As documented at revlog.__init__, revlog._cache stores raw text.
The current read and write usage of "_cache" in revlog.revision lacks of
raw=True check.
This patch fixes that by adding check about raw, and storing rawtext
explicitly in _cache.
Note: it may slow down cache hit code path when raw=False and flags=0. That
performance issue will be fixed in a later patch.
test-revlog-raw now points us to a new problem.
Narrowhg has been using "1 << 14" as its revlog flag value for a long
time. We (Google) have many repos with that value in production
already. When the same value was reserved for EXTSTORED, it made those
repos invalid. Upgrading them will be a little painful. We should
clearly have reserved the value for narrowhg a long time ago. Since
the EXTSTORED flag is not yet in any release and Facebook also says
they have not started using it in production, so it should be okay to
change it. This patch gives the current value (1 << 14) back to
narrowhg and gives a new value (1 << 13) to EXTSTORED.
The final part of integrating the compression manager APIs into
revlog storage is the plumbing for repositories to advertise they
are using non-zlib storage and for revlogs to instantiate a non-zlib
compression engine.
The main intent of the compression manager work was to zstd all
of the things. Adding zstd to revlogs has proved to be more involved
than other places because revlogs are... special. Very small inputs
and the use of delta chains (which are themselves a form of
compression) are a completely different use case from streaming
compression, which bundles and the wire protocol employ. I've
conducted numerous experiments with zstd in revlogs and have yet
to formalize compression settings and a storage architecture that
I'm confident I won't regret later. In other words, I'm not yet
ready to commit to a new mechanism for using zstd - or any other
compression format - in revlogs.
That being said, having some support for zstd (and other compression
formats) in revlogs in core is beneficial. It can allow others to
conduct experiments.
This patch introduces *highly experimental* support for non-zlib
compression formats in revlogs. Introduced is a config option to
control which compression engine to use. Also introduced is a namespace
of "exp-compression-*" requirements to denote support for non-zlib
compression in revlogs. I've prefixed the namespace with "exp-"
(short for "experimental") because I'm not confident of the
requirements "schema" and in no way want to give the illusion of
supporting these requirements in the future. I fully intend to drop
support for these requirements once we figure out what we're doing
with zstd in revlogs.
A good portion of the patch is teaching the requirements system
about registered compression engines and passing the requested
compression engine as an opener option so revlogs can instantiate
the proper compression engine for new operations.
That's a verbose way of saying "we can now use zstd in revlogs!"
On an `hg pull` conversion of the mozilla-unified repo with no extra
redelta settings (like aggressivemergedeltas), we can see the impact
of zstd vs zlib in revlogs:
$ hg perfrevlogchunks -c
! chunk
! wall 2.032052 comb 2.040000 user 1.990000 sys 0.050000 (best of 5)
! wall 1.866360 comb 1.860000 user 1.820000 sys 0.040000 (best of 6)
! chunk batch
! wall 1.877261 comb 1.870000 user 1.860000 sys 0.010000 (best of 6)
! wall 1.705410 comb 1.710000 user 1.690000 sys 0.020000 (best of 6)
$ hg perfrevlogchunks -m
! chunk
! wall 2.721427 comb 2.720000 user 2.640000 sys 0.080000 (best of 4)
! wall 2.035076 comb 2.030000 user 1.950000 sys 0.080000 (best of 5)
! chunk batch
! wall 2.614561 comb 2.620000 user 2.580000 sys 0.040000 (best of 4)
! wall 1.910252 comb 1.910000 user 1.880000 sys 0.030000 (best of 6)
$ hg perfrevlog -c -d 1
! wall 4.812885 comb 4.820000 user 4.800000 sys 0.020000 (best of 3)
! wall 4.699621 comb 4.710000 user 4.700000 sys 0.010000 (best of 3)
$ hg perfrevlog -m -d 1000
! wall 34.252800 comb 34.250000 user 33.730000 sys 0.520000 (best of 3)
! wall 24.094999 comb 24.090000 user 23.320000 sys 0.770000 (best of 3)
Only modest wins for the changelog. But manifest reading is
significantly faster. What's going on?
