Hi Christian and Elijah
On 07/04/2023 08:24, Christian Couder wrote:
# Intro
`git replay` has initially been developed entirely by Elijah Newren
mostly between June and July 2022 at:
https://github.com/newren/git/commits/replay
I took over a few months ago to polish and upstream it as GitLab is
interested in replacing libgit2, and for that purpose needs a command
to do server side (so without using a worktree) rebases, cherry-picks
and reverts.
I reduced the number of commits and features in this first patch
series, compared to what Elijah already developed. Especially I
stopped short of replaying merge commits and replaying
interactively. These and other features might be upstreamed in the
future after this patch series has graduated.
Thanks to Elijah, Patrick Steinhardt and Dscho for early reviews and
discussions.
Based on ae73b2c8f1 (The seventh batch, 2023-04-04)
Thanks to both of you for working on this it looks very interesting.
I've had a quick read over the patches and I ended up slightly confused
as to exactly what the aim of this series is. My main confusion is
whether "replay" is intended to be a plumbing command or a porcelain
command. The use case above suggests plumbing and there are patches that
take it in that direction by removing any diagnostic output and stopping
it update any refs. But then it is marked as porcelain in
command-list.txt and there are patches that do things like
unconditionally updating the index and worktree when there are conflicts
that stop it working in bare repositories. I've left some comments below
# Quick Overview (from Elijah)
`git replay`, at a basic level, can perhaps be thought of as a
"default-to-dry-run rebase" -- meaning no updates to the working tree,
or to the index, or to any references. However, it differs from
rebase in that it:
* Works for branches that aren't checked out
* Works in a bare repository
* Can replay multiple branches simultaneously (with or without common
history in the range being replayed)
* Preserves relative topology by default (merges are replayed too)
* Focuses on performance
* Has several altered defaults as a result of the above
I sometimes think of `git replay` as "fast-replay", a patch-based
analogue to the snapshot-based fast-export & fast-import tools.
# Reasons for diverging from cherry-pick & rebase (from Elijah)
There are multiple reasons to diverge from the defaults in cherry-pick and
rebase.
* Server side needs
* Both cherry-pick and rebase, via the sequencer, are heavily tied
to updating the working tree, index, some refs, and a lot of
control files with every commit replayed, and invoke a mess of
hooks[1] that might be hard to avoid for backward compatibility
reasons (at least, that's been brought up a few times on the
list).
* cherry-pick and rebase both fork various subprocesses
unnecessarily, but somewhat intrinsically in part to ensure the
same hooks are called that old scripted implementations would
have called.
To clarify, since 356ee4659bb (sequencer: try to commit without forking
'git commit', 2017-11-24) cherry-pick and rebase do not fork
subprocesses other than hooks for the cases covered by this patch series
(i.e. they do not fork "git commit" for simple picks).
* "Dry run" behavior, where there are no updates to worktree, index,
or even refs might be important.
* Should not assume users only want to operate on HEAD (see next
section)
* Decapitate HEAD-centric assumptions
* cherry-pick forces commits to be played on top of HEAD; inflexible.
* rebase assumes the range of commits to be replayed is
upstream..HEAD by default, though it allows one to replay
upstream..otherbranch -- but it still forcibly and needlessly
checks out otherbranch before starting to replay things.
I agree it would be nice to be able to restrict the range of commits
replayed, especially when replaying merges. The comment about checking
out other branch is out of date since 767a9c417eb (rebase -i: stop
checking out the tip of the branch to rebase, 2020-01-24)
* Assuming HEAD is involved severely limits replaying multiple
(possibly divergent) branches.
I'm not sure how true this is anymore, since 89fc0b53fdb (rebase: update
refs from 'update-ref' commands, 2022-07-19) the sequencer can update
multiple branches. The issue with divergent branch is with command line
arguments and the todo list generation rather than the capabilities of
the sequencer.
* Once you stop assuming HEAD has a certain meaning, there's not
much reason to have two separate commands anymore (except for the
funny extra not-necessarily-compatible options both have gained
over time).
I agree having a unified command at the plumbing level certainly makes
sense.
* (Micro issue: Assuming HEAD is involved also makes it harder for
new users to learn what rebase means and does; it makes command
lines hard to parse.
That's an interesting point, I wonder if operating on branches that are
not checked out is potentially confusing for new user though.
Not sure I want to harp on this too much, as
I have a suspicion I might be creating a tool for experts with
complicated use cases, but it's a minor quibble.)
* Performance
* jj is slaughtering us on rebase speed[2]. I would like us to become
competitive. (I dropped a few comments in the link at [2] about why
git is currently so bad.)
