The "if visible" guard is now pretty pointless, because the function
call it guards will do almost no work anyway when there are no visible
revisions.
We can also stop wrapping "visible" in a set since it just needs to be
an iterable now.
The pinned revs are simply revisions that should not be hidden even if
hideablerevs() says that should. Let's make that clear by simply
setting "hidden = hideablerevs() - pinnedrevs()" early on.
Roughly speaking, we currently do this to reveal hidden ancestors of
visible revisions:
1. Iterate over all visible non-public revisions and see if they have
hidden parents
2. For each revision found in step (1) walk the chain of hidden
commits and reveal it
We can simplify that by skipping step (1) and doing step (2) from all
visible non-public revisions instead.
This doesn't seem to have much impact on "perfvolatilesets".
Before:
! obsolete
! wall 0.004616 comb 0.000000 user 0.000000 sys 0.000000 (best of 570)
! visible
! wall 0.008235 comb 0.010000 user 0.010000 sys 0.000000 (best of 326)
After:
! obsolete
! wall 0.004727 comb 0.010000 user 0.010000 sys 0.000000 (best of 543)
! visible
! wall 0.008371 comb 0.000000 user 0.000000 sys 0.000000 (best of 324)
Instead of passing the domain into _consistencyblockers() and having
the function calculate the set of revisions to iterate over, let the
caller do it. This is just a minor refactoring to make future changes
simpler.
The function will stop searching as soon as it runs into a non-hidden
revision, so there is no need to restrict by the domain (of mutable
revisions) as well.
This doesn't seem to have much impact on "perfvolatilesets".
Before:
! obsolete
! wall 0.004903 comb 0.000000 user 0.000000 sys 0.000000 (best of 535)
! visible
! wall 0.008913 comb 0.010000 user 0.010000 sys 0.000000 (best of 300)
After:
! obsolete
! wall 0.004616 comb 0.000000 user 0.000000 sys 0.000000 (best of 570)
! visible
! wall 0.008235 comb 0.010000 user 0.010000 sys 0.000000 (best of 326)
This change makes the function actually reveal the ancestors by
removing them from the hidden set. This prepares for further
simplification.
Note that the function will now only reveal contiguous chains of
hidden revisions, but that's fine because we always pass it an
immediate child of any revision that should be revealed (or the
revision itself).
This doesn't seem to have much impact on "perfvolatilesets".
Before:
! obsolete
! wall 0.004672 comb 0.010000 user 0.010000 sys 0.000000 (best of 590)
! visible
! wall 0.008936 comb 0.010000 user 0.010000 sys 0.000000 (best of 322)
After:
! obsolete
! wall 0.004903 comb 0.000000 user 0.000000 sys 0.000000 (best of 535)
! visible
! wall 0.008913 comb 0.010000 user 0.010000 sys 0.000000 (best of 300)
E.g. tags and bookmarks can reveal revisions that would otherwise be
hidden. A revision can also be revealed because one if its descendants
is visible. Let's use the term "pinned" for the former case
(bookmarks etc.).
The improvement in time complexitty and the speed-up in computation is large
enough that the has little use now. Its update time can even gets in the way. So
we drop it.
This will allow us to unify the static/dynamic blockers logic in the next
changeset.
For a couple of years, we now have precomputed set for all mutable phases. We
can use this set restrict our search and quickly detect non-hideable children of
hideable changesets. This speeds up the hidden computation. See docstring of
the new function for details.
This new version reuses the '_domainancestors' function to keep the computation
of revealed changeset in O(len(visible))
Below are perfvolatilesets timing from two Mozilla repositories with different
contents. hidden cache is disabled while obtaining them.
1) Mozilla repository with:
* 400667 changesets
* 35 hidden changesets (first rev-268334)
* 288 visible drafts
* 1 unstable changeset
Before:
! visible
! wall 0.001744 comb 0.000000 user 0.000000 sys 0.000000 (best of 1563)
After:
! visible
! wall 0.000742 comb 0.000000 user 0.000000 sys 0.000000 (best of 3755)
The timing above include the computation of obsolete changeset:
! obsolete
! wall 0.000396 comb 0.000000 user 0.000000 sys 0.000000 (best of 6816)
So adjusted time give 1.3ms before versus 0.3ms after. A 4x speedup.