One reason might be data volume. zstd decompresses faster. So given
more bytes, it will put more distance between it and zlib.
Another reason is size. In the current design, zstd revlogs are
*larger*:
debugcreatestreamclonebundle (size in bytes)
zlib: 1,638,852,492
zstd: 1,680,601,332
I haven't investigated this fully, but I reckon a significant cause of
larger revlogs is that the zstd frame/header has more bytes than
zlib's. For very small inputs or data that doesn't compress well, we'll
tend to store more uncompressed chunks than with zlib (because the
compressed size isn't smaller than original). This will make revlog
reading faster because it is doing less decompression.
Moving on to bundle performance:
$ hg bundle -a -t none-v2 (total CPU time)
zlib: 102.79s
zstd: 97.75s
So, marginal CPU decrease for reading all chunks in all revlogs
(this is somewhat disappointing).
$ hg bundle -a -t <engine>-v2 (total CPU time)
zlib: 191.59s
zstd: 115.36s
This last test effectively measures the difference between zlib->zlib
and zstd->zstd for revlogs to bundle. This is a rough approximation of
what a server does during `hg clone`.
There are some promising results for zstd. But not enough for me to
feel comfortable advertising it to users. We'll get there...
Now that compression engines declare their header in revlog chunks
and can decompress revlog chunks, we refactor revlog.decompress()
to use them.
Making full use of the property that revlog compressor objects are
reusable, revlog instances now maintain a dict mapping an engine's
revlog header to a compressor object. This is not only a performance
optimization for engines where compressor object reuse can result in
better performance, but it also serves as a cache of header values
so we don't need to perform redundant lookups against the compression
engine manager. (Yes, I measured and the overhead of a function call
versus a dict lookup was observed.)
Replacing the previous inline lookup table with a dict lookup was
measured to make chunk reading ~2.5% slower on changelogs and ~4.5%
slower on manifests. So, the inline lookup table has been mostly
preserved so we don't lose performance. This is unfortunate. But
many decompression operations complete in microseconds, so Python
attribute lookup, dict lookup, and function calls do matter.
The impact of this change on mozilla-unified is as follows:
$ hg perfrevlogchunks -c
! chunk
! wall 1.953663 comb 1.950000 user 1.920000 sys 0.030000 (best of 6)
! wall 1.946000 comb 1.940000 user 1.910000 sys 0.030000 (best of 6)
! chunk batch
! wall 1.791075 comb 1.800000 user 1.760000 sys 0.040000 (best of 6)
! wall 1.785690 comb 1.770000 user 1.750000 sys 0.020000 (best of 6)
$ hg perfrevlogchunks -m
! chunk
! wall 2.587262 comb 2.580000 user 2.550000 sys 0.030000 (best of 4)
! wall 2.616330 comb 2.610000 user 2.560000 sys 0.050000 (best of 4)
! chunk batch
! wall 2.427092 comb 2.420000 user 2.400000 sys 0.020000 (best of 5)
! wall 2.462061 comb 2.460000 user 2.400000 sys 0.060000 (best of 4)
Changelog chunk reading is slightly faster but manifest reading is
slower. What gives?
On this repo, 99.85% of changelog entries are zlib compressed (the 'x'
header). On the manifest, 67.5% are zlib and 32.4% are '\0'. This patch
swapped the test order of 'x' and '\0' so now 'x' is tested first. This
makes changelogs faster since they almost always hit the first branch.
This makes a significant percentage of manifest '\0' chunks slower
because that code path now performs an extra test. Yes, I too can't
believe we're able to measure the impact of an if..elif with simple
string compares. I reckon this code would benefit from being written
in C...
