6.0 Performance Regression

[LPGhatguy]

A Roblox engineer working on an internal project recently reported a pretty bad performance regression for rojo build for version 6.0 vs 0.5.4 on macOS:

0.5.4:

real    0m0.071s
user    0m0.035s
sys     0m0.030s

6.0.0-rc.1

real    0m3.923s
user    0m0.852s
sys     0m2.858s

This is a 55x performance regression, which is incredibly bad. On my Windows machine with the same project, I had these performance numbers:

0.5.4:

real    0m0.134s
user    0m0.000s
sys     0m0.015s

6.0.0-rc.1:

real    0m0.377s
user    0m0.000s
sys     0m0.015s

0.5.4 is around twice as slow on my machine, but 6.0.0-rc.1 is around ten times faster on my machine.

After bisecting and probing the code a little bit, I narrowed this issue down to #297. Two specific issues I believe we regressed on:

1. rojo build now watches all files, instead of only happening when running rojo serve. We can introduce this as a configurable feature of the VFS again.
2. Removal of the file cache means that fs::metadata is invoked several times for each path. While repeated reads are not a factor for performance when building, it may be a good idea to cache more.

I tested this mitigations in the sketchy-perf-changes branch by removing file watching completely and implementing a permanent metadata cache. Both of these changes break serve, but I was able to print 6.0.0-rc.1 in line with the performance numbers for 0.5.4 on my machine.

When running the patched binary on the original reporter's machine, however, performance is around 2x worse than 0.5.4 still:

real    0m0.165s
user    0m0.050s
sys     0m0.059s

This warrants some serious digging I think.

[LPGhatguy]

Fixed after merging #324 and introducing 782b054!