view tests/test-largefiles-small-disk.t @ 38893:23d582caae30

changegroup: capture revision delta in a data structure

The current changegroup generation code is tightly coupled to
the revlog API. This tight coupling makes it difficult to implement
alternate storage backends without requiring a large surface area
of the revlog API to be exposed. This is not desirable.

In order to support changegroup generation with non-revlog storage,
we'll need to abstract the concept of delta generation.

This commit is the first step down that road. We introduce a
data structure for representing a delta in a changegroup.

The API still leaves a lot to be desired. But at least we now
have separation between data and actions performed on it.

As part of this, we tweak behavior slightly: we no longer
concatenate the delta prefix with the metadata header. Instead,
we track and emit the prefix as a separate chunk. This shouldn't
have any meaningful impact since all the chunks just get sent to
the wire, the compressor, etc.

Because we're introducing a new object, this does add some
overhead to changegroup execution. `hg perfchangegroupchangelog`
on my clone of the Mercurial repo (~40,000 visible revisions in
the changelog) slows down a bit:

! wall 1.268600 comb 1.270000 user 1.270000 sys 0.000000 (best of 8)
! wall 1.419479 comb 1.410000 user 1.410000 sys 0.000000 (best of 8)

With for `hg bundle -t none-v2 -a /dev/null`:

before: real 6.610 secs (user 6.460+0.000 sys 0.140+0.000)
after:  real 7.210 secs (user 7.060+0.000 sys 0.140+0.000)

I plan to claw back this regression in future commits. And I may
even do away with this data structure once the refactor is complete.
For now, it makes things easier to comprehend.

Differential Revision: https://phab.mercurial-scm.org/D4075

Test how largefiles abort in case the disk runs full

  $ cat > criple.py <<EOF
  > from __future__ import absolute_import
  > import errno
  > import os
  > import shutil
  > from mercurial import util
  > #
  > # this makes the original largefiles code abort:
  > _origcopyfileobj = shutil.copyfileobj
  > def copyfileobj(fsrc, fdst, length=16*1024):
  >     # allow journal files (used by transaction) to be written
  >     if b'journal.' in fdst.name:
  >         return _origcopyfileobj(fsrc, fdst, length)
  >     fdst.write(fsrc.read(4))
  >     raise IOError(errno.ENOSPC, os.strerror(errno.ENOSPC))
  > shutil.copyfileobj = copyfileobj
  > #
  > # this makes the rewritten code abort:
  > def filechunkiter(f, size=131072, limit=None):
  >     yield f.read(4)
  >     raise IOError(errno.ENOSPC, os.strerror(errno.ENOSPC))
  > util.filechunkiter = filechunkiter
  > #
  > def oslink(src, dest):
  >     raise OSError("no hardlinks, try copying instead")
  > util.oslink = oslink
  > EOF

  $ echo "[extensions]" >> $HGRCPATH
  $ echo "largefiles =" >> $HGRCPATH

  $ hg init alice
  $ cd alice
  $ echo "this is a very big file" > big
  $ hg add --large big
  $ hg commit --config extensions.criple=$TESTTMP/criple.py -m big
  abort: No space left on device
  [255]

The largefile is not created in .hg/largefiles:

  $ ls .hg/largefiles
  dirstate

The user cache is not even created:

  >>> import os; os.path.exists("$HOME/.cache/largefiles/")
  False

Make the commit with space on the device:

  $ hg commit -m big

Now make a clone with a full disk, and make sure lfutil.link function
makes copies instead of hardlinks:

  $ cd ..
  $ hg --config extensions.criple=$TESTTMP/criple.py clone --pull alice bob
  requesting all changes
  adding changesets
  adding manifests
  adding file changes
  added 1 changesets with 1 changes to 1 files
  new changesets 390cf214e9ac
  updating to branch default
  getting changed largefiles
  abort: No space left on device
  [255]

The largefile is not created in .hg/largefiles:

  $ ls bob/.hg/largefiles
  dirstate