decentraland / atlas-server / 39866997711

93%

master: 94%

Pull #155

github

fix: tighten the render pool's stop, abort, sizing and metrics

- A render is tracked as soon as it is admitted, so stop() waits for one
  admitted right before it instead of failing it with a 503.
- A client that leaves while its job is built no longer gets a worker.
- Threads default to min(CPUs, 2) and each worker's V8 heap is capped at
  256 MiB; canvas pixels are native memory and not covered by it.
- One thread is kept and warmed up on start. Terminating a thread while
  it loads canvas aborts the process, and the first render no longer pays
  for loading it. Thread errors outside a render are logged.
- Render times all come from the worker's clock, failures included.
  Failures before the worker runs are logged instead, and building the job
  is timed in dcl_map_render_prepare_seconds.

1441 of 1592 branches covered (90.52%)

Branch coverage included in aggregate %.

719 of 765 new or added lines in 16 files covered. (93.99%)

4966 of 5292 relevant lines covered (93.84%)

2032.91 hits per line