Git Scaling: What factors most affect Git performance for a large repo?

I posted this to comp.version-control.git.user and didn't get any response. I think the question is plumbing-related enough that I can ask it here.

I'm evaluating the feasibility of moving my team from SVN to git. We have a very large repo. [1] We will have a central repo using GitLab (or similar) that everybody works with. Forks, code sharing, pull requests etc. will be done through this central server.

By 'performance', I guess I mean speed of day to day operations for devs.

   * (Obviously, trivially, a (non-local) clone will be slow with a large repo.)
   * Will a few simultaneous clones from the central server also slow down
     other concurrent operations for other users?
   * Will 'git pull' be slow?
   * 'git push'?
   * 'git commit'? (It is listed as slow in reference [3].)
   * 'git stautus'? (Slow again in reference 3 though I don't see it.)
   * Some operations might not seem to be day-to-day but if they are called
     frequently by the web front-end to GitLab/Stash/GitHub etc then
     they can become bottlenecks. (e.g. 'git branch --contains' seems terribly
     adversely affected by large numbers of branches.)
   * Others?

Assuming I can put lots of resources into a central server with lots of CPU,
RAM, fast SSD, fast networking, what aspects of the repo are most likely to
affect devs' experience?
   * Number of commits
   * Sheer disk space occupied by the repo
   * Number of tags.
   * Number of branches.
   * Binary objects in the repo that cause it to bloat in size [1]
   * Other factors?

Of the various HW items listed above --CPU speed, number of cores, RAM, SSD, networking-- which is most critical here?

(Stash recommends 1.5 x repo_size x number of concurrent clones of available RAM. I assume that is good advice in general.)

Assume ridiculous numbers. Let me exaggerate: say 1 million commits, 15 GB repo, 50k tags, 1,000 branches. (Due to historical code fixups, another 5,000 "fix-up branches" which are just one little dangling commit required to change the code a little bit between a commit a tag that was not quite made from it.)

While there's lots of information online, much of it is old [3] and with git
constantly evolving I don't know how valid it still is. Then there's anecdotal
evidence that is of questionable value.[2]
    Are many/all of the issues Facebook identified [3] resolved? (Yes, I
understand Facebook went with Mercurial. But I imagine the git team nevertheless
took their analysis to heart.)

Thanks, Steve

[1] (Yes, I'm investigating ways to make our repo not so large etc. That's
    beyond the scope of the discussion I'd like to have with this
    question. Thanks.)
[2] The large amounts of anecdotal evidence relate to the "why don't you try it
    yourself?" response to my question. I will I I have to but setting up a
    properly methodical study is time consuming and difficult --I don't want to
    produce poor anecdotal numbers that don't really hold up-- and if somebody's
    already done the work, then I should leverage it.
[3] http://thread.gmane.org/gmane.comp.version-control.git/189776

On Thu, Feb 19, 2015 at 1:26 PM, Stephen Morton <stephen.c.morton@gmail.com> wrote:

What do you mean by large?
* lots of files
* large files
* or even large binary files (bad to diff/merge)
* long history (i.e. lots of small changes)
* impactful history (changes which rewrite nearly everything from scratch)

For reference, the linux
* has 48414 files, in 3128 directories
* the largest file is 1.1M, the whole repo is 600M
* no really large binary files
* more than 500051 changes/commits including merges
* started in 2004 (when git was invented essentially)
* the .git folder is 1.4G compared to the 600M files,
   indicating it may have been rewritting 3 times (well this
   metric is bogus, there is lots of compression
   going on in .git)

and linux seems to be doing ok with git.

So as long as you cannot pinpoint your question on what you are exactly concerned about, there will be no helpful answer I guess.

linux is by no means a really large project, there are other projects way larger than that (I am thinking about the KDE project for example) and they do fine as well.

Thanks, Stefan

On Thu, Feb 19, 2015 at 5:21 PM, Stefan Beller <sbeller@google.com> wrote:

Hi Stefan,

I think I addressed most of this in my original post with the paragraph

 "Assume ridiculous numbers. Let me exaggerate: say 1 million commits,
15 GB repo,
  50k tags, 1,000 branches. (Due to historical code fixups, another
5,000 "fix-up
  branches" which are just one little dangling commit required to
change the code
  a little bit between a commit and a tag that was not quite made from it.)"

To that I'd add 25k files, no major rewrites, no huge binary files, but lots of a few MB binary files with many revisions.

But even without details of my specific concerns, I thought that perhaps the git developers know what limits git's performance even if large projects like the kernel are not hitting these limits.

Steve

On Thu, Feb 19, 2015 at 3:06 PM, Stephen Morton <stephen.c.morton@gmail.com> wrote:

I did not realize you gave numbers below, as I started answering after reading the first paragraphs. Sorry about that.