2) Mozilla repository with:
* 405645 changesets
* 4312 hidden changesets (first rev-326004)
* 264 visible drafts
* 1 unstable changeset
Before:
! visible
! wall 0.025476 comb 0.030000 user 0.030000 sys 0.000000 (best of 111)
After
! visible
! wall 0.007703 comb 0.010000 user 0.010000 sys 0.000000 (best of 358)
The timing above include the computation of obsolete changeset:
! obsolete
! wall 0.006408 comb 0.010000 user 0.010000 sys 0.000000 (best of 404)
So adjusted time give 19ms before versus 1.3ms after. A 17x speedup.
The complexity of computing the revealed changesets is now 'O(revealed)'.
This massively speeds up the computation on large repository. Moving it to the
millisecond range.
Below are timing from two Mozilla repositories with different contents:
1) mozilla repository with:
* 400667 changesets
* 35 hidden changesets (first rev-268334)
* 288 visible drafts
* obsolete working copy (dynamicblockers),
Before:
! visible
! wall 0.030247 comb 0.030000 user 0.030000 sys 0.000000 (best of 100)
After:
! visible
! wall 0.000585 comb 0.000000 user 0.000000 sys 0.000000 (best of 4221)
The timing above include the computation of obsolete changeset:
! obsolete
! wall 0.000396 comb 0.000000 user 0.000000 sys 0.000000 (best of 6816)
So adjusted time give 30ms before versus 0.2ms after. A 150x speedup.
2) mozilla repository with:
* 405645 changesets
* 4312 hidden changesets (first rev-326004)
* 264 visible drafts
* obsolete working copy (dynamicblockers),
Before:
! visible
! wall 0.168658 comb 0.170000 user 0.170000 sys 0.000000 (best of 48)
After
! visible
! wall 0.008612 comb 0.010000 user 0.010000 sys 0.000000 (best of 325)
The timing above include the computation of obsolete changeset:
! obsolete
! wall 0.006408 comb 0.010000 user 0.010000 sys 0.000000 (best of 404)
So adjusted time give 160ms before versus 2ms after. A 75x speedup.
Recent mailing list discussion made me realised we could clarify these. We make
the function "public" to encourage extensions to wrap it and we use a more
explicit name that mirror "hideablerevs".
All versions of Python we support or hope to support make the hash
functions available in the same way under the same name, so we may as
well drop the util forwards.
The atomictemp.close() file attempts to do a rename, which can fail.
Moving the close inside the exception handler fixes it.
This doesn't fit well with the with: pattern, as it's the finalizer
that's failing.
Before this patch, revisions rollbacked at failure of previous
transaction might be visible at subsequent operations unintentionally,
if repoview object is reused even after failure of transaction:
e.g. command server and HTTP server are typical cases.
'repoview' uses the tuple of values below of unfiltered changelog as
"the key" to examine validity of filtered changelog cache.
- length
- tip node
- filtered revisions (as hashed value)
- '_delayed' field
'repoview' compares between "the key" of unfiltered changelog at
previous caching and now, and reuses filtered changelog cache if no
change is detected.
But this comparison indicates only that there is no change between
unfiltered 'repo.changelog' at last caching and now, but not that
filtered changelog cache is valid for current unfiltered one.
'repoview' uses "shallow copy" of unfiltered changelog to create
filtered changelog cache. In this case, 'index' buffer of unfiltered
changelog is also referred by filtered changelog.
At failure of transaction, unfiltered changelog itself is invalidated
(= un-referred) on the 'repo' side (see b7829fc79508 also). But
'index' of it still contains revisions to be rollbacked at this
failure, and is referred by filtered changelog.
Therefore, even if there is no change between unfiltered
'repo.changelog' at last caching and now, steps below makes rollbacked
revisions visible via filtered changelog unintentionally.
1. instantiate unfiltered changelog as 'repo.changelog'
(call it CL1)
2. make filtered (= shallow copy of) CL1
(call it FCL1)
3. cache FCL1 with "the key" of CL1
4. revisions are appended to 'index', which is shared by CL1 and FCL1
5. invalidate 'repo.changelog' (= CL1) at failure of transaction
6. instantiate 'repo.changelog' again at next operation
(call it CL2)
CL2 doesn't have revisions added at (4), because it is
instantiated from '00changelog.i', which isn't changed while
failed transaction.
7. compare between "the key" of CL1 and CL2
8. FCL1 cached at (3) is reused, because comparison at (7) doesn't
detect change between CL1 at (1) and CL2
9. revisions rollbacked at (5) are visible via FCL1 unintentionally,
because FCL1 still refers 'index' changed at (4)
The root cause of this issue is that there is no examination about
validity of filtered changelog cache against current unfiltered one.
This patch discards filtered changelog cache, if its 'index' object
isn't shared with unfiltered one.
