Summary:
Like it says in the title. We'd like to ask for the exact size that was
configured, because this way we can set the chunk size to the LFS threshold and
it avoids overlapping any file chunks server side.
Reviewed By: DurhamG
Differential Revision: D27824418
fbshipit-source-id: 43f40eb87080ec58e813ba1f1dda5b6a5e9f98ee
Summary:
This replicates behaviour of Python code - if unknown file content matches content of the file to be check out, do not abort checkout
This is useful for resuming interrupted checkouts / clones
Reviewed By: DurhamG
Differential Revision: D27799147
fbshipit-source-id: 7d2582583525da18fba08dfcd8bced2b619304de
Summary: Currently native checkout aborts on unknown files even with --clean flag. It should not abort with --clean
Reviewed By: DurhamG
Differential Revision: D27779554
fbshipit-source-id: 2badc84c10eab28d2b1fc8840142ef883ac48c26
Summary: It's been showing up while building mononoke. Let's fix it
Reviewed By: sfilipco
Differential Revision: D27789928
fbshipit-source-id: a15912f66b9ad3370545aed88405dbeb800e63de
Summary: This seems to have broken the EdenFS HgPrefetch test.
Reviewed By: xavierd
Differential Revision: D27795192
fbshipit-source-id: 80a748036961aa6a5750182bf65637fb76825341
Summary: This will show proper checkout progress when using native checkout
Reviewed By: quark-zju
Differential Revision: D27775423
fbshipit-source-id: 79f2afa02bd1fab7d5f747da1c714d4d1126ce7c
Summary:
EdenAPI makes heavy use of streaming HTTP responses consisting of a series of serialized CBOR values. In order to process the data in a streaming manner, we use the `CborStream` combinator, which attempts to deserialize the CBOR values as they are received.
`CborStream` hits a pathological case when it receives a very large CBOR value. Previously, it would always buffer the input stream into 1 MB chunks, and attempt to deserialize whenever a new chunk was received. In the case of downloading values that are >1GB in size, this meant potentially thousands of wasted deserialization attempts. In practice, this meant that EdenAPI would hang when receiving the content of large files.
To address this problem, this diff adds a simple heuristic: If a partial CBOR value exceeds the current buffer size, double the size threshold before attempting to deserialize again. This reduces the pathological case from `O(n^2)` to `O(log(n))` (with some caveats, described in the comment in the code).
Reviewed By: krallin
Differential Revision: D27759698
fbshipit-source-id: 67882c31ce95a934b96c61f1c72bd97cad942d2e
Summary:
Previously we'd skip dynamicconfigs when there wasn't a repo available.
Now that dynamicconfig can represent the NoRepo state, let's load dynamicconfigs
in that situation.
This also supports the case where there is no user name.
Reviewed By: kulshrax
Differential Revision: D26801059
fbshipit-source-id: 377cfffb6695a7fbe31303868a88862259ebf8c4
Summary: Add a new `edenapi.maxrequests` config option to allow controlling the number of parallel in-flight requests. This can be used to bound resource usage of the client when requesting large amounts of data.
Reviewed By: sfilipco
Differential Revision: D27724817
fbshipit-source-id: 8d607efa83d8b0b94074d1a6e06f6c536cc0c791
Summary: Add a method to allow setting `CURLMOPT_MAX_TOTAL_CONNECTIONS`, which limits the number of concurrent requests within a curl multi session. If the number of requests in the session exceeds this number, they will be queued and sent once earlier requests have completed.
Reviewed By: sfilipco
Differential Revision: D27724818
fbshipit-source-id: 436384aed9d6d29f426e5e45aebb7a72c24ba667
Summary:
Without this, `make local` will build `hostcaps` without fb-specific logic and
cause wrong configs being used. `hg up master` will error out like:
File "treemanifest/__init__.py", line 690, in _httpgetdesignatednodes
self.edenapi.trees(dpack, self.name, keys)
RustError: Server reported an error (403 Forbidden)
Reviewed By: quark-zju
Differential Revision: D27759821
fbshipit-source-id: d42895f44bc53003f2578b65640ebe4ee05d52e6
Summary:
Right now, if prefetch fails, we just give the client back an error saying
"content not found".
This isn't super helpful, because usually the reason the content is not found
is because we cannot talk to the server that has the content, so showing the
user why we cannot talk to said server is more useful.
I'd like to ship this gradually, so I also added a config flag to turn it off.
