Matt Harbison <matt_harbison@yahoo.com> [Mon, 17 Dec 2018 17:44:45 -0500] rev 40993
setup: avoid attempting to invoke the system-wide hg.exe on Windows
On Windows, the executable in the current directory gets priority over anything
in $PATH (both for cmd.exe and MSYS). That means, the former code was launching
the local hg.exe instead of the system-wide one, if it was previously built. If
that failed, it then fell back to the local hg code, but run through python.exe.
I'm not sure what it is about ef7119cd4965, but that started throwing up a
messagebox that python37.dll couldn't be loaded. (And indeed, python37 is not
in $PATH by default.) Invoking the local hg via the current python avoids that.
Boris Feld <boris.feld@octobus.net> [Mon, 17 Dec 2018 10:46:37 +0100] rev 40992
delta: ignore base whose chains already don't match expectations
If we know the existing chain does not match our criteria, there is no point
to build a delta to append. This is especially useful when dealing with a full
text much smaller than its parent. In that case, the parent chain is probably
already too large.
example affected manifest write
before: 1.421005s
after: 0.815520s (-42%)
Boris Feld <boris.feld@octobus.net> [Mon, 17 Dec 2018 10:42:19 +0100] rev 40991
delta: exclude base candidate much smaller than the target
If a revision's full text is that much bigger than a base candidate full text,
we no longer consider that candidate.
This solves a pathological case we encountered on a very specify repository.
It contains a long series of changesets with a very small manifest (one file)
co-existing with others changesets using a very large manifest.
Without this filtering, we ended up considering a large number of tiny full
snapshots as a potential base. It resulted in very large delta (the size of
the full text) and mercurial spending 99% of its time compressing these
deltas.
The timing of a commit moved from about 400s to about 10s (still slow, but not
ridiculously slow).
Boris Feld <boris.feld@octobus.net> [Mon, 17 Dec 2018 10:37:22 +0100] rev 40990
perfrevflogwrite: clear revlog cache between each write
We want to measure write time from a cold cache (similar to commit). So we
need to clear the cache to prevent computation from rev N-1 to interfere with
rev N.