This commit swaps in the just-added revlog compressor API into
the revlog class.
Instead of implementing zlib compression inline in compress(), we
now store a cached-on-first-use revlog compressor on each revlog
instance and invoke its "compress()" method.
As part of this, revlog.compress() has been refactored a bit to use
a cleaner code flow and modern formatting (e.g. avoiding
parenthesis around returned tuples).
On a mozilla-unified repo, here are the "compress" times for a few
commands:
$ hg perfrevlogchunks -c
! wall 5.772450 comb 5.780000 user 5.780000 sys 0.000000 (best of 3)
! wall 5.795158 comb 5.790000 user 5.790000 sys 0.000000 (best of 3)
$ hg perfrevlogchunks -m
! wall 9.975789 comb 9.970000 user 9.970000 sys 0.000000 (best of 3)
! wall 10.019505 comb 10.010000 user 10.010000 sys 0.000000 (best of 3)
Compression times did seem to slow down just a little. There are
360,210 changelog revisions and 359,342 manifest revisions. For the
changelog, mean time to compress a revision increased from ~16.025us to
~16.088us. That's basically a function call or an attribute lookup. I
suppose this is the price you pay for abstraction. It's so low that
I'm not concerned.
Upcoming patches will convert revlogs to use the compression engine
APIs to perform all things compression. The yet-to-be-introduced
APIs support a persistent "compressor" object so the same object
can be reused for multiple compression operations, leading to
better performance. In addition, compression engines like zstd
may wish to tweak compression engine state based on the revlog
(e.g. per-revlog compression dictionaries).
A global and shared decompress() function will shortly no longer
make much sense. So, we move decompress() to be a method of the
revlog class. It joins compress() there.
On the mozilla-unified repo, we can measure the impact of this change
on reading performance:
$ hg perfrevlogchunks -c
! chunk
! wall 1.932573 comb 1.930000 user 1.900000 sys 0.030000 (best of 6)
! wall 1.955183 comb 1.960000 user 1.930000 sys 0.030000 (best of 6)
! chunk batch
! wall 1.787879 comb 1.780000 user 1.770000 sys 0.010000 (best of 6
! wall 1.774444 comb 1.770000 user 1.750000 sys 0.020000 (best of 6)
"chunk" appeared to become slower but "chunk batch" got faster. Upon
further examination by running both sets multiple times, the numbers
appear to converge across all runs. This tells me that there is no
perceived performance impact to this refactor.
revlog.compress() compares the compressed size to the input size
and throws away the compressed data if it is larger than the input.
This is the correct thing to do, as storing compressed data that
is larger than the input takes up more storage space and makes reading
slower.
However, the comparison was implemented inconsistently. For the
streaming compression mode, we threw away the result if it was
greater than or equal to the input size. But for the one-shot
compression, we threw away the compression only if it was greater
than the input size!
This patch changes the comparison for the simple case so it is
consistent with the streaming case.
As a few tests demonstrate, this adds 1 byte to some revlog entries.
This is because of an added 'u' header on the chunk. It seems
somewhat wrong to increase the revlog size here. However, IMO the cost
of 1 byte in storage is insignificant compared to the performance gains
of avoiding decompression. This patch should invite questions around
the heuristic for throwing away compressed data. For example, I'd argue
we should be more liberal about rejecting compressed data, additionally
doing so where the number of bytes saved fails to reach a threshold.
But we can have this discussion another time.
Upcoming patches will introduce functionality for in-place
repository/store "upgrades." Copying the contents of a revlog
feels sufficiently low-level to warrant being in the revlog
class. So this commit implements that functionality.
Because full delta recomputation can be *very* expensive (we're
talking several hours on the Firefox repository), we support
multiple modes of execution with regards to delta (re)use. This
will allow repository upgrades to choose the "level" of
processing/optimization they wish to perform when converting
revlogs.
It's not obvious from this commit, but "addrevisioncb" will be
used for progress reporting.
