mirror of
https://github.com/webstudio-is/webstudio.git
synced 2026-11-02 18:18:35 +09:00
Oleg Isonen
fc15ebf81a
## Description The cold `build.contentDatabasePublishDiagnostics` path was parsing the same reachable MDX documents while the publication plan converged, then parsing them a third time to produce omission diagnostics. Content-heavy projects can approach or exceed the 15-second serverless limit on that first request. This change: - creates a request-scoped MDX dependency resolver that reuses parse results by asset revision while the publication plan converges; - carries the final template omissions out of the existing publication pass, so the diagnostics endpoint does not walk the artifact again; - keeps the existing sequential asset validation/preparation/stability checks unchanged; - does not use module-level caching or depend on a warm serverless instance; - leaves the timeout error handling from #6406 unchanged. ## Performance evidence An isolated cold-path benchmark used the same MDX payload (2,000 sections) and compared the previous three independent dependency-resolution walks with the new request-scoped convergence path: | path | time | | --- | ---: | | before | 1,178 ms | | after | 404 ms | That is a 66% reduction in the measured MDX resolution phase. The tests also assert the underlying work reduction directly: an unchanged revision is parsed once across convergence passes, while a changed revision is parsed again. For a stable 214-document candidate set, this removes up to 428 of the previous 642 parse operations. ## Verification - `pnpm --filter @webstudio-is/project-build test` — 2,420 tests passed - `pnpm --filter @webstudio-is/builder exec vitest run --project core-server` — 336 tests passed, including server-side e2e runner tests - `pnpm --filter @webstudio-is/project-build typecheck` - `pnpm --filter @webstudio-is/builder typecheck` - `pnpm lint` - `pnpm check:package-boundaries` - `pnpm check:generated-api` ## Code Review - [x] made a self-review - [x] added comments for the request-scoped cache and revision safety - [x] added tests for parse reuse, revision invalidation, and the optimized endpoint integration path ## Before merging - [ ] tested on a deployment preview with a production-sized content project ## Additional content-read optimization Successive compilation plans now share already-fetched, revision-keyed content bytes within one callback-scoped preparation session. The byte store: - is created inside one asset-stability attempt and discarded when that attempt finishes; - is not stored on the repository, in a module, or outside the request; - remains bounded by the existing 2 MiB hydration limit; - refetches changed revisions and starts clean on stability retries. Tests verify that two overlapping plans read unchanged content once, a revision change forces another read, a new session reads again, and a stability retry creates a new session. A fully overlapping two-pass workload therefore performs up to 50% fewer object-storage content reads. Additional verification: - `pnpm --filter @webstudio-is/asset-uploader test` — 292 tests passed - `pnpm --filter @webstudio-is/asset-uploader typecheck` ## Remaining optimization assessment The percentages below estimate whole-request improvement and are workload-dependent; they are not additive. | optimization | estimated improvement | complexity | decision | | --- | ---: | ---: | --- | | Request-local content-byte reuse | 10–35% | 3/10 | Implemented | | Stop when the initial dependency plan is already a fixed point | 5–15% for static/known MDX; ~0% for dynamic discovery | 2/10 | Implemented | | Reuse a canonical source snapshot/materialized entries across different plans | 15–30% | 8/10 | Deferred: large semantic and invalidation surface | | Reuse the bundle's initial asset-data snapshot | 1–5% | 5/10 | Not worthwhile | | Add a diagnostics-specific bundle loader | 2–8% | 6/10 | Not worthwhile for this PR | | Parallelize or skip unrelated user/project metadata | 0–3% | 2/10 | Not meaningful | | Combine collection validation and reserved-asset inspection | 0–5% normally | 4/10 | Only meaningful for projects with many collections | The fixed-point fast path is covered by a test asserting one artifact preparation and one dependency-resolution call when the original plan is already stable.