Using fs.writeFileSync in a serial style becomes especially costly when
we're writing a lot of files. In a recent profiling exercise I did on
Windows, nearly 80% of the time taken by Builder#_copyDirectory was spent
just closing the written files. By using the asynchronous fs.writeFile
function, we should be able to parallelize at least some of this work, and
await all the promises at the very end of copying the directory.
* Avoid modifying unibuild.watchSet in PackageSourceBatch._watchOutputFiles.
Should fix#10736 by preventing old hashes from remaining in
unibuild.watchSet, which was sometimes causing isUpToDate to fail
immediately upon restart.
* Regression test for issue #10736.
Although I managed to make this work somewhat better in my previous
commit, we still have to create a .tar.gz archive of the dev bundle, and
7z appears not to handle junction points very well. I think it will be
much simpler to avoid the helpers/builtin symlinking hack on Windows.
In PR #10720, we introduced the makeCheapPathFunction in an effort to
reduce the caching overhead for very frequently called (and already pretty
quick) operations like files.stat.
However, the default maximum LRU cache size of Math.pow(2, 16) can cause
quite a bit of cache eviction churn for large applications, which @veered
has identified as a potential source of build performance problems.
By setting the maximum cache size to Math.pow(2, 20) instead, I am no
longer seeing any files.stat calls in the profiling output for rebuilding
a large internal app, saving several seconds of rebuild time. The obvious
downside is that this cache might accumulate more memory over time, which
is why I didn't just set the max to Infinity, though that might be a
viable option if the total set of paths ever stat'd is small enough to fit
into the available memory.
In the future, I hope to find ways of managing LRU cache size that respond
to actual memory pressure (relative to available memory), rather than
pruning the cache after an arbitrary numeric threshold is reached.
I noticed recently that progress messages were no longer displayed (though
the ASCII spinner was still animating). I bisected the problem to the
progress.ts conversion in #10650, and I attempted to reproduce the
conversion, step by step. The problem ended up being the renaming of
._title to .title, which was a mistake because external code accesses
progress._title. Once I updated those invasive access points, progress
messages began displaying once again.
Partial alternative to using json5, which we tried in commit
5124cb495c, which was recently reverted due
to performance concerns: 0e19c365a8.
Should solve the specific problem reported in #10688.
This reverts commit 5124cb495c.
Perhaps unsurprisingly, JSON5 is much slower than JSON, so much so that
"other optimisticReadJsonOrNull" has become the single most expensive leaf
when building one of our internal applications. We will have to find
another solution.