I think lots of files organized in a hierarchical fashion ranging in the small MB range is not a huge deal. Also history is a non issue

The problem arises with having lots of branches. "640 git branches ought to be enough for everybody -- Linus" (just kidding) Git doesn't really scale efficiently with lots of branches (second hand information except for fetch/pull where I did some patches on another topic recently).

Thanks, Stefan

On Thu, Feb 19, 2015 at 10:26 PM, Stephen Morton <stephen.c.morton@gmail.com> wrote:

Anecdotally I work on a repo at work (where I'm mostly "the Git guy") that's:

 * Around 500k commits
 * Around 100k tags
 * Around 5k branches
 * Around 500 commits/day, almost entirely to the same branch
 * 1.5 GB .git checkout.
 * Mostly text source, but some binaries (we're trying to cut down[1] on those)

The main scaling issues we have with Git are:

 * "git pull" takes around 10 seconds or so
 * Operations like "git status" are much slower because they scale
with the size of the work tree
 * Similarly "git rebase" takes a much longer time for each applied
commit, I think because it does the equivalent of "git status" for
every applied commit. Each commit applied takes around 1-2 seconds.
 * We have a lot of contention on pushes because we're mostly pushing
to one branch.
 * History spelunking (e.g. git log --reverse -p -G<str>) is taking
longer by the day

The obvious reason for why "git pull" is slow is because git-upload-pack spews the complete set of refs at you each time. The output from that command is around 10MB in size for us now. It takes around 300 ms to run that locally from hot cache, a bit more to send it over the network.

But actually most of "git fetch" is spent in the reachability check subsequently done by "git-rev-list" which takes several seconds. I haven't looked into it but there's got to be room for optimization there, surely it only has to do reachability checks for new refs, or could run in some "I trust this remote not to send me corrupt data" completely mode (which would make sense within a company where you can trust your main Git box).

The "git status" operations could be made faster by having something like watchman, there's been some effort on getting that done in Git, but I haven't tried it. This seems to have been the main focus of Facebook's Mercurial optimization effort.

Some of this you can "solve" mostly by doing e.g. "git status -uno", having support for such unsafe operations (e.g. teaching rebase and pals to use it) would be nice at the cost of some safety, but having something that feeds of inotify would be even better.

It takes around 3 minutes to reclone our repo, we really don't care (we rarely re-clone). But I thought I'd mention it because for some reason this is important to Facebook and along with inotify were the two major things they focused on.

As far as I know every day Git operations don't scale all that badly with a huge history. They will a bit since everything will live in the same pack file, and this'll become especially noticable when your packfiles are being evicted out of the page cache.

However operations like "git repack" seem to be quite bad at handling these sort of repos. It already takes us many GB of RAM to repack ours. I'd hate to do the same if it was 10x as big.

Overall I'd say Git would work for you for a repo like that, I'd certainly still take it over SVN any day. The main thing you might want to try out is partitioning out any binary assets you may have. Usually that's much easier than splitting up the source tree itself.

I haven't yet done this, but I was planning on writing something to start archiving our tags (mostly created by [2]) along with aggressively deleting branches in the repo. I haven't benchmarked that but I think that'll make the "pull" operations much faster, which in turn will make the push contention (lots of lemmings pushing to the same ref) better since the pull && push window is reduced.

I do ask myself what we're going to do if we just keep growing and all the numbers I cited will be multiplied by 10-50x. With the current Git limitations on the git implementation I think we'd need to split the repo. The main reason we don't do so is because we like being able to atomically change a library and its users.

However there's nothing in the basic Git file repository format that inherently limits Git from being smarter about large repos, it just seems to be hard to implement with the way the current client is structured. In particular nothing would stop a Git client from:

 * Partially cloning a history but still being able to push upstream.
You could just get a partial commit/tree/blob graph and fetch the rest
on-demand as needed.
 * Scaling up to multi TB or PB repos. We'd just have to treat blobs
as something fetched on-demand, sort of like what git-annex does, but
built-in. We'd also have to be less stupid about how we pack big blobs
(or not at all)
 * Being able to partially clone a Git working tree. You could ask the
server for the last N commit objects and what you need for some
subdirectory in the repo. Then when you commit you ask the server what
the other top-level tree objects you need to make a commit are.
 * Nothing in the Git format itself actually requires filesystem
access. Not having to deal with external things modifying the tree
would be another approach to what the inotify effort is trying to
solve.

Of course changes like that will require a major rehaul of the current codebase, or another implementation. Some of those require much more active client/server interaction than what we have now, but they are possible, which gives me some hope for the future.

Finally, I'd like to mention that if someone here on-list is interested in doing work on these scalability topics in Git we'd be open to funding that effort on some contract basis. Obviously the details would have to be worked out blah blah blah, but that came up the last time we had discussions about this internally. Myself and a bunch of other people at work /could/ work on this ourselves, but we're busy with other stuff and would much prefer just to pay someone to fix them.