Initially I'll have the flag set, but I did default it to not-set in the code
so that our tests run with the desired configuration.
Note: I initially contemplated adding logging for this here, but after
discussion with xavierd it looks like just failing instead of eating the error
is probably a better approach (and it's much simpler). This is also consistent
with what EdenAPI does.
Reviewed By: mzr
Differential Revision: D27761572
fbshipit-source-id: 3506d9c97a00e3f076bd346883e07f49194b0b06
Summary:
Right now, if the server ever tells us a file is missing, we fail the entire
batch download. This is a bit unfortunate because other objects could still be
downloaded, but also because we lose the distinction between "server crashed"
and "server told us the data we want does not exist".
Besides, it's actually a bit unnecessary. Indeed, this fails, we just ignore
the error anyway up the stack, so it never actually gets accessed.
Reviewed By: mzr
Differential Revision: D27761574
fbshipit-source-id: cb4fb0526a3bf19c04ecb81c05d44d4d8afb81ad
Summary: We can just return the actual error here now.
Reviewed By: sfilipco
Differential Revision: D27761573
fbshipit-source-id: 0866f976b4ed434deffd96be6820ad05d27b7b93
Summary:
If a operation can't proceed because we are in an interrupted update state,
indicate in the hint that `hg update` needs a destination.
Reviewed By: sfilipco
Differential Revision: D27764182
fbshipit-source-id: f0734a4929e34833c4bf84e148db04d57b779246
Summary:
NOTE: The revisionstore LFS tests talk to prod Mononoke LFS, so the test here
will fail until that is deployed.
Like it says in the title, this adds support for downloading content in chunks
instead of all in one go. The underlying goal here is to let us do better load
balancing server side. This earlier diff in this stack explains that in more
detail: D27188610 (820b538001).
Reviewed By: quark-zju
Differential Revision: D27191594
fbshipit-source-id: 19b1be9bddc6b32b1fabfe4ab5738dfcf71b1a85
Summary: Introduce `FetchKey` and `FetchValue` traits to simplify repeated trait bounds in many `ReadStore` implementations. We also newly require `Clone` for both keys and values, which was already required by the fallback combinator.
Reviewed By: DurhamG
Differential Revision: D27652499
fbshipit-source-id: 6a3d5eb18a904b982fdb9946b80fcc9025d391ea
Summary:
Extend debugscmstore command to fetch arbitrary files / trees by key.
Replace debugpyscmstore with a python fallback for debugscmstore (allowing you to test with the store as it is constructed for Python, with legacy fallback).
Refactor some functionality so it is shared between the rust and python versions of debugscmstore.
Currently the output is pretty ugly. It uses the `{:#?}` format for everything. In the next change, I propose modifying the `Debug` implementation for `minibytes::Bytes` to use ascii-escaped bytestrings rather than the default slice formatter to make things much nicer.
This new `debugscmstore` functionality should be useful in integration tests for testing scmstore under different repo configurations, and for test harnesses and performance testing (fetch a specific set of things easily, simulate delays in the key stream by delaying the input pipe, etc).
Reviewed By: andll
Differential Revision: D27351321
fbshipit-source-id: 8650480e3f5b045b279472643570309c48d7fe6b
Summary: Like `FileScmStoreBuilder`, but for trees. As LFS is not used for trees, `TreeScmStoreBuilder` defaults to `ExtStoredPolicy::Use` (pass along anything you find without LFS-specific checks).
Reviewed By: DurhamG
Differential Revision: D27641290
fbshipit-source-id: 637340a23cef058e7e37a41ae7f5b4fcc9481190
Summary: Introduce a new `FileScmStoreBuilder` structured much like `ContentStoreBuilder`, but supporting the features needed for the intermingling of `contentstore` and `filescmstore` construction (shared indexedlog, scmstore fallback to contentstore).
Reviewed By: DurhamG
Differential Revision: D27640702
fbshipit-source-id: e9771e6f61d80698a9dd761a0db66407b565c010
Summary: The previous change here wasn't sufficient. We need to wrap the handle fd in a Handle now as well.
