perf: start recording total time after warming The warming might be costly and this should not affect the "time profile" of the actual collection.

perf: run the gc before each run The python garbage collector is a large source of performance troubles, we run it right before the timed section to reduce the change for the gc to add noise to the benchmark.

perf: allow profiling of more than one run By default, we still profile the first run only. However profiling more run help to understand side effect from one run to the other. So we add an option to be able to do so.

profiler: flush after writing the profiler output Otherwise, the profiler output might only partially appears until the next flush of the buffer. Since profiling often happens for long operation, the next flush can be a long time away.

stream-clone: disable gc for the entry listing section for the v2 format This is similar to the change we did for the v3 format in 6e4c8366c5ce. The benchmark bellow show this gives us a notable gains, especially on larger repositories. ### benchmark.name = hg.perf.stream-locked-section # benchmark.name = hg.perf.stream-locked-section # bin-env-vars.hg.flavor = default # bin-env-vars.hg.py-re2-module = default # benchmark.variants.version = v2 ## data-env-vars.name = pypy-2018-08-01-zstd-sparse-revlog 5e931bf8707c: 0.503820 ~~~~~ 1106d1bf695e: 0.470078 (-6.70%, -0.03) ## data-env-vars.name = pypy-2024-03-22-zstd-sparse-revlog 5e931bf8707c: 0.535756 ~~~~~ 1106d1bf695e: 0.490249 (-8.49%, -0.05) ## data-env-vars.name = heptapod-public-2024-03-25-zstd-sparse-revlog 5e931bf8707c: 1.327041 ~~~~~ 1106d1bf695e: 1.174636 (-11.48%, -0.15) ## data-env-vars.name = netbeans-2018-08-01-zstd-sparse-revlog 5e931bf8707c: 2.439158 ~~~~~ 1106d1bf695e: 2.220515 (-8.96%, -0.22) ## data-env-vars.name = netbeans-2019-11-07-zstd-sparse-revlog 5e931bf8707c: 2.630794 ~~~~~ 1106d1bf695e: 2.261473 (-14.04%, -0.37) ## data-env-vars.name = mozilla-central-2018-08-01-zstd-sparse-revlog 5e931bf8707c: 5.769002 ~~~~~ 1106d1bf695e: 5.062000 (-12.26%, -0.71) ## data-env-vars.name = mozilla-try-2019-02-18-zstd-sparse-revlog 5e931bf8707c: 13.351750 ~~~~~ 1106d1bf695e: 12.346655 (-7.53%, -1.01) ## data-env-vars.name = mozilla-central-2024-03-22-zstd-sparse-revlog 5e931bf8707c: 10.772939 ~~~~~ 1106d1bf695e: 9.495407 (-11.86%, -1.28) ## data-env-vars.name = mozilla-unified-2024-03-22-zstd-sparse-revlog 5e931bf8707c: 10.864297 ~~~~~ 1106d1bf695e: 9.475597 (-12.78%, -1.39) ## data-env-vars.name = mozilla-try-2023-03-22-zstd-sparse-revlog 5e931bf8707c: 17.448335 ~~~~~ 1106d1bf695e: 16.027474 (-8.14%, -1.42)