1. https://github.com/avar/pre-receive-reject-binaries
2. https://github.com/git-deploy/git-deploy

On Fri, Feb 20, 2015 at 6:29 AM, Ævar Arnfjörð Bjarmason <avarab@gmail.com> wrote:

Would be nice if you could make an anonymized version of this repo public. Working on a "real" large repo is better than an artificial one.

I wonder if reachability bitmap could help here..

No, it's not just about trusting the server side, it's about catching data corruption on the wire as well. We have a trick to avoid reachability check in clone case, which is much more expensive than a fetch. Maybe we could do something further to help the fetch case _if_ reachability bitmaps don't help.

On Fri, Feb 20, 2015 at 1:04 AM, Duy Nguyen <pclouds@gmail.com> wrote:

Yeah, I'll try to do that.

I could have sworn I had that enabled already but evidently not. I did test it and it cut down on clone times a bit. Now our daily repacking is:

        git --git-dir={} gc &&
        git --git-dir={} pack-refs --all --prune &&
        git --git-dir={} repack -Ad --window=250 --depth=100
--write-bitmap-index --pack-kept-objects &&

It's not clear to me from the documentation whether this should just be enabled on the server, or the clients too. In any case I've enabled it on both.

Even then with it enabled on both a "git pull" that pulls down just one commit on one branch is 13s. Trace attached at the end of the mail.

Still, if that's indeed a big bottleneck what's the worst-case scenario here? That the local repository gets hosed? The server will still recursively validate the objects it gets sent, right?

I wonder if a better trade-off in that case would be to skip this in some situations and instead put something like "git fsck" in a cronjob.

Here's a "git pull" trace mentioned above:

