tree eec2fdbd162e5797eb509ce0d39cdfa21a5345a8
parent efdd2f0d4c4d6b1b1090171b6428919038bd2980
author Derrick Stolee <dstolee@microsoft.com> 1601037216 +0000
committer Junio C Hamano <gitster@pobox.com> 1601056384 -0700

maintenance: add incremental-repack task

The previous change cleaned up loose objects using the
'loose-objects' that can be run safely in the background. Add a
similar job that performs similar cleanups for pack-files.

One issue with running 'git repack' is that it is designed to
repack all pack-files into a single pack-file. While this is the
most space-efficient way to store object data, it is not time or
memory efficient. This becomes extremely important if the repo is
so large that a user struggles to store two copies of the pack on
their disk.

Instead, perform an "incremental" repack by collecting a few small
pack-files into a new pack-file. The multi-pack-index facilitates
this process ever since 'git multi-pack-index expire' was added in
19575c7 (multi-pack-index: implement 'expire' subcommand,
2019-06-10) and 'git multi-pack-index repack' was added in ce1e4a1
(midx: implement midx_repack(), 2019-06-10).

The 'incremental-repack' task runs the following steps:

1. 'git multi-pack-index write' creates a multi-pack-index file if
   one did not exist, and otherwise will update the multi-pack-index
   with any new pack-files that appeared since the last write. This
   is particularly relevant with the background fetch job.

   When the multi-pack-index sees two copies of the same object, it
   stores the offset data into the newer pack-file. This means that
   some old pack-files could become "unreferenced" which I will use
   to mean "a pack-file that is in the pack-file list of the
   multi-pack-index but none of the objects in the multi-pack-index
   reference a location inside that pack-file."

2. 'git multi-pack-index expire' deletes any unreferenced pack-files
   and updaes the multi-pack-index to drop those pack-files from the
   list. This is safe to do as concurrent Git processes will see the
   multi-pack-index and not open those packs when looking for object
   contents. (Similar to the 'loose-objects' job, there are some Git
   commands that open pack-files regardless of the multi-pack-index,
   but they are rarely used. Further, a user that self-selects to
   use background operations would likely refrain from using those
   commands.)

3. 'git multi-pack-index repack --bacth-size=<size>' collects a set
   of pack-files that are listed in the multi-pack-index and creates
   a new pack-file containing the objects whose offsets are listed
   by the multi-pack-index to be in those objects. The set of pack-
   files is selected greedily by sorting the pack-files by modified
   time and adding a pack-file to the set if its "expected size" is
   smaller than the batch size until the total expected size of the
   selected pack-files is at least the batch size. The "expected
   size" is calculated by taking the size of the pack-file divided
   by the number of objects in the pack-file and multiplied by the
   number of objects from the multi-pack-index with offset in that
   pack-file. The expected size approximates how much data from that
   pack-file will contribute to the resulting pack-file size. The
   intention is that the resulting pack-file will be close in size
   to the provided batch size.

   The next run of the incremental-repack task will delete these
   repacked pack-files during the 'expire' step.

   In this version, the batch size is set to "0" which ignores the
   size restrictions when selecting the pack-files. It instead
   selects all pack-files and repacks all packed objects into a
   single pack-file. This will be updated in the next change, but
   it requires doing some calculations that are better isolated to
   a separate change.

These steps are based on a similar background maintenance step in
Scalar (and VFS for Git) [1]. This was incredibly effective for
users of the Windows OS repository. After using the same VFS for Git
repository for over a year, some users had _thousands_ of pack-files
that combined to up to 250 GB of data. We noticed a few users were
running into the open file descriptor limits (due in part to a bug
in the multi-pack-index fixed by af96fe3 (midx: add packs to
packed_git linked list, 2019-04-29).

These pack-files were mostly small since they contained the commits
and trees that were pushed to the origin in a given hour. The GVFS
protocol includes a "prefetch" step that asks for pre-computed pack-
files containing commits and trees by timestamp. These pack-files
were grouped into "daily" pack-files once a day for up to 30 days.
If a user did not request prefetch packs for over 30 days, then they
would get the entire history of commits and trees in a new, large
pack-file. This led to a large number of pack-files that had poor
delta compression.

By running this pack-file maintenance step once per day, these repos
with thousands of packs spanning 200+ GB dropped to dozens of pack-
files spanning 30-50 GB. This was done all without removing objects
from the system and using a constant batch size of two gigabytes.
Once the work was done to reduce the pack-files to small sizes, the
batch size of two gigabytes means that not every run triggers a
repack operation, so the following run will not expire a pack-file.
This has kept these repos in a "clean" state.

[1] https://github.com/microsoft/scalar/blob/master/Scalar.Common/Maintenance/PackfileMaintenanceStep.cs

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>