Add the ability for revlog objects to process revision flags and apply
registered transforms on read/write operations.
This patch introduces:
- the 'revlog._processflags()' method that looks at revision flags and applies
flag processors registered on them. Due to the need to handle non-commutative
operations, flag transforms are applied in stable order but the order in which
the transforms are applied is reversed between read and write operations.
- the 'addflagprocessor()' method allowing to register processors on flags.
Flag processors are defined as a 3-tuple of (read, write, raw) functions to be
applied depending on the operation being performed.
- an update on 'revlog.addrevision()' behavior. The current flagprocessor design
relies on extensions to wrap around 'addrevision()' to set flags on revision
data, and on the flagprocessor to perform the actual transformation of its
contents. In the lfs case, this means we need to process flags before we meet
the 2GB size check, leading to performing some operations before it happens:
- if flags are set on the revision data, we assume some extensions might be
modifying the contents using the flag processor next, and we compute the
node for the original revision data (still allowing extension to override
the node by wrapping around 'addrevision()').
- we then invoke the flag processor to apply registered transforms (in lfs's
case, drastically reducing the size of large blobs).
- finally, we proceed with the 2GB size check.
Note: In the case a cachedelta is passed to 'addrevision()' and we detect the
flag processor modified the revision data, we chose to trust the flag processor
and drop the cachedelta.
Adding the ability to passing flags to addrevision instead of simply passing
default flags to _addrevision will allow extensions relying on flag transforms
to wrap around addrevision() in order to update revlog flags.
The first use case of this patch will be the lfs extension marking nodes as
stored externally when the contents are larger than the defined threshold.
One of the reasons leading to setting flags in addrevision() wrappers in the
flag processor design is that it allows to detect files larger than the 2GB
limit before the check is performed, which allows lfs to transform the contents
into metadata.
This patch introduces a new 'raw' argument (defaults to False) to revlog's
revision() and _addrevision() methods.
When the 'raw' argument is set to True, it indicates the revision data should be
handled as raw data by the flagprocessor.
Note: Given revlog.addgroup() calls are restricted to changegroup generation, we
can always set raw to True when calling revlog._addrevision() from
revlog.addgroup().
This patch adds a line that ensures we are not setting by mistake a set of flags
overlfowing the 2 bytes they are allocated. Given the way the data is packed in
the revlog header, overflowing 2 bytes will result in setting a wrong offset.
In many cases, _chunkraw() is called with startrev==endrev. When
this is true, we can avoid an extra index lookup and some other
minor operations.
On the mozilla-unified repo, `hg perfrevlogchunks -c` says this
has the following impact:
! read w/ reused fd
! wall 0.371846 comb 0.370000 user 0.350000 sys 0.020000 (best of 27)
! wall 0.337930 comb 0.330000 user 0.300000 sys 0.030000 (best of 30)
! read batch w/ reused fd
! wall 0.014952 comb 0.020000 user 0.000000 sys 0.020000 (best of 197)
! wall 0.014866 comb 0.010000 user 0.000000 sys 0.010000 (best of 196)
So, we've gone from ~25x slower than batch to ~22.5x slower.
At this point, there's probably not much else we can do except
implement an optimized function in the index itself, including in C.
When I implemented `hg perfrevlogchunks`, one of the things that
stood out was N * _chunk() calls was ~38x slower than 1
_chunks() call. Specifically, on the mozilla-unified repo:
N*_chunk: 0.528997s
1*_chunks: 0.013735s
This repo has 352,097 changesets. So the average time per changeset
comes out to:
N*_chunk: 1.502us
1*_chunks: 0.039us
If you extrapolate these numbers to a repository with 1M changesets,
that comes out to 1.502s versus 0.039s, which is significant.
At these latencies, Python attribute lookups and function calls
matter. So, this patch inlines some code to cut down on that overhead.