phases: rework the logic of _pushdiscoveryphase to bound complexity This rework the various graph traversal in _pushdiscoveryphase to keep the complexity in check. This is done though a couple of things: - first, limiting the space we have to explore, for example, if we are not in publishing push, we don't need to consider remote draft roots that are also draft locally, as there is nothing to be moved there. - avoid unbounded descendant computation, and use the faster "rev between" computation. This provide a massive boost to performance when exchanging with repository with a massive amount of draft, like mozilla-try: ### data-env-vars.name = mozilla-try-2023-03-22-zstd-sparse-revlog # benchmark.name = hg.command.push # bin-env-vars.hg.flavor = default # bin-env-vars.hg.py-re2-module = default # benchmark.variants.explicit-rev = all-out-heads # benchmark.variants.issue6528 = disabled # benchmark.variants.protocol = ssh # benchmark.variants.reuse-external-delta-parent = default ## benchmark.variants.revs = any-1-extra-rev before: 20.346590 seconds after: 11.232059 seconds (-38.15%, -7.48 seconds) ## benchmark.variants.revs = any-100-extra-rev before: 24.752051 seconds after: 15.367412 seconds (-37.91%, -9.38 seconds) After this changes, the push operation is still quite too slow. Some of this can be attributed to general phases slowness (reading all the roots from disk for example) and other know slowness (not using persistent-nodemap, branchmap, tags, etc. We are also working on them, but with this series, phase discovery during push no longer showing up in profile and this is a pretty nice and bit low-hanging fruit out of the way. ### (same case as the above) # benchmark.variants.revs = any-1-extra-rev pre-%ln-change: 44.235070 this-changeset: 11.232059 seconds (-74.61%, -33.00 seconds) # benchmark.variants.revs = any-100-extra-rev pre-%ln-change: 49.234697 this-changeset: 15.367412 seconds (-68.79%, -33.87 seconds) Note that with this change, the `hg push` performance is now much closer to the `hg pull` performance, even it still lagging behind a bit. (and the overall performance are still too slow). ### data-env-vars.name = mozilla-try-2023-03-22-ds2-pnm # benchmark.variants.explicit-rev = all-out-heads # benchmark.variants.issue6528 = disabled # benchmark.variants.protocol = ssh # benchmark.variants.pulled-delta-reuse-policy = default # bin-env-vars.hg.flavor = rust ## benchmark.variants.revs = any-1-extra-rev hg.command.pull: 6.517450 hg.command.push: 11.219888 ## benchmark.variants.revs = any-100-extra-rev hg.command.pull: 10.160991 hg.command.push: 14.251107 ### data-env-vars.name = mozilla-try-2023-03-22-zstd-sparse-revlog # bin-env-vars.hg.py-re2-module = default # benchmark.variants.explicit-rev = all-out-heads # benchmark.variants.issue6528 = disabled # benchmark.variants.protocol = ssh # benchmark.variants.pulled-delta-reuse-policy = default ## bin-env-vars.hg.flavor = default ## benchmark.variants.revs = any-1-extra-rev hg.command.pull: 8.577772 hg.command.push: 11.232059 ## bin-env-vars.hg.flavor = default ## benchmark.variants.revs = any-100-extra-rev hg.command.pull: 13.152976 hg.command.push: 15.367412 ## bin-env-vars.hg.flavor = rust ## benchmark.variants.revs = any-1-extra-rev hg.command.pull: 8.731982 hg.command.push: 11.178751 ## bin-env-vars.hg.flavor = rust ## benchmark.variants.revs = any-100-extra-rev hg.command.pull: 13.184236 hg.command.push: 15.620843

phases: introduce a performant efficient way to access revision in a set This will be useful in the next changesets.

phases: use revision number in `_pushdiscoveryphase` We now reach our target checkpoint in terms of rev-num conversion. The `_pushdiscoveryphase` function is now performing graph computation based on revision number only. Avoiding repeated conversion from node-id to rev-num. See previous changeset updated `new_heads` for rationnal. Again, time saved in the 100 milliseconds order of magnitude for the mozilla-try benchmark I have been using. However, wow that the logic is done using revision number, we can look into having better logic in the next changesets, which will provide a much bigger speedup.

phases: move RemotePhasesSummary to revision number This continue our quest to align more logic on revision number instead of node-ids. The motivation is similar to the change to `new_heads` and `analyze_remote_phases` a few changeset earlier. Again, we take this as an opportunity to rename the class, and the attribute to the new naming scheme. This will highlight the need for code update for any code using it an expecting node-ids. Many of the rev-num → node-id conversion we had to introduce in the previous changesets can now be removed. More will be removed in the future as we continue to align code toward rev-num usage. time saved in the 100 milliseconds order of magnitude for the mozilla-try benchmark I have been using.

phases: stop using `repo.set` in `remotephasessummary` The `repository.set` create changectx on the fly, an expensive operation. Using `repo.revs` and a direct rev-num → node-id translation will be significantly faster. This is especially true as we prepare ourself to no longer do the rev-num → node-id transalation there. The speedup is a bit lost in the overall noisyness of the slow phase discovery algorithm, but it save a small amount of time in my benchmark.

phases: use revision number in analyze_remote_phases Same logic as the previous change to `new_heads`, see rationnal there. This avoids a small number of `nodes -> revs` conversion speeding thing up in the 100 milliseconds order of magnitude for the worses cases. However, the rest of the logic is noisy enough that it hardly matters for now.