$ time GIT_TRACE=1 git pull
13:06:13.603781 git.c:555               trace: exec: 'git-pull'
13:06:13.603936 run-command.c:351       trace: run_command: 'git-pull'
13:06:13.620615 git.c:349               trace: built-in: git
'rev-parse' '--git-dir'
13:06:13.631602 git.c:349               trace: built-in: git
'rev-parse' '--is-bare-repository'
13:06:13.636103 git.c:349               trace: built-in: git
'rev-parse' '--show-toplevel'
13:06:13.641491 git.c:349               trace: built-in: git 'ls-files' '-u'
13:06:13.719923 git.c:349               trace: built-in: git
'symbolic-ref' '-q' 'HEAD'
13:06:13.728085 git.c:349               trace: built-in: git 'config'
'branch.trunk.rebase'
13:06:13.738160 git.c:349               trace: built-in: git 'config' 'pull.ff'
13:06:13.743286 git.c:349               trace: built-in: git
'rev-parse' '-q' '--verify' 'HEAD'
13:06:13.972091 git.c:349               trace: built-in: git
'rev-parse' '--verify' 'HEAD'
13:06:14.149420 git.c:349               trace: built-in: git
'update-index' '-q' '--ignore-submodules' '--refresh'
13:06:14.294098 git.c:349               trace: built-in: git
'diff-files' '--quiet' '--ignore-submodules'
13:06:14.467711 git.c:349               trace: built-in: git
'diff-index' '--cached' '--quiet' '--ignore-submodules' 'HEAD' '--'
13:06:14.683419 git.c:349               trace: built-in: git
'rev-parse' '-q' '--git-dir'
13:06:15.189707 git.c:349               trace: built-in: git
'rev-parse' '-q' '--verify' 'HEAD'
13:06:15.335948 git.c:349               trace: built-in: git 'fetch'
'--update-head-ok'
13:06:15.691303 run-command.c:351       trace: run_command: 'ssh'
'git.example.com' 'git-upload-pack '\''/gitrepos/core.git'\'''
13:06:17.095662 run-command.c:351       trace: run_command: 'rev-list'
'--objects' '--stdin' '--not' '--all' '--quiet'
remote: Counting objects: 6, done.
remote: Compressing objects: 100% (6/6), done.
3:06:20.426346 run-command.c:351       trace: run_command:
'unpack-objects' '--pack_header=2,6'
13:06:20.431806 exec_cmd.c:130          trace: exec: 'git'
'unpack-objects' '--pack_header=2,6'
13:06:20.437343 git.c:349               trace: built-in: git
'unpack-objects' '--pack_header=2,6'
remote: Total 6 (delta 0), reused 0 (delta 0)
Unpacking objects: 100% (6/6), done.
13:06:20.444196 run-command.c:351       trace: run_command: 'rev-list'
'--objects' '--stdin' '--not' '--all'
13:06:20.447135 exec_cmd.c:130          trace: exec: 'git' 'rev-list'
'--objects' '--stdin' '--not' '--all'
13:06:20.451283 git.c:349               trace: built-in: git
'rev-list' '--objects' '--stdin' '--not' '--all'
From ssh://git.example.com/gitrepos/core
   02d33d2..41e72c4  core      -> origin/core
13:06:22.559609 run-command.c:351       trace: run_command: 'gc' '--auto'
13:06:22.562176 exec_cmd.c:130          trace: exec: 'git' 'gc' '--auto'
13:06:22.565661 git.c:349               trace: built-in: git 'gc' '--auto'
13:06:22.594980 git.c:349               trace: built-in: git
'rev-parse' '-q' '--verify' 'HEAD'
13:06:22.845728 git.c:349               trace: built-in: git
'show-branch' '--merge-base' 'refs/heads/core'
'41e72c42addc5075e8009a3eebe914fa0ce98b27'
'02d33d2be7f8601c3502fdd89b0946447d7cdf15'
13:06:23.087586 git.c:349               trace: built-in: git 'fmt-merge-msg'
13:06:23.341451 git.c:349               trace: built-in: git
'rev-parse' '--parseopt' '--stuck-long' '--' '--onto'
'41e72c42addc5075e8009a3eebe914fa0ce98b27'
'41e72c42addc5075e8009a3eebe914fa0ce98b27'
13:06:23.350513 git.c:349               trace: built-in: git
'rev-parse' '--git-dir'
13:06:23.362011 git.c:349               trace: built-in: git
'rev-parse' '--is-bare-repository'
13:06:23.365282 git.c:349               trace: built-in: git
'rev-parse' '--show-toplevel'
13:06:23.372589 git.c:349               trace: built-in: git 'config'
'--bool' 'rebase.stat'
13:06:23.377056 git.c:349               trace: built-in: git 'config'
'--bool' 'rebase.autostash'
13:06:23.382102 git.c:349               trace: built-in: git 'config'
'--bool' 'rebase.autosquash'
13:06:23.389458 git.c:349               trace: built-in: git
'rev-parse' '--verify' '41e72c42addc5075e8009a3eebe914fa0ce98b27^0'
13:06:23.608894 git.c:349               trace: built-in: git
'rev-parse' '--verify' '41e72c42addc5075e8009a3eebe914fa0ce98b27^0'
13:06:23.894026 git.c:349               trace: built-in: git
'symbolic-ref' '-q' 'HEAD'
13:06:23.898918 git.c:349               trace: built-in: git
'rev-parse' '--verify' 'HEAD'
13:06:24.102269 git.c:349               trace: built-in: git
'rev-parse' '--verify' 'HEAD'
13:06:24.338636 git.c:349               trace: built-in: git
'update-index' '-q' '--ignore-submodules' '--refresh'
13:06:24.539912 git.c:349               trace: built-in: git
'diff-files' '--quiet' '--ignore-submodules'
13:06:24.729362 git.c:349               trace: built-in: git
'diff-index' '--cached' '--quiet' '--ignore-submodules' 'HEAD' '--'
13:06:24.938533 git.c:349               trace: built-in: git
'merge-base' '41e72c42addc5075e8009a3eebe914fa0ce98b27'
'02d33d2be7f8601c3502fdd89b0946447d7cdf15'
13:06:25.197791 git.c:349               trace: built-in: git 'diff'
'--stat' '--summary' '02d33d2be7f8601c3502fdd89b0946447d7cdf15'
'41e72c42addc5075e8009a3eebe914fa0ce98b27'
[details on updated files]
13:06:25.488275 git.c:349               trace: built-in: git
'checkout' '-q' '41e72c42addc5075e8009a3eebe914fa0ce98b27^0'
13:06:26.467413 git.c:349               trace: built-in: git
'update-ref' 'ORIG_HEAD' '02d33d2be7f8601c3502fdd89b0946447d7cdf15'
Fast-forwarded trunk to 41e72c42addc5075e8009a3eebe914fa0ce98b27.
13:06:26.716256 git.c:349               trace: built-in: git 'rev-parse' 'HEAD'
13:06:26.958595 git.c:349               trace: built-in: git
'update-ref' '-m' 'rebase finished: refs/heads/core onto
41e72c42addc5075e8009a3eebe914fa0ce98b27' 'refs/heads/core'
'41e72c42addc5075e8009a3eebe914fa0ce98b27'
'02d33d2be7f8601c3502fdd89b0946447d7cdf15'
13:06:27.205320 git.c:349               trace: built-in: git
'symbolic-ref' '-m' 'rebase finished: returning to refs/heads/core'
'HEAD' 'refs/heads/core'
13:06:27.208748 git.c:349               trace: built-in: git 'gc' '--auto'

On Fri, Feb 20, 2015 at 1:09 PM, Ævar Arnfjörð Bjarmason <avarab@gmail.com> wrote:

I forgot to include that this took:

real 0m13.630s user 0m10.739s sys 0m4.064s

on my local laptop with a ssd + hot cache it was:

real 0m7.513s user 0m3.796s sys 0m0.624s

So some of that we could speed up with faster systems, but we still have quite a bit of Git overhead.