The impact of this patch on N*_chunk() calls is clear:
! wall 0.528997 comb 0.520000 user 0.500000 sys 0.020000 (best of 19)
! wall 0.367723 comb 0.370000 user 0.350000 sys 0.020000 (best of 27)
So, we go from ~38x slower to ~27x. A nice improvement. But there's
still a long way to go.
It's worth noting that functionality like revsets perform changelog
lookups one revision at a time. So this code path is worth optimizing.
Index entries are ordered tuples. We have accessors in the revlog
class to map tuple offsets to names. To help reinforce the order,
reorder the methods so they match the order of elements in the
tuple. While I'm here, also sneak in some minimal documentation.
This allows revlog-style files to be written out with checkambig=True
easily.
Because avoiding file stat ambiguity is needed only for filecache-ed
manifest and changelog, this patch does:
- use False for default value of checkambig
- focus only on writing changes of index file out
This patch also adds optional argument checkambig to _divert/_delay
for changelog, to safely accept checkambig specified in revlog
layer. But this argument can be fully ignored, because:
- changes are written into other than index file, if name != target
- changes are never written into index file, otherwise
(into pending file by _divert, or into in-memory buffer by _delay)
This is a part of ExactCacheValidationPlan.
https://www.mercurial-scm.org/wiki/ExactCacheValidationPlan
Profiling using statprof revealed a hotspot during changegroup
application calculating delta chain bases on generaldelta repos.
Essentially, revlog._addrevision() was performing a lot of redundant
work tracing the delta chain as part of determining when the chain
distance was acceptable. This was most pronounced when adding
revisions to manifests, which can have delta chains thousands of
revisions long.
There was a delta chain base cache on revlogs before, but it only
captured a single revision. This was acceptable before generaldelta,
when _addrevision would build deltas from the previous revision and
thus we'd pretty much guarantee a cache hit when resolving the delta
chain base on a subsequent _addrevision call. However, it isn't
suitable for generaldelta because parent revisions aren't necessarily
the last processed revision.
This patch converts the delta chain base cache to an LRU dict cache.
The cache can hold multiple entries, so generaldelta repos have a
higher chance of getting a cache hit.
The impact of this change when processing changegroup additions is
significant. On a generaldelta conversion of the "mozilla-unified"
repo (which contains heads of the main Firefox repositories in
chronological order - this means there are lots of transitions between
heads in revlog order), this change has the following impact when
performing an `hg unbundle` of an uncompressed bundle of the repo:
before: 5:42 CPU time
after: 4:34 CPU time
Most of this time is saved when applying the changelog and manifest
revlogs:
before: 2:30 CPU time
after: 1:17 CPU time
That nearly a 50% reduction in CPU time applying changesets and
manifests!
Applying a gzipped bundle of the same repo (effectively simulating a
`hg clone` over HTTP) showed a similar speedup:
before: 5:53 CPU time
after: 4:46 CPU time
Wall time improvements were basically the same as CPU time.
I didn't measure explicitly, but it feels like most of the time
is saved when processing manifests. This makes sense, as large
manifests tend to have very long delta chains and thus benefit the
most from this cache.
So, this change effectively makes changegroup application (which is
used by `hg unbundle`, `hg clone`, `hg pull`, `hg unshelve`, and
various other commands) significantly faster when delta chains are
long (which can happen on repos with large numbers of files and thus
large manifests).
In theory, this change can result in more memory utilization. However,
we're caching a dict of ints. At most we have 200 ints + Python object
overhead per revlog. And, the cache is really only populated when
performing read-heavy operations, such as adding changegroups or
scanning an individual revlog. For memory bloat to be an issue, we'd
need to scan/read several revisions from several revlogs all while
having active references to several revlogs. I don't think there are
many operations that do this, so I don't think memory bloat from the
cache will be an issue.
