Stringify layers for faster worker transfer

[mourner]

Ref #3153, not the cleanest change but appears to reduce time-to-fist-render by ~150ms.

Couldn’t run benchmarks yet because they seem to be broken at the moment (only run on master).

cc @jfirebaugh @lucaswoj

[mourner]

Benchmarks

buffer

master 1f8351d: 916 ms
stringify-layers 6ad9724: 943 ms

frame-duration

master 1f8351d: 10.6 ms, 15% > 16ms
stringify-layers 6ad9724: 10.5 ms, 15% > 16ms

[lucaswoj]

LGTM

appears to reduce time-to-first-render by ~150ms

How many ms is the time-to-first-render now? Does the map-load bench capture this metric? Could we write a benchmark that does capture this metric?

[mourner]

@lucaswoj looking at the timeline, it's still around 2000ms, but I haven't captured it exactly yet, and the benchmarks don't capture this too — I thought about adding one, although this should be a separate PR.

[anandthakker]

Does the map-load bench capture this metric? Could we write a benchmark that does capture this metric?

@lucaswoj @mourner I think it probably makes sense to just change the map-load bench into a first-render, since the latter is what matters most?

[mourner]

Benchmarks barely show a ~50-100ms difference, so it might not be worth the added hackiness.

[anandthakker]

@mourner If this change speeds up the set layers and update layers actions, then it could also affect the performance of style mutation methods like map.setFilter(), etc., right? For those cases, I think even a few ms difference can matter.

[lucaswoj]

Benchmarks barely show a ~50-100ms difference, so it might not be worth the added hackiness.

This PR is mergeable. Your call on whether or not to merge it.

Thanks for testing this!

[lucaswoj]

Closing as stale. We can come back to this at any time.