Reviewed By: quark-zju, sfilipco
Differential Revision: D27753458
fbshipit-source-id: bd8ebbdcdc9acb411362795263b1419360f15e1a
Summary: This test fails without other diffs in stack because previously native checkout was overwriting untracked files
Reviewed By: DurhamG
Differential Revision: D27667151
fbshipit-source-id: 9b3aea37ba5c2d07fe4fc975dd40b4d7bea9d810
Summary: These tests were broken in D27710099 (876f812e4b), but they show as passing unless run in a particular environment, so it went unnoticed. This change reverts the tests to use the pre- D27710099 (876f812e4b) behavior, which should unbreak them until they can be updated correctly.
Reviewed By: quark-zju
Differential Revision: D27756348
fbshipit-source-id: cfa6c12871b6ac0d22b8c70400e72b3ec5dd83a3
Summary:
The `add_heads_and_flush` method might add new nodes in the master group,
and it should update `overlay_map_next_id` accordingly. Without it, it
might error out like:
RustError: ProgrammingError: Server returned x~n (x = 9ebc9ebc49f1819767b40f9ceb22c37547a10c37 8459584, n = 1411).
But x exceeds the head in the local master group {}. This is not expected and indicates some logic error on the server side.
Full error: P387088806
Reviewed By: sfilipco
Differential Revision: D27637278
fbshipit-source-id: b45370db0561dec52cd513a12e2fd0110f18e0e5
Summary:
The filternodes is an API that can batch hasnode checks. It is more efficient
if the commit hashes are lazy.
Reviewed By: sfilipco
Differential Revision: D27636338
fbshipit-source-id: 4eb2dcd20b939faa38611b82e32ed97cf0c8f175
Summary:
The filternodes is an API that can batch hasnode checks. It is more efficient
if the commit hashes are lazy.
Reviewed By: sfilipco
Differential Revision: D27636341
fbshipit-source-id: 69cd708a46c719624d654c53de3d92968392d572
Summary:
If a vertex was resolved via remote protocol, respect it and
avoid requesting the same thing twice.
Reviewed By: sfilipco
Differential Revision: D27636340
fbshipit-source-id: 43cf86010745a85cf622c67be4261ea47a33c802
Summary:
Many places use the `[n for n in nodes if hasnode(n)]` pattern, which
can be slow with a lazy graph due to N+1 problem. Add an API so the
Python world can use a more efficient way for the same query.
Reviewed By: sfilipco
Differential Revision: D27636339
fbshipit-source-id: 99ccb85b2266aed22f1cf87a5e2528106edb3783
Summary:
That could cause a slow loop testing node.__contains__ remotely.
This changes the behavior subtly - nodes added via addgroup will change `tip`
position regardless of whether the nodes exist. This might be more desirable,
since add or addgroup explicitly adding nodes should probably update the
tip position.
Offending test `test-globalrevs-requires.t` was removed since we have
forked the server-side codebase and do not need to maintain hg server
features here.
Reviewed By: sfilipco
Differential Revision: D27630090
fbshipit-source-id: cf7ecc44bf08ed756f0f1aece6655bf674171249
Summary:
The idset is not fully backed by Rust and do not batch resolve vertexes.
The nameset is backed by Rust NameSet and has proper batch prefetching.
Use nameset if possible but fallback to idset if the backend is not in
Rust (rare, only used by hgsql repos now).
Reviewed By: sfilipco
Differential Revision: D27630092
fbshipit-source-id: cf847dd1a78bd5273a8928ecb6616fe11f2c7026
Summary:
This will be useful to avoid suboptimal code paths that is very slow
with lazy vertexes.
Reviewed By: sfilipco
Differential Revision: D27630089
fbshipit-source-id: 35ee4ba79b551453de78fd22aecccf10bc43b08b
Summary:
While it is in theory "correct" going through the remote resolution even if the
protocol is "local". The overhead turns out to be something. And the tracing
message "resolve .. remotely" can be quite noisy. Let's just skip remote
resolutions early in IdConvert implementations.
Reviewed By: sfilipco
Differential Revision: D27630094
fbshipit-source-id: 7d87079876f040cf8f764f7362021dddba0d4723
Summary:
Currently the "contains vertex" check can trigger excessive
fetches for add_heads (and add_heads_and_flush used by flush).
Add a test to demonstrate the problem.
Reviewed By: sfilipco
Differential Revision: D27630091
fbshipit-source-id: ce3639c2a13226ba5681b4e8696edd7acbcb57f9
Summary:
Otherwise it can cause a lazy dag to think vertexes as "missing", insert
vertexes unnecessarily, and potentially break key graph properties (a
vertex should only have one Id).
