The main manifestctx implementation has a find function, so the hybrid one needs
it as well. I tried to add a test for this, but the only known case is in a
weird situation where there are tags, but no tag caches and some other criteria
I haven't figured out yet. So I gave up.
Summary:
hybridmanifest.__nonzero__() was explicitly trying to call __nonzero__() on the
underlying manifest object. This breaks when the underlying manifest is a
manifestdict. manifestdict does not implement __nonzero__ (it instead has a
__len__ method that gets used instead when evaluating the manifestdict as a
boolean).
Because of this issue I was getting a crash in a local script when calling
repo.commitctx() with a memctx object.
Test Plan: Confirmed that calling repo.commitctx() with a memctx no longer crashes.
Reviewers: #sourcecontrol, durham, quark
Reviewed By: quark
Subscribers: net-systems-diffs@, yogeshwer, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4191858
Signature: t1:4191858:1479330743:7adb668c9529a8c8abaf9a8a1e64c8ee78b64d86
Summary:
Upstream has refactored the manifest class into several classes, so we need to
update treemanifest to work with the new structure. Notably, the factory add
function previously relied on the ability for the revlog class to create a new
manifestdict (via manifest.maniest.read()), since this isn't possible anymore,
we have to construct the hybridmanifest ourselves and provide an appropriate
loadflat function to get the flat manifest if necessary.
Test Plan: Ran the tests
Reviewers: #mercurial, rmcelroy
Reviewed By: rmcelroy
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4180891
Signature: t1:4180891:1479285991:82bc546a1eb682d3cfd8b4724bda575410405d0f
Summary:
Upstream Mercurial has refactored the manifest to get rid of the manifest class.
This patch updates fastmanifest to work with the new class structure. In
particular, it removes the hacky wrapping of 3 different manifest construction
functions, with a single wrapping of manifestlog.get(), which makes the code
simpler and more robust.
Test Plan: Ran the tests
Reviewers: #mercurial, rmcelroy
Reviewed By: rmcelroy
Subscribers: rmcelroy, stash, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4180874
Signature: t1:4180874:1479286125:bbd2a36aa86237d68036b9d7b0a580829219d869
Summary:
fastmanifest makes bundling slower because fastmanifest `readshallowfast` always returns full manifest.
That's a big problem for infinitepush.
Let's copy-paste readshallowfast implementation from upstream. It uses readshallowdelta() if possible.
Test Plan:
1) Run all the tests for fb-hgext
2) Run infinitepush with this extension enabled. Make it is fast
Reviewers: durham, simonfar, rmcelroy, quark
Reviewed By: quark
Subscribers: mjpieters, #sourcecontrol
Differential Revision: https://phabricator.intern.facebook.com/D4088360
Tasks: 13907166
Signature: t1:4088360:1477580931:746e4054380403abbc52d1922583021b81f31bb6
Summary: Tests are failing because readdelta() in fastmanifest doesn't work after an upstream change. Copy and fix the upstream code to work in this case.
Test Plan: Run the test on my devserver
Reviewers: #mercurial, quark
Reviewed By: quark
Subscribers: mjpieters, #sourcecontrol
Differential Revision: https://phabricator.intern.facebook.com/D4074517
Tasks: 14070483
Signature: t1:4074517:1477412975:bdfb6e81e4ea3bccafa59a5cb0a6e71fcb1e7753
Summary:
This makes local commits get written to trees (as well as flat manifest) when a
commit happens where the parent commit has a tree manifest already. During a
transaction where multiple trees are written (like when rebasing multiple
nodes), we reuse the same pack file for all the trees produced by tieing into
the transaction abort and close hooks.
Test Plan:
Ran the tests. Ran hg commit with the extension enabled. A future
patch will add an integration test for the treemanifest extension.
Reviewers: #mercurial, quark
Reviewed By: quark
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4055851
Signature: t1:4055851:1477059659:91b1c2f93ef986e910cea752ebf2466cb20ac921
Summary:
In a future diff we will start allowing local trees to be written as part of
transactions. Part of this involves making those written trees available to
readers in this process before the transaction is committed. Since all manifest
reads go through hybridmanifest, we can insert a cache at this layer (just like
fastmanifest already has a cache at this layer).
The future patch which adds new manifests will populate this cache, and clear
this cache when the transaction completes.
Test Plan:
Ran the tests. Also tested the actual flow with the future patch that
causes commit to write to the cache
Reviewers: #mercurial, quark
Reviewed By: quark
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4055838
Signature: t1:4055838:1477059568:0bf1d79cc1ca6f9bcefb881bab35c61c13641f4e
Summary:
It's possible for treemanifests to be created that aren't backed by a flat
manifest (like if you did `foo = treemf.matches(...)`). In these cases, if we
ever need to diff it with a flat text, we need to be able to convert it to flat.