phases: use revision number in new_heads All graph operations will be done using revision numbers, so passing nodes only means they will eventually get converted to revision numbers internally. As part of an effort to align the code on using revision number we make the `phases.newheads` function operated on revision number, taking them as input and using them in returns, instead of the node-id it used to consume and produce. This is part of multiple changesets effort to translate more part of the logic, but is done step by step to facilitate the identification of issue that might arise in mercurial core and extensions. To make the change simpler to handle for third party extensions, we also rename the function, using a more modern form. This will help detecting the different between the node-id version and the rev-num version. I also take this as an opportunity to add some comment about possible performance improvement for the future. They don't matter too much now, but they are worse exploring in a while.

phases: convert remote phase root to node while reading them This is currently a bit silly as we will convert them back to node right after, but that is an intermediate step before doing more disruptive changes.

phases: more compact error handling in analyzeremotephases using an intermediate variable result in more readable code, so let us use it.

push: rework the computation of fallbackheads to be correct The previous computation tried to be smart but ended up being wrong. This was caught by phase movement test while reworking the phase discovery logic to be faster. The previous logic was failing to catch case where the pushed set was not based on a common heads (i.e. when the discovery seemed to have "over discovered" content, outside the pushed set) In the following graph, `e` is a common head and we `hg push -r f`. We need to detect `c` as a fallback heads and we previous failed to do so:: e | d f |/ c | b | a The performance impact of the change seems minimal. On the most impacted repository at hand (mozilla-try), the slowdown seems mostly mixed in the overall noise `hg push` but seems to be in the hundred of milliseconds order of magnitude. When using rust, we seems to be a bit faster, probably because we leverage more accelaratd internals. I added a couple of performance related common for further investigation later on.

revset: stop serializing node when using "%ln" Turning hundred of thousand of node from node to hex and back can be slow… what about we stop doing it? In many case were we are using node id we should be using revision id. However this is not a good reason to have a stupidly slow implementation of "%ln". This caught my attention again because the phase discovery during push make an extensive use of "%ln" or huge set. In absolute, that phase discovery probably should use "%ld" and need to improves its algorithmic complexity, but improving "%ln" seems simple and long overdue. This greatly speeds up `hg push` on repository with many drafts. Here are some relevant poulpe benchmarks: ### data-env-vars.name = mozilla-try-2023-03-22-zstd-sparse-revlog # benchmark.name = hg.command.push # bin-env-vars.hg.flavor = default # bin-env-vars.hg.py-re2-module = default # benchmark.variants.explicit-rev = all-out-heads # benchmark.variants.issue6528 = disabled # benchmark.variants.protocol = ssh # benchmark.variants.reuse-external-delta-parent = default ## benchmark.variants.revs = any-1-extra-rev before: 44.235070 after: 20.416329 (-53.85%, -23.82) ## benchmark.variants.revs = any-100-extra-rev before: 49.234697 after: 26.519829 (-46.14%, -22.71) ### benchmark.name = hg.command.bundle # bin-env-vars.hg.flavor = default # bin-env-vars.hg.py-re2-module = default # benchmark.variants.revs = all # benchmark.variants.type = none-streamv2 ## data-env-vars.name = heptapod-public-2024-03-25-zstd-sparse-revlog before: 10.138396 after: 7.750458 (-23.55%, -2.39) ## data-env-vars.name = mercurial-public-2024-03-22-zstd-sparse-revlog before: 1.263859 after: 0.700229 (-44.60%, -0.56) ## data-env-vars.name = mozilla-try-2023-03-22-zstd-sparse-revlog before: 399.484481 after: 346.5089 (-13.26%, -52.98) ## data-env-vars.name = pypy-2024-03-22-zstd-sparse-revlog before: 4.540080 after: 3.401700 (-25.07%, -1.14) ## data-env-vars.name = tryton-public-2024-03-22-zstd-sparse-revlog before: 2.975765 after: 1.870798 (-37.13%, -1.10)

bundlespec: drop unused _bundlespecvariants dictionary Why do we have a `_bundlespecvariants`?

bundlespec: type the _bundlespeccontentopts dictionary If only we had a tool to detect the kind of stupid error we just fixed… ho wait.

bundlespec: fix the "streamv2" and "streamv3-exp" variant In c4aab3661f25, we broken this feature by adding unicode instead of bytes to the dictionary. On the other hand, this feature was never tested, so augment the tests to tests this.

perf: create the temporary target next to the source in stream-consume See inline comment for rational.

bundlespec: rationalize the way we specify stream bundle version Instead of having weird dedicated option for each version (v2, v3, etc) we reuse the same "stream" parameters. This is consistent with the ability to request a stream clone using "none-v2;stream=v2". This changeset introduce no user visible change, this is pure internal cleaning.