* From [3], there was a simple 4-patch series in linux.git that took
53 seconds to rebase. Switching to ort dropped it to 16 seconds.
While that sounds great, only 11 *milliseconds* were needed to do
the actual merges. That means almost *all* the time (>99%) was
overhead! Big offenders:
* --reapply-cherry-picks should be the default
I agree that can be a performance hit if there are a lot of upstream
commits, but it is also a usability feature as it means we don't stop
and ask the user what to do with the commits that have been upstreamed
which wastes more of their time. I think maybe we want different
defaults for the server use case than the user replaying commits or
perhaps default to dropping commits that become empty.
* can_fast_forward() should be ripped out, and perhaps other extraneous
revision walks
We should look at doing that at least for the merge backend which has
skip_unnecessary_picks(). I think it is useful to tell the user that the
branch was not updated by the rebase though.
* avoid updating working tree, index, refs, reflogs, and control
structures except when needed (e.g. hitting a conflict, or operation
finished)
Not writing to disc unless we need to is sensible. Having said that for
interactive rebases I do find having HEAD's reflog record all the picks
useful to unpick what went wrong if mess something up.
* Other performance ideas:
* single-file control structures instead of directory of files
I like the idea as it should make it easier to keep the on disc state
consistent, but I'm not sure how much of an issue that is in practice as
we only read/write the files once each time git is run. The bigger slow
down is writing the author script, commit message, list of rewritten
commits, todo list and done files with each pick.
* avoid forking subprocesses unless explicitly requested (e.g.
--exec, --strategy, --run-hooks). For example, definitely do not
invoke `git commit` or `git merge`.
Good, that matches what the sequencer does for non-merge commits when
we're not editing the commit message.
* Sanitize hooks:
* dispense with all per-commit hooks for sure (pre-commit,
post-commit, post-checkout).
I agree we should not be running those (we don't run the pre-commit hook
anyway). However we had a bug report when cherry-pick stopped running
the "prepare-commit-msg" hook (see
https://lore.kernel.org/git/CAKdAkRQuj1hfKeckjuR2oP+8C1i+ZR36O-+aRYif4ufaS_zs+w@xxxxxxxxxxxxxx/).
That shouldn't matter for the server but we should bear it in mind when
it comes to other use cases.
* pre-rebase also seems to assume exactly 1 ref is written, and
invoking it repeatedly would be stupid. Plus, it's specific
to "rebase". So...ignore? (Stolee's --ref-update option for
rebase probably broke the pre-rebase assumptions already...)
If replay is a plumbing command then skipping the pre-rebase hook makes
sense as scripts can call it themselves if they want to. For a porcelain
command keeping a hook that can prevent it from rewriting commits that
are already upstream (which I think is one of the main uses of the
pre-rebase hook) would be good.
* post-rewrite hook might make sense, but fast-import got
exempted, and I think of replay like a patch-based analogue
to the snapshot-based fast-import.
If we don't call the hook it would be good to have a option that outputs
that information so scripts can request it if they want. Also we should
think about if/when we want to update the notes associated with replayed
commits.
* When not running server side, resolve conflicts in a sparse-cone
sparse-index worktree to reduce number of files written to a
working tree. (See below as well)
* [High risk of possible premature optimization] Avoid large
numbers of newly created loose objects, when replaying large
numbers of commits. Two possibilities: (1) Consider using
tmp-objdir and pack objects from the tmp-objdir at end of
exercise, (2) Lift code from git-fast-import to immediately
stuff new objects into a pack?
* Multiple branches and non-checked out branches
* The ability to operate on non-checked out branches also implies
that we should generally be able to replay when in a dirty working
tree (exception being when we expect to update HEAD and any of the
dirty files is one that needs to be updated by the replay).
* Also, if we are operating locally on a non-checked out branch and
hit a conflict, we should have a way to resolve the conflict without
messing with the user's work on their current branch.
That sounds tricky to do in a user friendly way.
* Idea: new worktree with sparse cone + sparse index checkout,
containing only files in the root directory, and whatever is
necessary to get the conflicts
If the user has not asked for a sparse checkout then this could be
surprising. Sometimes I find it helpful to be able to poke about in
other source files when resolving a conflict. I also often build and
test after resolving a conflict which requires more than just the
conflict to be checked out.
* Companion to above idea: control structures should be written to
$GIT_COMMON_DIR/replay-${worktree}, so users can have multiple
replay sessions, and so we know which worktrees are associated
with which replay operations.