BTW, at the time of this patch, redundant truncation of
'00changelog.i' at failure of transaction (see b7829fc79508 for
detail) often prevents "hg serve" from making already rollbacked
revisions visible, because updating timestamps of '00changelog.i' by
truncation makes "hg serve" discard old repoview object with invalid
filtered changelog cache.
This is reason why this issue is overlooked before this patch, even
though test-bundle2-exchange.t has tests in similar situation: failure
of "hg push" via HTTP by pretxnclose hook on server side doesn't
prevent subsequent commands from looking up outgoing revisions
correctly.
But timestamp on the filesystem doesn't have enough resolution for
recent computation power, and it can't be assumed that this avoidance
always works as expected.
Therefore, without this patch, this issue might appear occasionally.
Before this patch if the hiddencache existed but was empty, it would crash
mercurial. This patch adds exception handling when reading the hiddencache to
avoid the issue.
When encountering a corrupted cache file we print a devel warning. There would
be no point in issuing a normal warning as the user wouldn't be able to do
anything about the situation.
The warning looks like:
devel-warn: corrupted hidden cache, removing it at: /path/to/repoview.py
Getting the data necessary for the cache key using the changelog/revlog method
adds a significant overhead. Given how simple the underlying implementation is
and often this code is ran, it makes sense to violate layering and directly
compute the data.
Testing `hg log` on Mozilla-central, this reduce the time spent on changelog
cache validation by an extra half:
before: 12.2s of 69s
after: 6.1s of 62s
Total speed up from this patch and it's parent is 3x
(With stupid python profiler overhead)
The global speedup without profiler overhead is still there,
Before: 51s
After: 39s (-23%)
As explained in the comment, we were computing the key of the cache value every
time because of some obscure MQ test failure. I've dropped that code and ran the
test again that failure is gone. I assume some transaction cleanup got rid of
it.
So we are dropping that code. This provide a significant speedup.
Testing `hg log` on Mozilla-central this reduce the time spent on changelog
cache validation by a third:
before: 19.5s of 80s
after: 12.2s of 69s
(With stupid python profiler overhead)
Since all children are processed before their parents, we can apply the following algorithm:
For each rev (descending order):
* If I'm still hidden, no children will block me,
* If I'm not hidden, I must remove my parent from the hidden set,
This allows us to dynamically change the set of 'hidden' revisions, dropping the
need for the 'actuallyhidden' dictionary and the 'blocked' boolean in the queue.
As before, we start iterating from all heads and stop at the first public
changesets. This ensures the hidden computation is 'O(not public())' instead of
'O(len(min(not public()):))'.
Since we want to process all non-public changesets from top to bottom, a heap
seems more appropriate. This will ensure any revision is processed after all
its children, opening the way to code simplification.
Previously we would compute the repoview's static blockers by finding all the
children of hidden commits that were not hidden. This was O(number of commits
since first hidden change) since 'children' requires walking every commit from
tip until the first hidden change.
The new algorithm walks all heads down until it sees a public commit. This makes
the computation O(number of draft) commits, which is much faster in large
repositories with a large number of commits and a low number of drafts.
On a large repo with 1000+ obsolete markers and the earliest draft commit around
tip~200000, this improves computehidden perf by 200x (2s to 0.01s).
Starting with b1e85ff3a7fc, when a clone reached the checkout stage,
the cached changelog in the filtered view was still seeing the
_delayed flag, even though the changelog had already been finalized.
Monkey patching repoview does not really work and making it really work will
be really hard. So we better have it broken without complexity than broken with
extra complexity.
It doesn't seem to be a common idiom for repo instances, but the status() method
is replaced in largefiles' purge() override. Since __setattr__ is implemented
in repoview to setattr() on the unfiltered repo, the replacement method wouldn't
get called unless it was invoked with the unfiltered repo, because the filtered
repo remains unchanged.
Since this doesn't seem to be commonly used, I didn't bother to filter out
methods that perhaps shouldn't be replaced, such as changelog().
We have been running hiddencache verification since 3.1.1 and so
far not received a bug report concerning it. Therefore we remove
the verification code and make the hiddencache authoritive. That
way we get the intended speedup.
Incidentally, this avoids the changelog cache being invalidated each time
it's accessed on a repoview.
On a filtering experiment on a repository the size of mozilla-central,
this makes a significant difference:
Before, running hg log -l 10 --time with about 8k changesets filtered out:
time: real 1.490 secs (user 1.450+0.000 sys 0.040+0.000)
After:
time: real 0.540 secs (user 0.530+0.000 sys 0.010+0.000)