Follow-up to f81b114b, addressing the three ways that commit could make things
worse rather than better. All three were verified against the real database or
by breaking the test and watching it fail.
- The grace window is now capped at 120s. A window is a cushion for the TTL
boundary, not a second TTL, but the call sites treated it as the latter: the
5 min values/staff routes and the 10 min teams route asked for a window as
long as or longer than their own TTL, so a single large staleMs silently
doubled how far behind a value could be served. Nothing marked those as
unsafe, because nothing looked wrong. The cap lives in the cache rather than
at the call sites so no future route can reintroduce it. Routes that asked
for less than 120s (the 10s online poll, the 20s news cache) are unchanged,
so their intended cushion still does its job.
- A request no longer queues behind an arbitrarily old render. Sharing a render
is what collapses a cold-cache stampede into one render, but a hung render
used to hold up everyone who arrived after it. A newcomer past 2s now serves
the placeholder instead of waiting, reusing the ImagerUnavailableError path
that "both upstreams down" already takes. The caller that actually started
the render keeps waiting, which is correct: it is the one whose image this
is. When the join window is removed the new test hangs for the full 10s it
was meant to prevent, which is the tail this bounds.
- The information_schema row-count estimate is gone; the counters are exact
again. Running it against the live database: users 165, rooms 92, camera_web
0, and the estimate was 0.00% off on all three. At 165 rows an index scan is
cheaper than the extra round trip the estimate needed, so the optimisation
bought nothing and traded a guaranteed-correct member count for an
approximation that InnoDB would only make less accurate as the table grows.
The exactness is now pinned by tests: a real zero stays zero, a database
error propagates instead of becoming a number, and each counter counts the
table it claims to. The module stays, because the homepage and the boot
warm-up writing different values to the same cache key is its own bug.
The module comment records the measured numbers, because "COUNT(*) is too slow"
sounds true in the abstract and is false here.
3223 tests pass.
Three separate things that were each costing more than they needed to on the
hot path.
- Single-flight avatar renders. The disk cache was checked first and a miss
went straight to the upstream, with nothing shared between callers, so a page
requesting dozens of avatars at once turned N concurrent requests for one
figure into N renders. A render is the most expensive operation this app
does, and the duplication happened exactly when the cache had nothing to
offer. Eight concurrent requests now cause one render instead of eight. The
map lives on globalThis because Next can evaluate the module more than once
per process, and two copies would each start their own render.
- Let public read-only routes be cached by a shared cache. Every JSON response
was `cache-control: no-store`, so a CDN in front of the app could not answer
any of it and every request reached the origin. publicCacheControl() opts a
route in with s-maxage and stale-while-revalidate, using the same TTL as the
server-side cache so the two layers cannot disagree. The default stays
no-store: most routes here are personalised, admin-only or auth-dependent.
/api/badges/leaderboard is deliberately left alone because it returns
per-viewer rank entries to signed-in callers.
Note this only takes effect once a cache rule exists for /api/* at the CDN, or
the explicit `cache: "no-store"` is dropped from the client fetches (24 files
do that today, including the /api/online poll). The headers alone are inert
until one of those happens.
- Take the homepage row counts from the storage engine estimate instead of
COUNT(*), which walks an index and gets slower as the tables grow. A missing
or zero estimate falls back to the exact count rather than ever showing a
wrong zero. The online count stays exact: it is an indexed read over a small
subset and a few seconds of drift reads as broken rather than approximate.
The counters move into one module because the homepage and the boot warm-up
populate the same cache keys, so two implementations would race to write
different values into the same entry.
3223 tests pass.
A fallback render drops the requested effect and is only a degraded
stand-in, so writing it to the 30 day disk cache kept serving the worse
image long after the local renderer recovered. Cache primary renders only
and let the next request pick up the real render.
Effect renders need a little over 4s, which the 4s primary timeout cut off,
so every avatar with the default effect fell through to an unreachable
public fallback and rendered as a placeholder. Raise the primary budget
above the observed render cost and shorten the fallback budget.
Also stop the proxy from stamping no-store over the avatar and media
responses, so browsers keep the long-lived Cache-Control the route already
sends, and recreate the imaging cache directories with the container user
on every deploy, since root ownership made those cache writes fail
silently.
- Add persistent disk cache for rendered avatars/badges (storage/imaging)
so repeats never touch the flaky local renderer and cached renders
survive upstream downtime
- Serve cache-first with stale-on-error; cut primary/fallback timeouts
from 10s/6s to 4s/4s so failing images cannot stall pages
- Avatar proxy now returns a graceful 200 silhouette instead of 502 when
no renderer can produce a figure, so no broken-image glyphs appear
- Badge endpoint becomes a caching proxy trying configured CDN, public
Habbo CDN and local /swf copy in order, and drops the fragile IP rate
limit that could blank badge streams
- Route all site badge images (profile, me, badges, apply pages) through
the cached proxy instead of hot-linking images.habbo.com
openhands