bundle: do no check the changegroup version if no changegroup is included We don't need to check the compatibility of something we will not use. In practice this was getting in the was of `streamv2` bundles on a narrow repository as the 'cg.version=02' value was rejected by this checks.

perf-stream-consume: use the source repository config when applying This might contains critical configuration for the benchmark, like enabling of extensions like narrow.

unbundle: move most of the logic on cmdutil to help debug::unbundle reuse This make sure `hg debug::unbundle` focus on the core logic.

postincoming: move to cmdutil This looks like a good place for it to live.

postincoming: avoid computing branchhead if no report will be posted This otherwise defeat some of the branch v3 optimization.

streamclone: stop listing files for entries that have no volatile files This will save a lot of python related time. This significantly boost performance. The following number comes from a large private repository using perf::stream-locked-section: base-line: 35.04 seconds prev-change: 24.51 seconds (-30%) prev-change: 20.88 seconds (-40%) prev-change: 14.22 seconds (-60%) this-change: 11.58 seconds (-67% from baseline; -18% from prev)

stream-clone: disable gc for the initial section for the v3 format The number of small container created turn Python in a gc-frenzy that seriously impact performance. This significantly boost performance. The following number comes from a large private repository using perf::stream-locked-section: base-line: 35.04 seconds prev-change: 24.51 seconds (-30%) prev-change: 20.88 seconds (-40%) this-change: 14.22 seconds (-60% from baseline; -31% from prev)

stream-clone: disable gc for `_entries_walk` duration The number of small container created turn Python in a gc-frenzy that seriously impact performance. This significantly boost performance. The following number comes from a large private repository using perf::stream-locked-section: base-line: 35.04 seconds prev-change: 24.51 seconds (-30%) this-change: 20.88 seconds (-40% from baseline; -15% from previous changes)

nocg: make the utility work are both a decorator and context manager In some case, the context manager version will be simpler.

stream-clone: stop getting the file size of all file in v3 The point of v3 is to do less work in the locked section. It was currently not the case. This significantly boost performance. The following number comes from a large private repository using perf::stream-locked-section: base-line: 35.03 seconds this-change: 24.50 seconds (-30%)

stream: in v3, skip the "size" fast path if the entries as some unknown size We are about to prefetch size during the lock less in the v3 case. So we need to avoid trying to use that prefetched size when it is not available. See next changeset for the motivation.

perf-stream-locked-section: advertise the right version key in the help As the v3 format is still experimental, its key is "v3-exp". The help text was not pointing that out. (we also fix `perf::stream-generate` in the process)

perf-stream-locked-section: fix the call to the v3 generator That generator simply return chunks so we should not assign the return to a tuple.

perf-stream-locked-section: actually use v1 generation when requested We were fetching a v1 generator but actually using the v2 function…

branching: merge stable into default

branching: merge stable into default

branchcache: allow to detect "pure topological case" for branchmap We don't rum this detection every time we run the branchcache, that would be costly. However we now do it when running `hg debugupdatecache`. This will help existing repository to benefit from the fastpath when possible.

branchcache: add a "pure topological head" fast path In a narrow but actually quick common case, all topological heads are all on the same branch and all open. In this case, computing the branch map is very simple. We can quickly detect situation where this situation will not change. So we update the V3 format to be able to express this situation and upgrade the update code to detect we remains in that mode. The branch cache is populated with the actual value when the branch map is accessed, but the update_disk method can do the update without needing to populate it.

branchcache: move the processing of the new data in a dedicated method In a future changeset, this will allow the V3 of the branch cache to use a fast path when possible.

branchcache: gather newly closed head in a dedicated set This is part of a series to more clearly split the update in two step. This will allow us to introduce a fast path during update in a future changeset.

branchcache: gather new obsolete revision in a set This is part of a series to more clearly split the update in two step. This will allow us to introduce a fast path during update in a future changeset.

branchcache: filter obsolete revisions sooner Since we won't do anything with the obsolete revisions, we can just ignore them sooner.

branchcache: skip entries that are topological heads in the on disk file In the majority of cases, topological heads are also branch heads. We have efficient way to get the topological heads and efficient way to retrieve their branch information. So there is little value in putting them in the branch cache file explicitly. On the contrary, writing them explicitly tend to create very large cache file that are inefficient to read and update. So the branch cache v3 format is no longer including them. This changeset focus on the format aspect and have no focus on the performance aspect. We will cover that later.