Before fd1bb7c, if the C index.partialmatch raises RevlogError, the Python
code raises "ambiguous identifier" error immediately, which is efficient.
fd1bb7c took hidden revisions into consideration and forced the slow path
enumerating the changelog to double-check hidden revisions. But it's not
necessary if we know the revlog has no hidden revisions.
This patch adds back the fast path for unfiltered revlogs.
Previously, we likely called _chunkraw() multiple times in order to
ensure it didn't change out from under us. I'm pretty certain this code
had its origins in the days where we attempted to have thread safety of
localrepository and thus revlog instances.
revlog instances are already not thread safe for writing. And, as of
Mercurial 3.6, hgweb uses a separate localrepository instance per
request, so there should only be a single thread reading a revlog at
a time. We more or less decided that attempting to make classes like
revlog thread safe is a lost cause.
So, this patch removes thread safety from _chunks. As a result, we make
one less call into _chunkraw() when the initial read isn't serviced
by the cache. This translates to savings of 4 function calls overall
and possibly prevents the creation of an additional buffer view into the
cache. I doubt this translates into any real world performance wins
because decompression will almost certainly dwarf time spent in
_chunks(). But it does make the code simpler, so it is an improvement.
A subsequent patch will refactor _chunks() and the calculation of the
offset will no longer occur in that function. Prepare by returning the
offset from _chunkraw().
Revlog chunks can be stored uncompressed. If the first byte of the
raw data is \0, we store the data as is. Else we prefix it with 'u'.
Before, we performed a string slice to strip out the 'u' prefix.
With this patch, we use a buffer to avoid an extra memory copy and
associated garbage collection overhead.
I was unable to verify any performance impact of this patch. For both
mozilla-central and the hg repos, the number of manifest revisions
with 'u' prefixes is very small - under 1%. So this change likely
isn't called enough to have an impact on manifest reading. However,
the reasoning behind this change is solid, so it should be safe.
This code is already written in multiple locations.
While this code needs to be fast and extracting it to its own function
adds overhead, code paths reading delta chains typically read,
decompress, and do binary patching on revlog data from the delta chain.
This other work (especially zlib decompression) almost certainly
accounts for a lot more time than the overhead of introducing a Python
function call. So I'm not worried about the performance impact of this
change.
clearcaches() was added several years ago in 1e47437a1ca7 as part
of implementing a perf command. Since revlog instances have many caches
and since the spirit of this mostly unused method is to facilitate
performance testing, I think it's appropriate for all the revlog's
caches to get cleared when it is called.
Revlogs were recently refactored to open file handles in "a+" and use a
persistent file handle for reading and writing. This drastically
reduced the number of file handles being opened.
Unfortunately, it appears that some versions of Solaris lose the file
offset when performing a write after the handle has been seeked.
The simplest workaround is to seek to EOF on files opened in a+ mode
before writing to them, which is what this patch does.
Ideally, this code would exist in the vfs layer. However, this would
require creating a proxy class for file objects in order to provide a
custom implementation of write(). This would add overhead. Since
revlogs are the only files we open in a+ mode, the one-off workaround
in revlog.py should be sufficient.
This patch appears to have little to no impact on performance on my
Linux machine.
In the most complex case, we try using the incoming delta base, then
we try both parents, and then we try the previous revlog entry. If
none of these result in a good delta, we natually use the null
revision as base. However, we sometimes consider the nullrev before we
have exhausted our other options. Specifically, when both parents are
null, we use the nullrev as delta base if it produces a good delta
(according to _isgooddelta()), and we fail to try the previous revlog
entry as delta base. After e60126c6093d (addrevision: use general
delta when the incoming base delta is bad, 2015-12-01), it can also
happen for non-merge commits when the incoming delta is not good.
The Firefox repo (from many months back) shrinks a tiny bit with this
patch: from 1.855GB to 1.830GB (1.4%). The hg repo itself shrinks even
less: by less than 0.1%. There may be repos that get larger instead.
This undoes the unexplained test change in e60126c6093d.
