mirror of
https://github.com/yu-i-i/overleaf-cep.git
synced 2026-11-12 13:13:45 +01:00
Jakob Ackermann
37e27bb22e
* [monorepo] reduce writes from chrome in cypress Chrome puts its shared memory in /dev/shm, and docker's default 64MB is far too small for it - which is why Cypress passes --disable-dev-shm-usage, spilling those segments to files on the container's writable layer instead. Under one spec that is a few hundred MB of write-then-delete churn; under the CI fan-out it was most of the disk bandwidth on the whole box. So every service built from dockerfiles/cypress gets shm_size 1gb and every cypress config drops the flag again: the SaaS e2e service, web's CI and local component-test services plus its standalone cypress one, and server-ce's e2e and native cypress services. Both e2e-open variants inherit the setting through `extends`; the native one is a service of its own, and would be left running Chrome against 64MB once the flag is gone. * [monorepo] adopt cypress-pnp-reporter Materializing the reporter packages copied 267 packages / 3779 files (59MB) into /tmp on every container start, once per spec. Resolving them through PnP inside Cypress's Electron process instead drops that to zero: the entrypoint shell went from 55.6MB of write_bytes to 0. * [web] serve built assets from a prepared buffer between builds Serving a chunk cost ~2ms whatever its size, and an editor page pulls ~144 of them, so webpack-dev-server burned 2-3 cores through an e2e run purely answering requests for output that had not changed since the last build. It is not bytes: cutting the bundle 2.6x with a cheaper devtool moved CPU by 8%, and an empty 304 cost nearly as much as the full body. It is the per-request work - reading the asset out of the in-memory filesystem and re-hashing the whole body for a strong ETag, on every request. Three things a build does that this has to respect, each found by tracing the hooks rather than by reading them: - `done` fires twice per build, for the compilation and for the manifest, both with the same hash. A flag would have been cleared by the second while the first was still running, so the gate is a counter. - An incremental `done` reports only the assets it rewrote - 16, not 630 - so asset names accumulate. Resetting them on `invalid` left the manifest's two behind and nothing was served from here again. - webpack-dev-middleware holds requests during a rebuild. This sits in front of it, so it steps aside for that window and drops the cache when a rebuild starts. Whether a response can be kept is then decided as it ends rather than as it began: an epoch bumped on `invalid` names the build its bytes were read from, and only a response from the current build with no rebuild in flight is stored. A rebuild starting mid-response can no longer write the previous build's bytes in behind the cache being dropped, which eviction by asset name could not catch anyway - an incremental build does not name the assets it left alone. Measured on the editor entrypoint, 16 concurrent clients: 0.26-0.31s -> 0.045-0.050s of webpack CPU per page load, and 0.0455s immediately after a rebuild. Same chunk graph, byte-identical output. GitOrigin-RevId: 08b4ab2d665414069bc01b4763dd64faa9b4e5cb
2 lines
16 B
JavaScript
2 lines
16 B
JavaScript
// no setup yet
|