This patches adds the ability to convert it to a flat manifest via the text()
function.
Test Plan: Ran the tests
Reviewers: #mercurial, quark
Reviewed By: quark
Subscribers: quark, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4055656
Signature: t1:4055656:1477056921:7417c9ca06d61c5a283c6cf9693d83650e39f196
Summary:
This fixes some (but not all) of the fastmanifest errors. The remaining ones are
from an import * and one that will require slightly more refactoring to fix.
Test Plan: Ran the tests
Reviewers: #mercurial, quark
Reviewed By: quark
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4055633
Signature: t1:4055633:1477050763:09e758b1bd72b18799cdabf510eb0b2332077833
Summary:
The fastmanifest code had a race condition where committing quickly in
a loop could cause a stack trace. It was caused by the code check m.incache()
and then using m._cachedmanifest(), and between those two lines a background
process could have cleared the entry fromm the cache. So now we just check if
_cachedmanifest() actually returns a value.
Test Plan:
With and without the fix:
```
cd ~/local && rm -rf temprepo ; hg init temprepo && cd temprepo && for j in `seq -f "%03g" 1 100`; do for i in `seq -f "%03g" 1 10`; do echo "line$j" >> file$i.txt; done; echo "Commit $j done"; hg ci -Am "Commit $j" ; done
```
It would stack trace reliably without the fix, and never with the fix.
Reviewers: ikostia, #mercurial, quark
Reviewed By: quark
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D4004376
Tasks: 13828535
Signature: t1:4004376:1476266037:0ff33a43fd35b5c7309fbfbf546799d75cc03ebc
Summary:
Upstream 4718718ed358 requires a `readfast` method for manifestctx.
`hybridmanifestctx` does not have such method so let's add it.
Test Plan: Added a test.
Reviewers: durham, ttung, #sourcecontrol, rmcelroy
Reviewed By: rmcelroy
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3955478
Signature: t1:3955478:1475582074:f80c9eb97bcd42846625727ef1da4bce3256c88c
Summary:
This fixes a crash that `hybridmanifestctx` does not have `readdelta`
after upstream's b404425704fa. `readdelta` should be fast for vanilla
manifest so just forward the call there.
Test Plan: Modify a file that `.` touched, run `hg amend` and it won't crash.
Reviewers: ttung, #sourcecontrol, durham
Reviewed By: durham
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3948989
Signature: t1:3948989:1475198696:d8ae194500fc093c03cc115f4613a9b8c8bfbff2
Summary:
Upstream no longer has manifestctx derive from manifestdict; instead it uses a
read function. Let's update hybridmanifestctx to do the same.
Test Plan: Ran the fastmanifest tests
Reviewers: #sourcecontrol, ttung
Reviewed By: ttung
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3863656
Signature: t1:3863656:1473878502:6fb01f9f23fe64aafdc366a7ae98d510f3fb7927
Summary: @durham's changes to manifest.py landed, so we need to update where we plug into fastmanifest.
Test Plan: `PYTHONPATH=~/work/mercurial/facebook-hg-rpms/fb-hgext/:~/work/mercurial/facebook-hg-rpms/lz4revlog/:~/work/mercurial/facebook-hg-rpms/remotenames/ /opt/local/bin/python2.7 ~/work/mercurial/facebook-hg-rpms/hg-crew/tests/run-tests.py -j32 test-fastmanifest*.{py,t}`
Reviewers: rmcelroy, durham
Reviewed By: durham
Subscribers: mitrandir, durham, mjpieters, lcharignon
Differential Revision: https://phabricator.intern.facebook.com/D3838608
Signature: t1:3838608:1473439470:1a689f60e72d351f1cf12d9e2cf1d2da9ab96bd9
Summary:
This adds support for the optional ctreemanifest type. If the module is
available and if `fastmanifest.usetree` is enabled (default: False), we'll use
tree manifests before falling back to flat manifests.
Test Plan: Ran the ctreemanifest perf suite and the fastmanifest test suite
Reviewers: #fastmanifest, ttung
Reviewed By: ttung
Subscribers: ttung, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3754309
Signature: t1:3754309:1471929802:3ccba8ce6160b7a6d20b599626efd5876bd969f6
Summary:
This assert is meant to check that at least one source was provided to the
hybridmanifest. The old version was broken though.
The test requires an update because it attempts to construct a hybridmanifest
with no source.
