Anton Shestakov <av6@dwimlabs.net> [Mon, 12 Dec 2022 17:49:48 +0400] rev 49792
tests: document what the other currently enabled pylint check ensures
Anton Shestakov <av6@dwimlabs.net> [Mon, 12 Dec 2022 17:42:30 +0400] rev 49791
hghave: detect newer pylint
Older versions (e.g. 2.7.2) say: "Usage: pylint [options]"
Newer versions (e.g. 2.15.5) say: "usage: pylint [options]"
Pierre-Yves David <pierre-yves.david@octobus.net> [Sat, 12 Nov 2022 00:18:41 +0100] rev 49790
emitrevision: consider ancestors revision to emit as available base
This should make more delta base valid. This notably affects:
* case where we skipped some parent with empty delta to directly delta against
an ancestors
* case where an intermediate snapshots is stored.
This change means we could sent largish intermediate snapshots over the wire.
However this is actually a sub goal here. Sending snapshots over the wire means
the client have a high odd of simply storing the pre-computed delta instead of
doing a lengthy process that will… end up doing the same intermediate snapshot.
In addition the overall size of snapshot (or any level) is "only" some or the
overall delta size. (0.17% for my mercurial clone, 20% for my clone of Mozilla
try). So Sending them other the wire is unlikely to change large impact on the
bandwidth used.
If we decide that minimising the bandwidth is an explicit goal, we should
introduce new logic to filter-out snapshot as delta. The current code has no
notion explicite of snapshot so far, they just tended to fall into the wobbly
filtering options.
In some cases, this patch can yield large improvement to the bundling time:
### data-env-vars.name = mozilla-try-2019-02-18-zstd-sparse-revlog
# benchmark.name = perf-bundle
# benchmark.variants.revs = last-100000
before: 68.787066 seconds
after: 47.552677 seconds (-30.87%)
That translate to large improvement to the pull time :
### data-env-vars.name = mozilla-try-2019-02-18-zstd-sparse-revlog
# benchmark.name = pull
# benchmark.variants.issue6528 = disabled
# benchmark.variants.revs = last-100000
before: 142.186625 seconds
after: 75.897745 seconds (-46.62%)
No significant negative impact have been observed.
Pierre-Yves David <pierre-yves.david@octobus.net> [Wed, 09 Nov 2022 13:54:15 -0500] rev 49789
sqlitestore: add an `ancestors` method
We will need it during bundling.
The implementation mirror the one in revlog.
Pierre-Yves David <pierre-yves.david@octobus.net> [Thu, 24 Nov 2022 04:04:19 +0100] rev 49788
emitrevision: if we need to compute a delta on the fly, try p1 or p2 first
Falling back to `prev` does not yield any real value on modern storage and
result in pathological changes to be created on the other side. Doing a delta
against a parent will likely be smaller (helping the network) and will be safer
to apply on the client (helping future pulls by Triggering intermediate
snapshop where they will be needed by later deltas).
Pierre-Yves David <pierre-yves.david@octobus.net> [Mon, 28 Nov 2022 16:27:23 +0100] rev 49787
emitrevision: simplify the fallback to computed delta
Not using the stored delta, or having a full snapshot on disk behave the same
ways, so lets use the same code path for that, this is simpler, and it update
will be simpler.
Pierre-Yves David <pierre-yves.david@octobus.net> [Mon, 28 Nov 2022 15:59:52 +0100] rev 49786
emitrevision: also check the parents in the availability closure
One of the point of having a closure is to gather the logic in it. So we gather
the logic.
The `parents[:]` part is a bit ugly but will be replaced by better code soon
anyway.
Pierre-Yves David <pierre-yves.david@octobus.net> [Mon, 28 Nov 2022 15:48:51 +0100] rev 49785
emitrevision: add a small closure to check if a base is usable
We will make more use of this and make it more complex too.
Pierre-Yves David <pierre-yves.david@octobus.net> [Thu, 08 Dec 2022 15:13:17 +0100] rev 49784
chg: scale the timeout in test with the rest
This should avoid some flakiness where the logs reports server shutting down.