Even with the hot cache on the ssd I get on this repo:

$ time (git log -1 >/dev/null)

real 0m0.938s user 0m0.916s sys 0m0.020s

v.s. the same on linux.git:

$ time (git log -1 >/dev/null)

real 0m0.016s user 0m0.008s sys 0m0.004s

Which I suspect is a function of the high ref count, but it could be something else...

On Fri, Feb 20, 2015 at 1:09 PM, Ævar Arnfjörð Bjarmason <avarab@gmail.com> wrote:

tl;dr: After some more testing it turns out the performance issues we have are almost entirely due to the number of refs. Some of these I knew about and were obvious (e..g. git pull), but some aren't so obvious (why does "git log" without "--all" slow down as a function of the overall number of refs?).

Rather than getting an anonymized version of the repo we have, a simpler isolated test case is just doing this on linux.git:

    $ git rev-list --all | perl -ne 'my $cnt; while (<>) {
s<([a-f0-9]+)><git tag -a -m"Test" TAG $1>gm; next unless int rand 10
== 1; $cnt++; s/TAG/tagnr-$cnt/; print }'  | sh -x

That'll create a tag for every 10th commit or so, which is around 50k tags for linux.git.

I actually ran this a few times while testing it, so this is a before and after on a hot cache of linux.git with 406 tags v.s. ~140k. I ran the gc + repack + bitmaps for both repos noted in an earlier reply of mine, and took the fastest run out of 3:

    $ time (git log master -100 >/dev/null)
    Before: real    0m0.021s
    After: real    0m2.929s
    $ time (git status >/dev/null)
    # Around 150ms, no noticeable difference
    $ time git fetch
    # I'm fetching from git@github.com:torvalds/linux.git here, the
    # cache is hot but upstream has *no* changes
    Before: real    0m1.826s
    After: real    0m8.458s

Details on why "git fetch" is slow in this situation:

    $ time GIT_TRACE=1 git fetch
    15:15:00.435420 git.c:349               trace: built-in: git 'fetch'
    15:15:00.654428 run-command.c:341       trace: run_command: 'ssh'
'git@github.com' 'git-upload-pack '\''torvalds/linux.git'\'''
    15:15:02.426121 run-command.c:341       trace: run_command:
'rev-list' '--objects' '--stdin' '--not' '--all' '--quiet'
    15:15:05.507327 run-command.c:341       trace: run_command:
'rev-list' '--objects' '--stdin' '--not' '--all'
    15:15:05.508329 exec_cmd.c:134          trace: exec: 'git'
'rev-list' '--objects' '--stdin' '--not' '--all'
    15:15:05.510490 git.c:349               trace: built-in: git
'rev-list' '--objects' '--stdin' '--not' '--all'
    15:15:08.874116 run-command.c:341       trace: run_command: 'gc' '--auto'
    15:15:08.879570 exec_cmd.c:134          trace: exec: 'git' 'gc' '--auto'
    15:15:08.882495 git.c:349               trace: built-in: git 'gc' '--auto'
    real    0m8.458s
    user    0m6.548s
    sys     0m0.204s

Even things you'd expect to not be impacted are, like a reverse log search on the master branch:

    $ time (git log --reverse -p --grep=arm64 origin/master >/dev/null)
    Before: real    0m4.473s
    After: real    0m6.194s

Or doing 10 commits and rebasing on the upstream:

    $ time (git checkout origin/master~ && for i in {1..10}; do echo
$i > file && git add file && git commit -m"moo" $file; done && git
rebase origin/master)
    Before: real    0m6.798s
    After: real    0m12.340s

The remaining slowdown comes from the size of the tree, which we can deal with by either reducing it in size (we have some copied JS libraries and whatnot) or trying the inotify-powered git-status.

In our case there's no good reason for why we have this many refs in the repository everyone uses. We basically just have a bunch of dated rollout tags that have been accumulating for years, and a bunch of mostly unused branches people just haven't cleaned up.

So I'm going to:

 1. Write a hook that rejects tags that aren't new (i.e. forbid
re-pushes of old tags)
 2. Create an archive repository that contains all the old tags (i.e.
just run "git fetch" on the main one from cron)
 3. Run a script to regularly delete tags from the main repo
 4. Run the same script on the clients that clone the repo

The branches are slightly harder, deleting those that are fully merged into the same branch is easy, deleting those whose contents 100% matches patch-id's already in the main branch is another thing we can do, and just clean up branches unconditionally after they've reached a certain age (they'll still be archived).