Test Plan: Ran the tests
Reviewers: #fastmanifest, ttung
Reviewed By: ttung
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3760229
Signature: t1:3760229:1472062113:695085444915cb15a93db7f317580dcffc2f115d
Summary:
This adds the `fastmaniest.usecache` config option (default: True) which lets
users disable the use of the fastmanifest cache without disabling the entire
extension. This will be useful in a future patch where we'll use this same
hybrid manifest infrastructure to allow swapping in tree manifests (even on
repos that don't want fast manifest caches).
Also contains a minior fix to an assert that didn't appear to make sense (I
think it should be asserting that at least one data source is provided).
Test Plan: Ran the ctreemanifest perf suite and the fastmanifest test suite
Reviewers: #fastmanifest, ttung
Reviewed By: ttung
Subscribers: ttung, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3754301
Signature: t1:3754301:1471990417:0cde06f271d10c7f0c839e4f644310a4f525f3f6
Summary:
Because we now prune in parallel to priming the cache, we no longer need the prune at the end. The only scenario where we still need to prune is where we never enter the priming loop, i.e., when the revset is empty.
Depends on D3545267, D3544997
Test Plan: since we don't prune in many circumstances any more, the test output is slighty affected. otherwise, the tests pass.
Reviewers: #fastmanifest, durham
Reviewed By: durham
Subscribers: durham, mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3547219
Signature: t1:3547219:1468373775:e4535e3a66fb362385a23b992749eb7c3fcf7ea1
Summary: vfs.join(None) will return its base. doesn't matter if it's nested or whatever.
Test Plan:
cloned fbjava, then stripped a revision (with evolve off). then added a new revision, and ran `hg incoming`. this resulted in the same stacktrace!
then with the fix run `hg incoming` again. no crash this time.
Reviewers: #fastmanifest, simonfar
Reviewed By: simonfar
Subscribers: mitrandir, simonfar, mjpieters, lcharignon
Differential Revision: https://phabricator.intern.facebook.com/D3598348
Tasks: 12305684
Signature: t1:3598348:1469093045:440cf3d314589f46dd8083901b584d7cfd54a95a
Summary: Sometimes opener.vfs is an _fncachevfs, which has a .vfs member, not a .base member. When you get one of these, go deeper.
Test Plan:
Run testrunner.py flib/intern/sandcastle/vcs/ without this change, see it traceback.
Rerun testrunner.py flib/intern/sandcastle/vcs/ with this change, see it fail in the same way as the current released version.
Reviewers: durham, lcharignon, ttung, rmcelroy
Reviewed By: rmcelroy
Subscribers: rmcelroy, mjpieters, #sourcecontrol
Differential Revision: https://phabricator.intern.facebook.com/D3593097
Tasks: 12305684
Signature: t1:3593097:1469033487:aec4c4a5de6f92a64f4730f03e9780e02f086819
Summary: Rather than have fastmanifestcache as a singleton, we attach it to each opener.
Test Plan: pass existing unit tests.
Reviewers: #fastmanifest, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3553542
Tasks: 12169797
Signature: t1:3553542:1468373683:b0a7fcdf5dbf12046e5b4be9aa0bc5d46e55ce06
Summary: `prune` is essentially `makeroomfor(0, set())`.
Test Plan: pass existing unit tests. output is slightly different, but that's just because we no longer output that debugging line.
Reviewers: #fastmanifest, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3547069
Signature: t1:3547069:1468352305:6185132648f871e26d3cfd449c059523b5eb6d6b
Summary:
It's kind of silly to write the FM cache entry, and then immediately blow away the entry if it doesn't fit in the limit. Instead, just calculate if we have enough space, and raise CacheFullException if we don't.
Depends on D3538944
Test Plan: pass existing tests.
Reviewers: lcharignon, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters, quark
Differential Revision: https://phabricator.intern.facebook.com/D3545446
Signature: t1:3545446:1468283437:7e4167e3b5a13bf3d01fca45a965d7aaa8843166
Summary:
There's two issues with `setwithlimit`:
1) It does not have a docblock, nor an API that fully covers all the possible conditions.
2) It returns None if the cache entry is already present.
This modifies `setwithlimit` to return True if the cache entry ultimately makes it (whether it previously existed or the write was successful), False if the write fails, and raises CacheFullException if the cache is full.
Test Plan: pass existing tests. the already-in-cache case is a race condition that is not practical to reproduce.
Reviewers: lcharignon, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3545407
Signature: t1:3545407:1468283390:64d6cdba7a17cca9bd4465c6e03b9143e4b4cc4b
Summary:
Instead of pruning at the end, we need to prune as we are writing entries.
