GET /api/photos sent only X-Next-Cursor — no ETag — so every gallery re-mount rebuilt up to 500 PhotoDtos, serde-serialized the whole vector, and shipped the full body even when nothing changed. The handler now emits a lightweight content-derived ETag (hash of before + limit + max(modified_at) + row count) and honours If-None-Match, with Cache-Control: private, no-cache so the SPA's default fetch cache mode always revalidates. An unchanged "navigate away and back" becomes an empty 304 instead of a full rebuild + reserialize + transfer. The DB query still runs (the cheap part); the win is skipping the DTO build, serialization, and body bytes. Proven end-to-end (throwaway Postgres + server, 7 images): 1st GET (no If-None-Match) -> 200 4586 bytes + ETag 2nd GET (If-None-Match matches) -> 304 0 bytes 3rd GET (If-None-Match stale) -> 200 4586 bytes (correctly invalidated) ~655 B/photo, so a full 500-row first page saves ~320 KB + a 500-DTO build/serialize per unchanged revalidation. Unlike a cold load this is the common gallery-navigation path, so it hits real user-facing latency. Regression test: tests/api/photos_etag.hurl (added to the api-test suite). Methodology in benches/PHOTOS-ETAG.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2.0 KiB
Photos timeline — ETag / 304 conditional revalidation
GET /api/photos returned only X-Next-Cursor — no ETag. So every time the
gallery re-mounts (navigate away and back), the server rebuilds up to 500
PhotoDtos (each a FileDto::from with several allocations), serde-serializes
the whole vector, and ships the full JSON body — even when nothing changed.
The handler now emits a lightweight, content-derived ETag and honours
If-None-Match, with Cache-Control: private, no-cache so the browser always
revalidates. An unchanged gallery re-mount becomes an empty 304 instead of a
full rebuild + reserialize + re-transfer.
- ETag = hash of
(before, limit, max(modified_at), row count)over the page — page identity plus a freshness signal, mirroring the file-list endpoint. Any upload/edit/delete that changes the page bumpsmax(modified_at)or the count, so the ETag changes and the client gets the fresh body. - The browser revalidates because the SPA's
apiFetchuses the default fetch cache mode andno-cacheforces a conditional request. - The DB query still runs (it's the cheap part); the win is skipping the DTO build + serialization + body bytes.
Reproduce / proof
tests/api/photos_etag.hurl (run by the api-test suite) asserts: first GET → 200
- ETag +
Cache-Control: no-cache; conditional GET with the matching ETag → empty 304; a staleIf-None-Match→ full 200. End-to-end measurement against a throwaway Postgres + server (7 uploaded images):
1st GET (no If-None-Match) : 200 4586 bytes + ETag
2nd GET (If-None-Match matches) : 304 0 bytes ← the win
3rd GET (If-None-Match stale) : 200 4586 bytes ← correctly invalidated
~655 B/photo on the wire. At a full 500-row first page that's ~320 KB + 500-DTO build/serialize saved per unchanged "navigate away and back", plus the response bytes. Steady-state gallery navigation is the common case, so this hits real user-facing latency (unlike a one-time cold load).