We certainly want some way of making sure we only update a given ref in
one replay session, and have checks to for whether the ref is checked
out anywhere as we do now for rebase --update-refs. That seems to be
lacking in the patches adding ref updating in this series.
- [1] https://lore.kernel.org/git/pull.749.v3.git.git.1586044818132.gitgitgadget@xxxxxxxxx/
- [2] https://github.com/martinvonz/jj/discussions/49
- [3] https://lore.kernel.org/git/CABPp-BE48=97k_3tnNqXPjSEfA163F8hoE+HY0Zvz1SWB2B8EA@xxxxxxxxxxxxxx/
# Important limitations
* The code die()s if there are any conflict. No resumability. No nice
output. No interactivity.
I can see that on a server you might not want any output, but if I run
it locally it would be nice to have a message saying which paths have
conflicts. Maybe we could add a --quiet flag for the server rather than
removing the existing messages?
* No replaying merges, nor root commits. Only regular commits.
That is a reasonable place to start.
* Signed commits are not properly handled. It's not clear what to do
to such commits when replaying on the server side.
Yes on the server where you don't have access to the signing key there
is not much need for replay to have a signing option.
# Commit overview
* 1/14 replay: introduce new builtin
This creates a minimal `git replay` command by moving the code
from the `fast-rebase` test helper from `t/helper/` into
`builtin/` and doing some renames and a few other needed changes.
* - 2/14 replay: start using parse_options API
- 3/14 replay: die() instead of failing assert()
- 4/14 replay: introduce pick_regular_commit()
- 5/14 replay: don't simplify history
- 6/14 replay: add an important FIXME comment about gpg signing
- 7/14 replay: remove progress and info output
- 8/14 replay: remove HEAD related sanity check
These slowly change the command to make it behave more like a
regular commands and to start cleaning up its output.
* 9/14 replay: very coarse worktree updating
Make it handle conflicts in a very coarse way. This might not
work on bare repos, but it allows existing tests to pass and it's
nice to help cli users a bit when they get conflicts.
* 10/14 replay: make it a minimal server side command
After the cleaning up in previous ommits, it's now time to
radically change the way it works by stopping it to do ref
updates, to update the index and worktree, to consider HEAD as
special. Instead just make it output commands that should be
passed to `git update-ref --stdin`.
* - 11/14 replay: use standard revision ranges
- 12/14 replay: introduce guess_new_base()
- 13/14 replay: add different modes
- 14/14 replay: stop assuming replayed branches do not diverge
These finish the clean up and add new interesting features at
the same time, as well as related documentation and tests.
# Note about tests and documentation
Note that the `fast-rebase` test helper was used before this series in
t6429-merge-sequence-rename-caching.sh
So when `git replay` is created from `fast-rebase` in patch 1/14, this
test script is also converted to use `git replay`. This ensures that
`git replay` doesn't break too badly during the first 10 patches in
this patch series.
Tests and documentation are introduced specifically for `git replay`
only in 11/14 and later patches as it doesn't make much sense to
document and test behavior that we know is going to change soon. So
it's only when the command is crystalizing towards its final form that
we start documenting and testing it.
# Possibly controversial issues
* bare or not bare: this series works towards a command with the end
goal of it being usable and used on bare repos, contrary to existing
commands like `git rebase` and `git cherry-pick`, but the tests
currently don't check that, and in case of conflicts it won't
currently work on bare repos. One reason for that is that existing
tests in t6429 should continue to work, and one of these tests
requires some output in case of conflict. And it's nice for users to
get some help in case of conflict. It's also nice for users if
commands that should work on both bare and non bare repos work well
on non bare repos first as they are less likely to use them on bare
repos. So let's have a command that works well on non-bare repos
first, even if its end goal is to work fine on bare repos too. We
plan to improve things for bare repos soon after this first patch
series graduates.
* exit status: a successful, non-conflicted replay exits with code
0. When the replay has conflicts, the exit status is 1. If the
replay is not able to complete (or start) due to some kind of error,
the exit status is something other than 0 or 1. It has been
suggested in an internal review that conflicts might want to get a
more specific error code as an error code of 1 might be quite easy
to return by accident. It doesn't seem to me from their docs (which
might want to be improved, I didn't look at the code) that other
commands like `git merge` and `git rebase` exit with a special error
code in case of conflict.
I don't think we've ever had a special "conflict" error code but it
would be useful for scripts if replay had one. Does replay return a
different exit code for "merge conflicts" and "cannot merge because it
would overwrite an untracked file"? Does it have an exit code for "the
commit becomes empty" or are those patches unconditionally dropped?