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

Do you force --decorate with some config? Or do you see similar performance difference with "git rev-parse master", too?

This is understandable, as it will not look at any ref other than HEAD.

On Fri, Feb 20, 2015 at 10:04 PM, Junio C Hamano <gitster@pobox.com> wrote:

Yes, I had log.decorate=short set in my config. With --no-decorate:

    $ time (git log --no-decorate -100 >/dev/null)
    # Before: real    0m0.010s
    # After: real    0m0.065s

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:

There you have the answer to your earlier question, then, which was:

On 20.02.2015 15:25, Ævar Arnfjörð Bjarmason wrote:

Interesting. We currently have similar performance issues when pushing to a Git repo hosted on Gerrit. The only difference of our repo compared to others on the same server that do not have any performance issues is the large number of refs (about 40k) in refs/changes.

I still wonder why that would slow down a push as we have not configured our clients to fetch / push refs/changes.

On 02/20/2015 03:25 PM, Ævar Arnfjörð Bjarmason wrote:

I'm assuming that you pack your references periodically. (If not, you should, because reading lots of loose references is very expensive for the commands that need to iterate over all references!)

On the other hand, packed refs also have a downside, namely that whenever even a single packed reference has to be read, the whole packed-refs file has to be read and parsed. One way that this can bite you, even with innocuous-seeming commands, is if you haven't disabled the use of replace references (i.e., using "git --no-replace-objects <CMD>" or GIT_NO_REPLACE_OBJECTS). In that case, almost any Git command has to read the "refs/replace/*" namespace, which, in turn, forces the whole packed-refs file to be read and parsed. This can take a significant amount of time if you have a very large number of references.

So try your experiments with replace references disabled. If that helps, consider disabling them on your server if you don't need them.

Michael

On Tue, Feb 24, 2015 at 1:44 PM, Michael Haggerty <mhagger@alum.mit.edu> wrote:

Yes, as mentioned in another reply of mine, like this:

    git --git-dir={} gc &&
    git --git-dir={} pack-refs --all --prune &&
    git --git-dir={} repack -Ad --window=250 --depth=100
--write-bitmap-index --pack-kept-objects &&

Interesting. I tried the rough benchmarks I posted above with GIT_NO_REPLACE_OBJECTS=1 and couldn't see any differences, although as mentioned in another reply --no-decorate had a big effect on git-log.

On Fri, Feb 20, 2015 at 7:09 PM, Ævar Arnfjörð Bjarmason <avarab@gmail.com> wrote:

Pack bitmaps matter most on the server side. What I was not sure was whether it helped the client side as well because you do rev-list on the client side for reachability test. But thinking again, I don't think enabling pack bitmaps on the client side helps much. The "--not --all" part in rev-list basically just traverses commits, not trees and objects (where pack bitmaps shine). The big problem here is "--all" which will go examine all refs. So big ref number problem again..

The server is under pressure to pack and send data fast so it does not validate as heavily as the client. When deltas are reused, only crc32 is verified. When deltas are generated, the server must unpack some objects for deltification, but I don't think it rehashes the content to see if it produces the same SHA-1. Single bit flips could go unnoticed..

Either that or be optimistic, accept the pack (i.e. git-fetch returns quickly) and validate it in the background. If the pack is indeed good, you don't have to wait until validation is done. If the pack is bad, you would know after a minute or two, hopefully you can still recover from that point.

On Fri, Feb 20, 2015 at 4:26 AM, Stephen Morton <stephen.c.morton@gmail.com> wrote:

There are no locks in server when cloning, so in theory cloning does not affect other operations. Cloning can use lots of memory though (and a lot of cpu unless you turn on reachability bitmap feature, which you should).

If we exclude the server side, the size of your tree is the main factor, but your 25k files should be fine (linux has 48k files).

This one is not affected by how deep your repo's history is, or how wide your tree is, so should be quick..

Ah the number of refs may affect both git-push and git-pull. I think Stefan knows better than I in this area.

(also git-add)

Again, the size of your tree. I'm trying to address problems in [3], but at your repo's size, I don't think you need to worry about it.

git-blame could be slow when a file is modified a lot.

On Fri, 2015-02-20 at 06:38 +0700, Duy Nguyen wrote:

I can tell you that this is a bit of a problem for us at Twitter. We have over 100k refs, which adds ~20MiB of downstream traffic to every push.

I added a hack to improve this locally inside Twitter: The client sends a bloom filter of shas that it believes that the server knows about; the server sends only the sha of master and any refs that are not in the bloom filter. The client uses its local version of the servers' refs as if they had just been sent. This means that some packs will be suboptimal, due to false positives in the bloom filter leading some new refs to not be sent. Also, if there were a repack between the pull and the push, some refs might have been deleted on the server; we repack rarely enough and pull frequently enough that this is hopefully not an issue.

