300 forks of

git/torvalds/linux.git

Each linux.git is about 1.8GB

Forks are efficient... sorta

• git clone --local
  • uses hardlinks
  • safe to use everywhere
  • saves space until next repack
• git clone --shared
  • sets up objects/info/alternates
  • can result in repo corruption
  • saves lots of space

git.kernel.org is 30GB

without alternates, 400GB

Using alternates carefully

• Avoid object cleanups in the "mother" repository
  • "git repack -Ad" leaves loose objects intact
  • Do not run "git gc", just "git repack"
• Object cleanups in "daughter" repositories are OK
  • "git gc" is okay and encouraged
  • "git repack -adl" to save space
• Grandchildren are a recipe for disaster

Grokmirror

replicating git repos sanely

(it can be done!)

Grokmirror highlights

• Works via git hooks
• Creates a manifest file that replicas can pull
• Updates only repos that changed
  • Runs "git remote update" for changes
  • Runs "git clone" for new repos
  • Prunes junk
• Parallelizing and very efficient
• Will even fsck and repack for you

github.com/mricon/grokmirror

Still adding features

• Automatically recognize nearly-identical repos
  • And set up alternates when warranted
• Recognize when running "git gc" is safe
• Add "--dissociate" clone support (new in git 2.3)
• Support both python 2.7 and 3.5
• Other features upon request

git-daemon tweaks

Because cloning linux.git eats up 1GB RAM

Hardware

• RAM, lots and lots of RAM
  • Git-daemon will eat all of it, and then some
• Use haproxy in the front to combat abuse
  • Don't use http caches, as they will likely break repos after gc/repack
• Use fast disks with good seek times
  • Active repos will quickly create lots of loose objects all over the disk
• Repacking (deltas and compression) eats processing power

Useful repack flags

• -b --pack-kept-objects are your best friends
  • creates a bitmap index
  • cuts down on "counting objects" time
  • Git 2.0+ only
• -f in limited cases (we don't use it)
• Will not repack refs (tags, etc)
  • Use a separate "git pack-refs" command

Repo and bundles

Android-specific

Repo

• A tool to keep track of hundreds of repos, using a manfiest (that is a git repo itself)
• Not for the same purposes as grokmirror (dev-oriented, not mirroring-oriented)
• Parallelizes clones and updates
  • (resulting in abuse of your servers)

Repo bundles

• Neat feature that can use lookaside "git bundle" packs to offload most traffic to http cache
  • Can be placed on akamai or other accelerators
• Only works with http:// clone URLs
• Can save you tons of expensive bandwidth if your project uses on repo at all
• See "git-bundle(1)" and Google's "repo" tool