Pierre-Yves David <pierre-yves.david@octobus.net> [Wed, 07 Dec 2022 20:12:23 +0100] rev 49783
bundle: emit full snapshot as is, without doing a redelta
With the new `forced` delta-reused policy, it become important to be able to
send full snapshot where full snapshot are needed. Otherwise, the fallback delta
will simply be used on the client side… creating monstrous delta chain, since
revision that are meant as a reset of delta-chain chain becoming too complex are
simply adding a new full delta-tree on the leaf of another one.
In the `non-forced` cases, client process full snapshot from the bundle
differently from deltas, so client will still try to convert the full snapshot
into a delta if possible. So this will no lead to pathological storage
explosion.
I have considered making this configurable, but the impact seems limited enough
that it does not seems to be worth it. Especially with the current
sparse-revlog format that use "delta-tree" with multiple level snapshots, full
snapshot are much less frequent and not that different from other intermediate
snapshot that we are already sending over the wire anyway.
CPU wise, this will help the bundling side a little as it will not need to
reconstruct revisions and compute deltas. The unbundling side might save a tiny
amount of CPU as it won't need to reconstruct the delta-base to reconstruct the
revision full text. This only slightly visible in some of the benchmarks. And
have no real impact on most of them.
### data-env-vars.name = pypy-2018-08-01-zstd-sparse-revlog
# benchmark.name = perf-bundle
# benchmark.variants.revs = last-40000
before: 11.467186 seconds
just-emit-full: 11.190576 seconds (-2.41%)
with-pull-force: 11.041091 seconds (-3.72%)
# benchmark.name = perf-unbundle
# benchmark.variants.revs = last-40000
before: 16.744862
just-emit-full:: 16.561036 seconds (-1.10%)
with-pull-force: 16.389344 seconds (-2.12%)
# benchmark.name = pull
# benchmark.variants.revs = last-40000
before: 26.870569
just-emit-full: 26.391188 seconds (-1.78%)
with-pull-force: 25.633184 seconds (-4.60%)
Space wise (so network-wise) the impact is fairly small. When taking compression into
account.
Below are tests the size of `hg bundle --all` for a handful of benchmark repositories
(with bzip, zstd compression and without it)
This show a small increase in the bundle size, but nothing really significant
except maybe for mozilla-try (+12%) that nobody really pulls large chunk of anyway.
Mozilla-try is also the repository that benefit the most for not having to
recompute deltas client size.
### mercurial:
bzip-before: 26 406 342 bytes
bzip-after: 26 691 543 bytes +1.08%
zstd-before: 27 918 645 bytes
zstd-after: 28 075 896 bytes +0.56%
none-before: 98 675 601 bytes
none-after: 100 411 237 bytes +1.76%
### pypy
bzip-before: 201 295 752 bytes
bzip-after: 209 780 282 bytes +4.21%
zstd-before: 202 974 795 bytes
zstd-after: 205 165 780 bytes +1.08%
none-before: 871 070 261 bytes
none-after: 993 595 057 bytes +14.07%
### netbeans
bzip-before: 601 314 330 bytes
bzip-after: 614 246 241 bytes +2.15%
zstd-before: 604 745 136 bytes
zstd-after: 615 497 705 bytes +1.78%
none-before: 3 338 238 571 bytes
none-after: 3 439 422 535 bytes +3.03%
### mozilla-central
bzip-before: 1 493 006 921 bytes
bzip-after: 1 549 650 570 bytes +3.79%
zstd-before: 1 481 910 102 bytes
zstd-after: 1 513 052 415 bytes +2.10%
none-before: 6 535 929 910 bytes
none-after: 7 010 191 342 bytes +7.26%
### mozilla-try
bzip-before: 6 583 425 999 bytes
bzip-after: 7 423 536 928 bytes +12.76%
zstd-before: 6 021 009 212 bytes
zstd-after: 6 674 922 420 bytes +10.86%
none-before: 22 954 739 558 bytes
none-after: 26 013 854 771 bytes +13.32%