branchcache: simplify the branch rev cache test We don't need that many content dump and this gets in the way in change in access pattern (e.g. accessing revision in a different order change the order of branches in the "names" file). So we simplify this test in advance.

branchcache: store filtered hash and obsolete hash independently for V3 This will avoid the bug covered in tests/test-branches-obsolete.t when we stop storing all heads explicitly in V3.

branchcache: show the cache file content in test-branches-obsoletes.t This help to track the changes in format between v2 and v3.

branchcache: rework the `filteredhash` logic to be more generic We now have a more flexible `key_hashes` tuple. We duplicated various logic in the V2 and V3 version of the cache as the goal is to start changing the logic for V3 in the next few changesets.

filteredhash: rename the filteredhash function The new name is less ambiguous, as we are about to introduce an alternative function it seems like a good idea to have clearer name to distinct the two.

filteredhash: split the computation of revision sets The branch2's filteredhash combines the filtered revisions and the obsolete ones, this will creates issue for implicit reference to heads we want to introduce for the v3 of the branch cache format. So we isolate this logic for alternative use.

filteredhash: move the hashing in its own function This will help us to reuse this logic in variants of the hashes used for branch cache validation.

branchcache: cleanup the final key generation after update A lot of duplicated work seemed to be done, as we already update the tiprev and tipnode when needed right before. So we simplify that part to focus on the filtered hash. See inline comment for details.

branchcache: add more test for the logic around obsolescence and branch heads While working on branch-cache-v3, we noticed some ambiguity in the filtered+obsolete hash. However this was only caught by a rebase test by chance. It seems important to explicitly tests these cases.

branchcache-v3: use more explicit header line The key-value approach is clearer and gives more rooms to have the format evolve in a clear way. It also provides extension (like topic) simpler way to extend the validation scheme. This is just a small evolution, the V3 format is still a work in progress.

branchcache-v3: introduce a v3 format For now the format is the very same, however we will start changing it in future changesets.

branchcache: use an explicit class for the v2 version This prepare the introduction of an experimental v3 format version. In the process, we move the description of the format in that new class.

branchcache: add some blank line in a test This helps each section to stand out.

branching: merge stable into default

branching: merge stable into default

branchcache: use update_disk to refresh 'served' and 'served.hidden' The `update_disk` method is dedicated to this kind of usecase. Now that the writting patterns are more consistent, we can use it to warm these two important cache. I am dropping the first comment about "refreshing all the others" because it is false. If a branchmap already exist for "served", none of the subset will be updated.

branchcache: explictly update disk state only if no transaction exist If a transaction exist the `write_dirty` call will eventually be done and the state will be synched on disk. It is better to no interfer with that.

branchcache: do not use `__getitem__` in updatecache The `update_disk` method uses `updatecache` and the point of `update_disk` is to be able to do alternative processing to the one we do in `__getitem__`. So we calling `__getitem__` in `updatecache` defeat this purpose. Instead we do the equivalent explicitly to preserve the spirit of `update_disk` (that we will actually put to use soon, I promise)

branchcache: explicitly track inheritence "state" We move from a binary "dirty" flag to a three value "state": "clean", "inherited", "dirty". The "inherited" means that the branch cache is not only "clean", but it is a duplicate of its parent filter. If a branch cache is "inherited", we can non only skip writing its value on disk, but it is a good idea to delete any stale value on disk, as those will just waste time (and possibly induce bug) in the future. We only do this in the update related to transaction or explicit cache update (e.g `hg debugupdatecache`). Deleting the file when we simply detected a stall cache during a read only operation seems more dangerous. We rename `copy` to `inherit_for` to clarify we associate a stronger semantic to the operation.

branchcache: stop writing more branchcache file on disk than needed Before this change, we were unconditionally writing a branchmap file for the filter level passed to `update_disk`. This is actually counter productive if no update were needed for this filter level. In many case, the branch cache for a filter level is identical to its parent "subset" and it is better to simply keep the subset update and reuse it every time instead of having to do identical work for similar subset. So we change the `update_disk` method to only write a file when that filter level differ from its parent. This removes many cases where identical files were written, requiring multiple boring update in the test suite. The only notable changes is the change to `test-strip-branch-cache.t`, this case was checking a scenario that no longer reproduce the bug as writing less branchmap file result in less stalled cache on disk. Strictly speaking, we could create a more convoluted scenario that create a similar issue. However the next changeset would also cover that scenario so we directly updated that test case to a "no longer buggy" state.