Depends on D3545060
Test Plan: pass test in D3545060
Reviewers: lcharignon, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3545267
Signature: t1:3545267:1468279898:5332838efa52bb55733d0e6dcf613b27e71e9d3e
Summary:
Previously, depending on the code path, the limit specified would not actually take effect. For instance, if we came in from debugmanifestcache, and attempted to populated the cache, we would use `systemawarecachelimit` when filling the cache, and the fixedsize limit specified by the user when pruning.
With this change, we unify the all the cache limit decisions to `fastmanifestcache`. If the user actually overrides the limit, we set the limit in `fastmanifestcache` and let that make the decisions.
We also change the definitions of limit in `hg debugcachemanifest` to:
1) >0 => it's the limit.
2) =0 => use systemawarecachelimit
3) <0 => no limit!
Test Plan: pass existing unit tests. there's a small change in the test output, because we always evaluate the limit now, plus we remove the test for limit=0, since it means something different now.
Reviewers: lcharignon, durham
Reviewed By: durham
Subscribers: trunkagent, mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3544997
Signature: t1:3544997:1468281604:8f78f00ebf2afd8f3f1fbefbd82316b97cc4b193
Summary:
We need a fastmanifest object in order to size it. Once we know its size, we can make room in the cache.
This slightly affects one of the tests, as we request the manifest text earlier than we previously did.
Depends on D3537904
Test Plan: used in later diff.
Reviewers: lcharignon, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters, quark
Differential Revision: https://phabricator.intern.facebook.com/D3538991
Signature: t1:3538991:1468280648:41c65d91529babe0559eac7b75509481adf2765f
Summary: This allows us to convert a manifest once, and try to repeatedly jam it into the cache (presumably while evicting things).
Test Plan: pass existing tests.
Reviewers: lcharignon, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters, quark
Differential Revision: https://phabricator.intern.facebook.com/D3537904
Signature: t1:3537904:1468268209:6f1746d3cee5071321d830bb91efd591a31a8616
Summary: We don't need this since we loosely lock the cache worker.
Test Plan: pass existing tests.
Reviewers: lcharignon, durham
Reviewed By: durham
Subscribers: mitrandir, mjpieters, quark
Differential Revision: https://phabricator.intern.facebook.com/D3537355
Signature: t1:3537355:1468268248:ebf7de5bf5030b90039823e1111a073d7eacc447
Summary:
Before this patch we were doing two mistakes in the ratio computation.
- We were not recording global cache hit or miss, it was always a hit because
a function is truthy is python
- We were not deduping cache miss for the same manifest multiple times
This patch fixes these two mistakes.
It also changes the logging of cache hit and miss to include the name of the
operation that triggered the cache hit or miss (diff or filesnotin).
Test Plan: Test output changed and commented
Reviewers: durham, ttung
Differential Revision: https://phabricator.intern.facebook.com/D3484658
Summary:
Before this patch we were using config instead of configbool for reading
the debugmetrics config causing "False" to be evaluated as a truthy value
for the config. This patch fixes the issue and sets the config to false for
some of the tests to reduce the noise of the output.
Test Plan: Tests pass
Reviewers: ttung
Differential Revision: https://phabricator.intern.facebook.com/D3524501
Summary:
There's no need for the `setup()` to be called in `reposetup()` any more.
Depends on D3510966
Test Plan:
1) pass existing unit tests.
2) kill chg server instances, then run `USE_CHG= hg diff -c . --debug` repeatedly. (USE_CHG tells my hg wrapper to use chg)
Reviewers: quark, lcharignon
Reviewed By: lcharignon
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3511004
Signature: t1:3511004:1467410145:0761c80da77d5a193fbdb08527fcc4aae9f8065e
Summary: Whenever fastmanifest needs an ui object, we pass it a global proxy object. Whenever someone calls `reposetup(..)`, we update the backing object.
Test Plan:
1) pass existing unit tests
2) enable chg, then run `hg diff -c . --debug`. saw fastmanifest debug output in the console.
Reviewers: lcharignon, quark
Reviewed By: lcharignon
Subscribers: trunkagent, mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3510966
Signature: t1:3510966:1467410140:88fe810d03b9a9aa6818f17eb605a70eee931364
Summary:
There are two parts to add logging to blackbox. Logging events
and changing the configuration of blackbox. This diff does the former.
We don't reuse all the metrics event as:
- blackbox does not support wildcard for events like fastmanifest-*
- they are not meant to be human readable, they are for performance logging
We instead log a few new events:
- What triggered a caching operation?
