :)

quoted

I couldn't find a link to them anywhere (even if they are only
available in your fork in a pre-polished state) despite looking, but
my apologies if they are available and I'm just missing them.

FWIW it's avar-szeder/progress-bar-assertions in https://github.com/avar/git.git, that repo contains various functioning and not-so-
functioning code.

https://github.com/avar/git/tree/meta/ is my version of the crappy scripts we probably all have some version of for building my own git,
things that are uncommented in series.conf is what I build my own git from.

quoted

In general, I think that these patches are clear and are helpful in
pinning down issues with the progress API (which I have made a hadnful
of times in the past), so I would be happy to see them picked up.

Here's all 25 patches (well, around 20 before) that I had queued up locally and fixed up a bit.

The 01/25 is something I submitted already as https://lore.kernel.org/git/patch-1.1-cba5d88ca35-20210621T070114Z-avarab@gmail.com (local);
hoping to get this in incrementally.

The 12/25 is my own version of that "global progress struct, 11/25 is the first of many bugs SZEDER missed in his :)

18/25 is the first step of the UI I was going for, the signal handler can now drive the progress bar, so e.g. during "git gc" we show (at least
for me, on git.git), a "stalled" message just before we start the actual count of "Enumerating Objects".

After that was in I was planning on adding config-driven support to show a "spinner" when we stalled in that way, config-driven because
you could just scrape e.g. https://github.com/sindresorhus/cli-spinners/blob/main/spinners.json
into your own config. See
https://jsfiddle.net/sindresorhus/2eLtsbey/embedded/result/ :)

19-23/25 is my grabbing of SZEDER's patches that I'm comfortable labeling as "PATCH", I think they work, but no BUG() assertions yet. I
left out the GIT_TEST_CHECK_PROGRESS parts, since my earlier works set things up to do any BUG() we trust by default.

22/25 is what I think we should do instead of SZEDER's 6/7
(http://lore.kernel.org/git/20210620200303.2328957-7-szeder.dev@gmail.com (local))
I don't think this "our total doesn't match at the end" is something we should always BUG() on, for reasons explained there.

I am sympathetic to doing it by default though, hence the
stop_progress_early() API, that's there to allow select callers to bypass his BUG(...) assertion.

24/25 and 25/25 are "RFC" and a rebased+modified version of SZEDER's
BUG(...) assertions.

His series passes the test suite, but actually severely break things things. It'll make e.g. "git commit-graph write" BUG(...) out. The reason
the tests don't catch it is because we have a blind spot in the tests.

Namely, that most things that use the progress bar API use isatty() to check if they should start_progress(). If you run the tests as e.g.
(better ways to do this, especially in parallel, most welcome):

   for t in t[0-9]*.sh; do if ! ./$t -vixd; then echo $t bad; break; fi; done

You can discover various things that his series BUG()'s on, I fixed a couple of those myself, it's an early part of this series.

But we'll still have various untested for BUG()'s even then, this is because you *also* have to have the test actually emit a "naked"
progress bar on stderr, if the test itself e.g. pipes fd 2 to a file it won't work.

I created a shitty-and-mostly-broken throwaway change to search-replace all the guards of "start_progress(...)" to run unconditionally, and
convert all the "delayed" to the non-delayed version. That'll find even more BUG()'s where SZEDER's series still needs to be fixed (and also
some unrelated segfaults, I gave up on it soon after).

Even if we fix that I wouldn't trust it, because a lot of the progress bars we have depend on the size and shape of the data we're
processing, e.g. the bug I fixed in 11/25. If people find this BUG() approach worth pursuing I think it would be better to make it an opt-in
flag we convert one caller at a time to.

For some it's really clear that we could assert it, for others such as the commit-graph it's much more subtle, we're in some callback after
setting a "total", that callback does a "break", "continue" etc. in various places, all depending on repository data.

It's not easy to reason about that and be certain that we can hold to the estimate. If we get it wrong someone's repo in the wild won't fully
GC because of the overly eager BUG().

If SZEDER wants to pursue it I think it'll be easier on top of this series, but personally I really don't see the point of spending effort on it.

We should really be going in the other direction, of having more fuzzy ETAs, not less.

E.g. we often have enough data at the start of "Enumerating Objects"
to give a good-enough target value, that it's 5-10% off isn't really the point, but that the user looking at it sees something better than a
dumb count-up, and can instead see that they'll probably be looking at it for about a minute. Now our API is to give no ETA/target if we're
not 100% sure, it's not good UX.

So trying to get the current exact count/exact percentage right seems like a distraction to me in the longer term. If anything we should
just be rounding those numbers, showing fuzzy ETAs instead of percentages if we can etc.

SZEDER Gábor (4):
 commit-graph: fix bogus counter in "Scanning merged commits" progress
   line
 entry: show finer-grained counter in "Filtering content" progress line
 progress: assert last update in stop_progress()
 progress: assert counting upwards in display()

Ævar Arnfjörð Bjarmason (21):
 progress.c tests: fix breakage with COLUMNS != 80
 progress.c tests: make start/stop verbs on stdin
 progress.c tests: test some invalid usage
 progress.c tests: add a "signal" verb
 progress.c: move signal handler functions lower
 progress.c: call progress_interval() from progress_test_force_update()
 progress.c: stop eagerly fflush(stderr) when not a terminal
 progress.c: add temporary variable from progress struct
 midx perf: add a perf test for multi-pack-index
 progress.c: remove the "sparse" mode nano-optimization
 pack-bitmap-write.c: add a missing stop_progress()
 progress.c: add & assert a "global_progress" variable
 progress.[ch]: move the "struct progress" to the header
 progress.[ch]: move test-only code away from "extern" variables
 progress.c: pass "is done?" (again) to display()
 progress.[ch]: convert "title" to "struct strbuf"
 progress.c: refactor display() for less confusion, and fix bug
 progress.c: emit progress on first signal, show "stalled"
 midx: don't provide a total for QSORT() progress
 progress.c: add a stop_progress_early() function
 entry: deal with unexpected "Filtering content" total

cache.h                          |   1 -
commit-graph.c                   |   2 +-
csum-file.h                      |   2 -
entry.c                          |  12 +-
midx.c                           |  25 +-
pack-bitmap-write.c              |   1 +
pack.h                           |   1 -
parallel-checkout.h              |   1 -
progress.c                       | 391 ++++++++++++++++++-------------
progress.h                       |  50 +++-
reachable.h                      |   1 -
t/helper/test-progress.c         |  54 +++--
t/perf/p5319-multi-pack-index.sh |  21 ++
t/t0500-progress-display.sh      | 247 ++++++++++++++-----
14 files changed, 537 insertions(+), 272 deletions(-)  create mode 100755 t/perf/p5319-multi-pack-index.sh

Is there provision for disabling progress on a per-command basis? My
use case is specifically in a CI/CD script, being able to suppress
progress handling. The current Jenkins plugin does not appear to have
provision for hooking into a mechanism, which makes things get a bit
wonky when a job runs with a pseudo-tty (as provided by Jenkins
through SSH/RMI).
-Randall

There isn't, some commands support --no-progress, but it's hit and miss.

You can then set the undocumented GIT_PROGRESS_DELAY=99999999 (or some
really big number) to suppress more of them.

We could just add it as a top-level "git --no-progress" option I
suppose...

Probably better would be to detect such not-a-terminals somehow, I think
at some point our own gc.log was a victim of this.