Hi Ævar
On 22/03/2021 11:48, Ævar Arnfjörð Bjarmason wrote:
I recently started using rebase.rescheduleFailedExec=true and noticed
this bug in its implementation. It's conceptually a relatively simple
fix, but as noted in 3/3 rebase being a "start this operation, run
other command verbs later" has an unintuitive interaction with our
usual "command-line options override config".
!README FIRST!
Everthing after this line has no relevance to this series, it's just a
side musing on another (mis-)feature of --reschedule-failed-exec.
!/README FIRST!
There's another bug/misfeature I've noticed with
rebase.rescheduleFailedExec=true (although maybe it'll be argued by
someone that it's a feature). Let's say you run:
git rebase -x false --reschedule-failed-exec HEAD~2
You'll now land in a state of (according to our helpful PS1 code, but
also the rebase state) of 2/4. Aside: Because you're rebasing 2
commits, we have 4 because of the 2x exec. So far so good and all of
that's expected.
But now when you do "git rebase --continue" we'll fail as expected,
but not as expected (at least I find it funny) bump the count to 3/5,
then 4/6, 5/7 etc.
With my "I just read the code" hat on this makes perfect sense. Every
time we process a TODO item we bump the count of processed items, and
under --reschedule-failed-exec we simply push the command that just
failed onto the list, hence the increasing done/TODO count when a
command fails.
But I don't see how it makes any sense to expose this as UI via "git
status" and __git_ps1. I asked git to process this item of 4 sequencer
TODO items. If I fail an item it makes more sense that I just don't
get past it, not that under the hood a new identical item is
rescheduled for me. Just don't advance past the current item!
Now if I've tried X times to make the "make test" pass for each commit
in my 3-commit series I'm going to be on item 10/12 or
whatever. There's no way to just look at that and see where I'm at in
the sequence.
As an aside it would arguably make more sense to report 1/3 instead of
2/6 for the first commit of 3 with a failing -x "make test", but you
can have X number of "exec" items and a manually edited list etc.
I think counting the number of picks rather than total commands is
probably better especially now that we have reset and label commands as
well. Grouping all the execs with the previous pick doesn't seem that
unreasonable to me (though I should note that I don't use the prompt
from git.git and my prompt prints the command that we've stopped at on a
line above the prompt so it is easy for me to see where I am in the list).
So that's probably a no-go but at least once I'm used to it I know if
i'm on 4/6 I'm on commit 2/3, with --reschedule-failed-exec you'll
have no idea what 12/14 or whatever means for where you are in your
3-patch sequence, it has no relation to the TODO list you edited, just
rebase-merge's own internal state.
Getting back on topic: This just seems like needlessly exposing an
implementation detail, I also asked to "pick" a commit, but if that
item "fails" e.g. due to:
$ git rebase --continue
You must edit all merge conflicts and then
mark them as resolved using git add
We don't push a new "pick" item on the list and inflate the count, so
why would we do that for "exec"?
We do push a new pick if a pick/merge/reset fails because it would
overwrite untracked files.
As an aside I wish I could skip a rescheduled pick or exec with 'git
rebase --skip' rather than having to edit the todo list (skipping reset
commands is a bad idea though)
Just say the command failed, return
its non-zero status from "git rebase --continue", and don't advance.
Maybe it is useful to keep track of the N number of failures, and
e.g. report in __git_ps1:
master|REBASE 2/6
master|REBASE 2(tries: 1)/6
master|REBASE 2(tries: 2)/6
Something like that would be nice but I think just not incrementing the
count when we reschedule a command would be fine.
Instead of the current:
master|REBASE 2/6
master|REBASE 3/7
master|REBASE 4/8
To say I'm on 2/6, but that I've tried 3 times already and failed to
advance past it.
Anyway, I don't have patches for this side-report/rant. Looking at the
implementation it seemed more non-trivial than I was willing to
quickly fix.
Yeah, I think the implementation would need a bit of thought.
In summary I agree that it would be good to improve the count so it
makes more sense to the user
Best Wishes
Phillip
We bump the count fairly early before we even get to there being an
"exec" item, to implement this proposed view of the world we'd need to
defer that (or go back and edit it once we see "failed exec" and that
we're using --reschedule-failed-exec).
I'm not familiar enough with the sequencer internals to know if trying
that would lead us down a path of e.g. having inconsistent or bad
state if we'd die in the middle of all of that.
Ævar Arnfjörð Bjarmason (3):
rebase tests: camel-case rebase.rescheduleFailedExec consistently
rebase tests: use test_unconfig after test_config
rebase: don't override --no-reschedule-failed-exec with config
Documentation/git-rebase.txt | 8 ++++++++
sequencer.c | 5 +++++
t/t3418-rebase-continue.sh | 30 ++++++++++++++++++++++++++++--
3 files changed, 41 insertions(+), 2 deletions(-)