- What command we run to cache the manifest?
- What revisions are to be cached?
- What revisions are actuallt cached?
- Do we overflow the cache?
This will allow us to understand all the cases where caching didn't trigger.
Like in t11877434.
Test Plan:
Test output does not change
Verified manually by changing the blackbox configuration that events get logged:
CHGDISABLE=1 hg book food --config blackbox.track='command, commandfinish, commandexception, exthook, fsmonitor, pythonhook, fastmanifest' --config extensions.fastmanifest=/home/lcharignon/facebook-hg-rpms/fb-hgext/fastmanifest --config extensions.blackbox=/home/lcharignon/facebook-hg-rpms/hg-crew/hgext/blackbox.py
Reviewers: ttung
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3504158
Tasks: 11877434
Summary: If the key as always 'ratio', then all the ratios are aliased when we send off aggregate stats. That means we end up sending the last entry, rather than each individual entry.
Test Plan:
along with D3504940
run FB_HG_DIAGS= hg diff -c .:
```
"int": {
"builddate": 1467289392,
"cachehitratio": 100,
"consumed": 489,
"diffcachehitratio": -1,
"elapsed": 280,
"errorcode": 0,
"filesnotincachehitratio": 0,
"time": 1467318481
},
```
Reviewers: lcharignon
Reviewed By: lcharignon
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3505464
Tasks: 12019647
Signature: t1:3505464:1467320144:327e0d306b9afa90ed9fc2d704a8ca02a79f3038
Summary:
This diff removes `self.ui = ui` in metricscollector to address stale ui
issues with chg. Also clean up unnecessary sample merges since there is
only one collector per process.
Test Plan:
Run with `fastmanifest.debugmetrics=1` and confirm see lines like:
[FM-METRICS] Begin metrics
[FM-METRICS] kind: cachehitratio, kwargs: [('ratio', 100.0)]
[FM-METRICS] kind: diffcachehitratio, kwargs: [('ratio', -1)]
[FM-METRICS] kind: filesnotincachehitratio, kwargs: [('ratio', 0.0)]
[FM-METRICS] End metrics
Reviewers: #mercurial, ttung, lcharignon
Reviewed By: lcharignon
Subscribers: mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3504442
Signature: t1:3504442:1467311563:5904bd06b210fa8550b2e1f7e85c2133ad7a1a0e
Summary:
chg runs extsetup and uisetup with dummy ui objects. Before this
patch we were relying on ui objects in extsetup and uisetup. After this
patch, we move the logic to reposetup. We do things a little differently
because reposetup can be called multiple times and extsetup and uisetup are
called only once.
This caused one problem for `hg clone` in the test as reposetup is not
called for clone operation. I adapted the test a little bit to cover this case.
Test Plan: Modified tests pass
Reviewers: mitrandir, quark, lcharignon
Reviewed By: lcharignon
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3499176
Signature: t1:3499176:1467244335:a70d718725c1d19ab6b6feb4558a8ffd1a020ab2
Summary:
Since metrics doesn't send a string, let's just stub out the call.
The alternative is to put a '' in every `recordsample()` call.
Test Plan: pass existing tests.
Reviewers: lcharignon
Reviewed By: lcharignon
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3479138
Tasks: 11900487
Signature: t1:3479138:1466787619:910917093cf0361ed181af5597ab702b936f8d1e
Summary: It's actually most relevant to least relevant. Add a docblock.
Test Plan: pass unit tests
Reviewers: lcharignon
Reviewed By: lcharignon
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3478616
Signature: t1:3478616:1466787566:b38eceddf91c3e88341f9173e07244ac33ab6345
Summary:
Since we no longer need to unlock in only one fork, we can use the `with lock() as xxx:` syntax
Depends on D3468830
Test Plan: pass existing unit tests
Reviewers: durham, lcharignon
Reviewed By: lcharignon
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3468832
Tasks: 11683504
Signature: t1:3468832:1466694908:6e1d3beff673df5b70ece250af72ad9be66545a6
Summary:
Introduce a cacher class, which encapsulates all the logic for actually caching manifests. Triggers can either directly invoke cacher, or invoke `hg cachemanifest`, which then calls cacher.
Depends on D3468828
Test Plan: pass existing unit tests
Reviewers: durham, lcharignon
Reviewed By: lcharignon
Subscribers: mitrandir, mjpieters
Differential Revision: https://phabricator.intern.facebook.com/D3468830
Tasks: 11683504
Signature: t1:3468830:1466695022:9a470fcb06badd9e90931cd931a945e8039d71ea