view tests/test-narrow-commit.t @ 40326:fed697fa1734

sqlitestore: file storage backend using SQLite

This commit provides an extension which uses SQLite to store file
data (as opposed to revlogs).

As the inline documentation describes, there are still several
aspects to the extension that are incomplete. But it's a start.
The extension does support basic clone, checkout, and commit
workflows, which makes it suitable for simple use cases.

One notable missing feature is support for "bundlerepos." This is
probably responsible for the most test failures when the extension
is activated as part of the test suite.

All revision data is stored in SQLite. Data is stored as zstd
compressed chunks (default if zstd is available), zlib compressed
chunks (default if zstd is not available), or raw chunks (if
configured or if a compressed delta is not smaller than the raw
delta). This makes things very similar to revlogs.

Unlike revlogs, the extension doesn't yet enforce a limit on delta
chain length. This is an obvious limitation and should be addressed.
This is somewhat mitigated by the use of zstd, which is much faster
than zlib to decompress.

There is a dedicated table for storing deltas. Deltas are stored
by the SHA-1 hash of their uncompressed content. The "fileindex" table
has columns that reference the delta for each revision and the base
delta that delta should be applied against. A recursive SQL query
is used to resolve the delta chain along with the delta data.

By storing deltas by hash, we are able to de-duplicate delta storage!
With revlogs, the same deltas in different revlogs would result in
duplicate storage of that delta. In this scheme, inserting the
duplicate delta is a no-op and delta chains simply reference the
existing delta.

When initially implementing this extension, I did not have
content-indexed deltas and deltas could be duplicated across files
(just like revlogs). When I implemented content-indexed deltas, the
size of the SQLite database for a full clone of mozilla-unified
dropped:

before: 2,554,261,504 bytes
after:  2,488,754,176 bytes

Surprisingly, this is still larger than the bytes size of revlog
files:

revlog files: 2,104,861,230 bytes
du -b:        2,254,381,614

I would have expected storage to be smaller since we're not limiting
delta chain length and since we're using zstd instead of zlib. I
suspect the SQLite indexes and per-column overhead account for the
bulk of the differences. (Keep in mind that revlog uses a 64-byte
packed struct for revision index data and deltas are stored without
padding. Aside from the 12 unused bytes in the 32 byte node field,
revlogs are pretty efficient.) Another source of overhead is file
name storage. With revlogs, file names are stored in the filesystem.
But with SQLite, we need to store file names in the database. This is
roughly equivalent to the size of the fncache file, which for the
mozilla-unified repository is ~34MB.

Since the SQLite database isn't append-only and since delta chains
can reference any delta, this opens some interesting possibilities.
For example, we could store deltas in reverse, such that fulltexts
are stored for newer revisions and deltas are applied to reconstruct
older revisions. This is likely a more optimal storage strategy for
version control, as new data tends to be more frequently accessed
than old data. We would obviously need wire protocol support for
transferring revision data from newest to oldest. And we would
probably need some kind of mechanism for "re-encoding" stores. But
it should be doable.

This extension is very much experimental quality. There are a handful
of features that don't work. It probably isn't suitable for day-to-day
use. But it could be used in limited cases (e.g. read-only checkouts
like in CI). And it is also a good proving ground for alternate
storage backends. As we continue to define interfaces for all things
storage, it will be useful to have a viable alternate storage backend
to see how things shake out in practice.

test-storage.py passes on Python 2 and introduces no new test failures on
Python 3. Having the storage-level unit tests has proved to be insanely
useful when developing this extension. Those tests caught numerous bugs
during development and I'm convinced this style of testing is the way
forward for ensuring alternate storage backends work as intended. Of
course, test coverage isn't close to what it needs to be. But it is
a start. And what coverage we have gives me confidence that basic store
functionality is implemented properly.

Differential Revision: https://phab.mercurial-scm.org/D4928

#testcases flat tree

  $ . "$TESTDIR/narrow-library.sh"

#if tree
  $ cat << EOF >> $HGRCPATH
  > [experimental]
  > treemanifest = 1
  > EOF
#endif

create full repo

  $ hg init master
  $ cd master

  $ mkdir inside
  $ echo inside > inside/f1
  $ mkdir outside
  $ echo outside > outside/f1
  $ hg ci -Aqm 'initial'

  $ echo modified > inside/f1
  $ hg ci -qm 'modify inside'

  $ echo modified > outside/f1
  $ hg ci -qm 'modify outside'

  $ cd ..

(The lfs extension does nothing here, but this test ensures that its hook that
determines whether to add the lfs requirement, respects the narrow boundaries.)

  $ hg --config extensions.lfs= clone --narrow ssh://user@dummy/master narrow \
  >    --include inside
  requesting all changes
  adding changesets
  adding manifests
  adding file changes
  added 3 changesets with 2 changes to 1 files
  new changesets *:* (glob)
  updating to branch default
  1 files updated, 0 files merged, 0 files removed, 0 files unresolved
  $ cd narrow

  $ hg update -q 0

Can not modify dirstate outside

  $ mkdir outside
  $ touch outside/f1
  $ hg debugwalk -v -I 'relglob:f1'
  * matcher:
  <includematcher includes='(?:(?:|.*/)f1(?:/|$))'>
  f  inside/f1  inside/f1
  $ hg add .
  $ hg add outside/f1
  abort: cannot track 'outside/f1' - it is outside the narrow clone
  [255]
  $ touch outside/f3
  $ hg add outside/f3
  abort: cannot track 'outside/f3' - it is outside the narrow clone
  [255]

But adding a truly excluded file shouldn't count

  $ hg add outside/f3 -X outside/f3

  $ rm -r outside

Can modify dirstate inside

  $ echo modified > inside/f1
  $ touch inside/f3
  $ hg add inside/f3
  $ hg status
  M inside/f1
  A inside/f3
  $ hg revert -qC .
  $ rm inside/f3

Can commit changes inside. Leaves outside unchanged.

  $ hg update -q 'desc("initial")'
  $ echo modified2 > inside/f1
  $ hg manifest --debug
  4d6a634d5ba06331a60c29ee0db8412490a54fcd 644   inside/f1
  7fb3bb6356d28d4dc352c5ba52d7350a81b6bd46 644   outside/f1 (flat !)
  d0f2f706468ab0e8bec7af87446835fb1b13511b 755 d outside/ (tree !)
  $ hg commit -m 'modify inside/f1'
  created new head
  $ hg files -r .
  inside/f1
  $ hg manifest --debug
  3f4197b4a11b9016e77ebc47fe566944885fd11b 644   inside/f1
  7fb3bb6356d28d4dc352c5ba52d7350a81b6bd46 644   outside/f1 (flat !)
  d0f2f706468ab0e8bec7af87446835fb1b13511b 755 d outside/ (tree !)
Some filesystems (notably FAT/exFAT only store timestamps with 2
seconds of precision, so by sleeping for 3 seconds, we can ensure that
the timestamps of files stored by dirstate will appear older than the
dirstate file, and therefore we'll be able to get stable output from
debugdirstate. If we don't do this, the test can be slightly flaky.
  $ sleep 3
  $ hg status
  $ hg debugdirstate --no-dates
  n 644         10 set                 inside/f1