* to guess or not to guess: commit 12/14 introduces the
guess_new_base() function which tries to find a base to rebase onto
when the --onto option is not provided, making this option actually
optional instead of mandatory. Given that it's an heuristic and the
command end goal is to be used on server side, we might want to
introduce this as an iterative improvement later. I still think it's
interesting to have it in for now though, as it shows that --onto
and --advance (which is introduced in the following commit) should
indeed be options. If --onto was always mandatory in the series,
people could argue that it shouldn't be an option and its argument
should always be the first (unconditional) argument of the command.
I think it comes down to "what's the aim of this series?" is it focused
on the bare repository server use case or is it trying to add a general
purpose cli tool.
* make worktree and index changes optional: commit 10/14 stops
updating the index and worktree, but it might be better especially
for cli users to make that optional. The issue is that this would
make the command more complex while we are developing a number of
important features. It seems to me that this should rather be done
in an iterative improvement after the important features have
landed.
I'm confused by this as patch 9 seems to start updating the index and
worktree when there are conflicts but patch 10 stops updating the index
and worktree if the replay is successful.
* when and where to add tests and docs: although t6429 has tests that
are changed to use the new command instead of the fast-rebase
test-tool command as soon as the former is introduced, there is no
specific test script and no doc for the new command until commit
11/14 when standard revision ranges are used. This is done to avoid
churn in tests and docs while the final form of the command hasn't
crystalized enough. Adding tests and doc at this point makes this
commit quite big and possibly more difficult to review than if they
were in separate commits though. On the other hand when tests and
docs are added in specific commits some reviewers say it would be
better to introduce them when the related changes are made.
* --advance and --contained: these two advanced options might not
belong to this first series and could perhaps be added in a followup
series in separate commits. On the other hand the code for
--contained seems involved with the code of --advance and it's nice
to see soon that git replay can indeed do cherry-picking and rebase
many refs at once, and this way fullfil these parts of its promise.
Once I understood what these options did the names made sense, but I
could not tell from the names what they were going to do. For me
"--cherry-pick" and "--update-refs" would have been clearer. It might be
worth splitting this patch so the individual options are added separately.
* replaying diverging branches: 14/14 the last patch in the series,
which allow replaying diverging branches, can be seen as a
fundamental fix or alternatively as adding an interesting
feature. So it's debatable if it should be in its own patch along
with its own tests as in this series, or if it should be merged into
a previous patch and which one.
It might make sense to add this in the same patch as you add --contained.
* only 2 patches: this patch series can be seen from a high level
point of view as 1) introducing the new `git replay` command, and 2)
using `git replay` to replace, and get rid of, the fast-rebase
test-tool command. The fact that not much of the original
fast-rebase code and interface is left would agree with that point
of view. On the other hand, fast-rebase can also be seen as a first
iteration towards `git replay`. So it can also make sense to see how
`git replay` evolved from it.
Starting with fast-rebase means one has to checkout the first patch to
see what code we're adding to replay.c and to make sense of the later
patches that remove code. It would be interesting to compare this series
to one that started from scratch but I guess that would be quite a bit
of work.
Thanks for working on this, I'm interested to see where it goes
Best Wishes
Phillip
Elijah Newren (14):
replay: introduce new builtin
replay: start using parse_options API
replay: die() instead of failing assert()
replay: introduce pick_regular_commit()
replay: don't simplify history
replay: add an important FIXME comment about gpg signing
replay: remove progress and info output
replay: remove HEAD related sanity check
replay: very coarse worktree updating
replay: make it a minimal server side command
replay: use standard revision ranges
replay: introduce guess_new_base()
replay: add different modes
replay: stop assuming replayed branches do not diverge
.gitignore | 1 +
Documentation/git-replay.txt | 130 +++++++
Makefile | 2 +-
builtin.h | 1 +
builtin/replay.c | 419 +++++++++++++++++++++++
command-list.txt | 1 +
git.c | 1 +
t/helper/test-fast-rebase.c | 233 -------------
t/helper/test-tool.c | 1 -
t/helper/test-tool.h | 1 -
t/t3650-replay-basics.sh | 160 +++++++++
t/t6429-merge-sequence-rename-caching.sh | 43 ++-
12 files changed, 739 insertions(+), 254 deletions(-)
create mode 100644 Documentation/git-replay.txt
create mode 100644 builtin/replay.c
delete mode 100644 t/helper/test-fast-rebase.c
create mode 100755 t/t3650-replay-basics.sh