We're still testing to see if this works. But due to the number of assumptions it makes, it's probably not that great an idea for general use.

There are probably more complex schemes to compute minimal (or small-enough) packs; in particular, if the patch is just a few megs off of master, it's better to just send the whole pack. That doesn't work for us because we've got a log-based replication scheme that the pack appends to, and we don't want the log to get too big; we want more-minimal packs than that. But it might work for others.

David Turner <dturner@twopensource.com> writes:

Interesting.

Care to extend the discussion to improve the protocol exchange, which starts at $gmane/263932 [*1*], where I list the known issues around the current protocol (and a possible way to correct them in footnotes)?

[Footnote]

*1* http://thread.gmane.org/gmane.comp.version-control.git/263898/focus=263932

On Fri, 2015-02-20 at 12:59 -0800, Junio C Hamano wrote:

At Twitter, we changed to an entirely different clone strategy for our largest repo: instead of using git clone, we use bittorrent (on a tarball of the repo). For git pull, we maintain a journal of all pushes ever made to the server (data and ref updates); each client keeps track of their location in that journal. So now pull does not require any computation on the server; the client just requests the segment of the journal that they don't have. Then the client replays the journal. This scheme isn't perfect: clients end up with data about even transitory and long-dead branches, and there is presently no way to redact data (although that would be possible to add). And of course shallow and sparse clones are impossible. But it works quite well for Twitter's needs. As I understand it, the hope is to implement redaction and then submit patches upstream.

I say "we", but I personally did not do any of the above work. Because I haven't looked into most of these issues personally, I'm reluctant to say too much on protocol improvements. I would want to better understand the constraints. I do think there is value in having a diversity of possible protocols to handle different use cases. As repositories grow, traditional full-repo clones become less viable. Network transfer and client-side performance both suffer. In a repo the size of (say) WebKit, the traditional model works. In a repo the size of Facebook's monorepo, it starts to break down. So Facebook does entirely shallow clones (using hg, but the problems are similar in git). Commands like log and blame instead call out to a server to gather history data. At Google, whose repo is I think two or three orders of magnitude larger than WebKit, all local copies are both shallow and sparse; there is also support for "sparse commits" -- so that a commit that affects (say) ten thousand files across the entire tree can be kept to a reasonable size.

<end digression>

Twitter's journal scheme explains why I implemented bloom filter pushes -- the number of refs does not significantly affect pull performance, but pushes still go through the normal git machinery, so we wanted an optimization to reduce latency there.

On Fri, Feb 20, 2015 at 7:42 AM, David Turner <dturner@twopensource.com> wrote:

I wonder how efficient rsync is for transferring these refs: the client generates a "file" containing all refs, the server does the same with their refs, then the client rsync their file to the server.. The changes between the server and the client files are usually small, I'm hoping rsync can take advantage of that.

On Sat, Feb 21, 2015 at 11:01 AM, Duy Nguyen <pclouds@gmail.com> wrote:

Some numbers without any actual coding. After the initial clone, we store the server's refs in a file called base-file at client. At the next push or pull, the server saves its refs in 'new-file'. Using rsync to avoid initial ref advertisement would involve these steps (rdiff command is from librsync)

client> rdiff signature base-file signature (client sends "signature" file to server) server> rdiff delta signature new-file delta (server sends "delta" file back to client) client> rdiff patch base-file delta new-file

The exchanged files over network are "signature" and "delta". I used my git.git's packed-refs as the base-file (1416 refs, 78789 bytes) and modifies three lines to create new-file. That produced a signature file of 480 bytes and delta file of 6163 bytes. That's 7% the size of the new file. Good.

When I modified more lines in new-file (15 lines), the delta file grew to 26644 bytes ("sig" file remains the same because it only depends on base-file). Total transferred bytes were 60% the size of new-file. Less impressive. Maybe there's some tuning options for better results...

The same process could be used to transfer the whole client refs to server instead of sending lots of "have" lines. I suspect there will be more changes between client's "have" file and server ref list. If the changes spread out and cause a lot of blocks to be sent, the saving would not be as high as I wanted. I guess that's it for rsync idea.

On Thu, Feb 19, 2015 at 04:26:58PM -0500, Stephen Morton wrote:

This hasn't been a problem for us at $DAYJOB. Git doesn't lock anything on fetches, so each process is independent. We probably have about sixty developers (and maybe twenty other occasional users) that manage to interact with our Git server all day long. We also have probably twenty smoker (CI) systems pulling at two hour intervals, or, when there's nothing to do, every two minutes, plus probably fifteen to twenty build systems pulling hourly.

I assume you will provide adequate resources for your server.