Reviewed By: sfilipco
Differential Revision: D27629315
fbshipit-source-id: 1688d13cb94015bbc675613ecf9225556ff48372
Summary:
Also move related functions.
Similar to D27547584 (af3c3b3fd0), this allows `add_heads_and_flush` use `IdConvert`
on the `NameDag`, instead of the `IdMap` to trigger remote fetches properly.
This diff is easier to view with whitespace changes ignored.
Reviewed By: sfilipco
Differential Revision: D27629314
fbshipit-source-id: 8f79223c5d324aabfc5ab9813a9f65400fc533ec
Summary:
See the previous diff for context. Drop Locked and related APIs (prepare_filesystem_sync).
This makes it easier to use operate on a mut NameDag on flush because it does not need
to use separate types (Locked) for writing which has issues of not having the remote protocol
involved.
Reviewed By: sfilipco
Differential Revision: D27629306
fbshipit-source-id: 301445b242321ad5f424741ea91ebf6c075bff1c
Summary:
See the previous diff for context. Drop SyncableIdMap so we are one step
closer to using mut NameDag directly on add_heads, which knows when and how to
do remote fetching properly.
Reviewed By: sfilipco
Differential Revision: D27629310
fbshipit-source-id: 883606e40bb83907dfa6142ddd2c3030de61698f
Summary:
By using SyncableIdDag and SyncableIdMap, it's harder to use extra features
around them (ex. remote fetching). Drop SyncableIdDag so we are one step
closer to using mut NameDag directly on add_heads, which knows when and how to
do remote fetching properly.
Reviewed By: sfilipco
Differential Revision: D27629307
fbshipit-source-id: 8e9a5a4348a42b9751752b82feb3f3d2d7c4ba45
Summary:
The `Parents` trait is used for input of adding (non-lazy) vertexes to the
graph. The API will be used to provide extra hints to optimize network
fetches.
With the current logic, `assign_head` will ask the server to resolve the heads
first, to check if it is already assigned, then to resolve the parents, etc. to
the roots (in the "to assign" set). Ideally the `assign_head` logic can ask
the server to resolve the roots first, and if that's unassigned, then just mark
all descendants of the roots as unassigned, do not send more requests.
Note: the current pull logic has all the hashes ready (hashes are known).
But whether the hashes have Id assigned are unknown. It is more tricky
taking the "lock" and "reload" into consideration - hashes without Ids might
"become" having ids assigned after we obtain the lock to write data to disk.
Practically, `pull` using the current logic would take a very long time because
it tries to resolve things remotely for every "to assign" commits.
Reviewed By: sfilipco
Differential Revision: D27629317
fbshipit-source-id: e02f54f43ef65228ce6e3a8a8723dd9ae0a07008
Summary: This just simplifies the test code a bit.
Reviewed By: sfilipco
Differential Revision: D27629308
fbshipit-source-id: 04eac5cd045c71123e7fc410af74609dbadb8fb7
Summary: This avoids triggering remote lookups if an unknown name was looked up multiple times.
Reviewed By: sfilipco
Differential Revision: D27629316
fbshipit-source-id: 64c1ce5e872a5ce4f14c650a946ae8396f4cc74c
Summary:
When translating RequestNameToLocation to ResponseIdNamePair. If "heads" are
known, but some "names" aren't. Do not treat it as an error. This will be
used by the client-side to properly handle the "contains" check.
Reviewed By: sfilipco
Differential Revision: D27629309
fbshipit-source-id: 206ec5df956b33f4e816ab8d6a67ce776fd7bd74
Summary: This will make it easier to test client / server dags in upcoming changes.
Reviewed By: sfilipco
Differential Revision: D27629318
fbshipit-source-id: e3137654613aa3208a8f2e4b9f4ddfe73871f2c5
Summary: This will be used in upcoming changes. It just delegates to the Arc inner.
Reviewed By: sfilipco
Differential Revision: D27629313
fbshipit-source-id: ba6cd7cac2b9f5c1a2898c439c53768995a2dc42
Summary: This will be used by upcoming changes.
Reviewed By: sfilipco
Differential Revision: D27629312
fbshipit-source-id: 6c56e73caf4e1a398ac3a8b4294bd9f380af3764
Summary: This will be used by upcoming changes.
Reviewed By: sfilipco
Differential Revision: D27629319
fbshipit-source-id: d19e490268561f3154642e5bb1e415da4c5d03ee