The most pathological case I've seen for git push is a branch with a single commit merged into the main development branch. As of Git 2.3.0, the performance regression here is fixed.

Obviously, the speed of your network connection will affect this. Even at 30 MB/s, cloning several gigabytes of data takes time. Git tries hard to eliminate sending a lot of data, so if your developers keep reasonably up-to-date, the cost of establishing the connection will tend to dominate.

I see pull and push times that are less than 2 seconds in most cases.

These can be slow with slow disks or over remote file systems. I recommend not doing that. I've heard rumbles that disk performance is better on Unix, but I don't use Windows so I can't say.

You should keep your .gitignore files up-to-date to avoid enumerating untracked files. There's some work towards making this less of an issue.

git blame can be somewhat slow, but it's not something I use more than about once a day, so it doesn't bother me that much.

The number of files can impact performance due to the number of stat()s required.

The number of tags and branches individually is really less relevant than the total number of refs (tags, branches, remote branches, etc). Very large numbers of refs can impact performance on pushes and pulls due to the need to enumerate them all.

If you want good performance, I'd recommend the latest version of Git both client- and server-side. Newer versions of Git provide pack bitmaps, which can dramatically speed up clones and fetches, and Git 2.3.0 fixes a performance regression with large numbers of refs in non-shallow repositories.

It is totally worth it to roll your own packages of git if your vendor provides old versions.

I generally find that having a good disk cache is important with large repositories. It may be advantageous to make sure the developer machines have adequate memory. Performance is notably better on development machines (VMs) with 2 GB or 4 GB of memory instead of 1 GB.

I can't speak to the server side, as I'm not directly involved with its deployment.

I routinely work on a repo that's 1.9 GB packed, with 25k (and rapidly growing) refs. Other developers work on a repo that's 9 GB packed, with somewhat fewer refs. We don't tend to have problems with this.

Obviously, performance is better on some of our smaller repos, but it's not unacceptable on the larger ones. I generally find that the 940 KB repo with huge numbers of files performs worse than the 1.9 GB repo with somewhat fewer. If you can split your repository into multiple logical repositories, that will certainly improve performance.

If you end up having pain points, we're certainly interested in working through those. I've brought up performance problems and people are generally responsive.

This is fantastic. I really appreciate all the answers. And it's great that I think I've sparked some general discussion that could lead somewhere too.

Notes:

I'm currently using 2.1.3. I'll move to 2.3.x

I'm experimenting with git-annex to reduce repo size on disk. We'll see.

I could remove all tags older than /n/ years old in the active repo and just maintain them in the historical repo. (We have quite a lot of CI-generated tags.) It sounds like that might improve performance.

Questions:

1. Ævar : I'm a bit concerned by your statement that git rebases take
about 1-2 s per commit. Does that mean that a "git pull --rebase", if
it is picking up say 120 commits (not at all unrealistic), could
potentially take 4 minutes to complete? Or have I misinterpreted your
comment.

2. I'd not heard about bitmap indexes before this thread but it sounds
like they should help me. In limited searching I can't find much
useful documentation about them. It is also not clear to me if I have
to explicitly run "git repack --write-bitmap-indexes" or if git will
automatically detect when they're needed; first experiments seem to
indicate that I need to explicitly generate them. I assume that once
the index is there, git will just use it automatically.

Steve

On Thu, Feb 19, 2015 at 7:03 PM, brian m. carlson <sandals@crustytoothpaste.net> wrote:

Stephen Morton <stephen.c.morton@gmail.com> writes:

Ævar talked about "applied commits" during rebase. When you "git pull --rebase", you fast-forward the history you just fetched, which is almost instantaneous, and then you reapply your local history on top of it.

So, the performance depends on how long your local history is, not on how many commits you're fetching.

On Fri, Feb 20, 2015 at 11:06:44AM -0500, Stephen Morton wrote:

Pack bitmaps are a way of speeding up clones and fetches by precomputing reachability information. Practically, this means that the initial "Counting objects" phase is instantaneous for clones and much faster for fetches.

The way I've done it in the past is to set repack.writeBitmaps = true in /etc/gitconfig on the server. (I highly recommend enabling reflogs in the same place.) Then you'll need to ensure that git gc runs periodically so that bitmaps are generated.

On 20.02.2015 01:03, brian m. carlson wrote:

Do you happen now which version, if at all, of JGit and Gerrit support pack bitmaps?

Do you also know in what Git version the regression was introduced?

On Fri, Feb 20, 2015 at 11:08:55PM +0100, Sebastian Schuberth wrote:

They were originally implemented in JGit, but I don't know what version, sorry. Some googling tells me that it's probably version 3.0.

v1.8.4-rc3-8-gfbd4a70.  It was fixed in v2.2.1-65-g2dacf26.