Vercel 504s: CDN Caching, Timeout Guards, DID Caching

Second pass at the 504 problem.

The earlier fix — mapWithConcurrency and client-side offloading of the home page — helped with the index. Blog post pages and the subscriptions/recommendations endpoints were still timing out. The underlying cause was obvious in hindsight: the in-memory Map in cache.ts is cold on every Vercel function invocation. Module scope resets between serverless calls. The cache has never actually done anything on production.

Without any CDN caching headers, every visitor triggers a full network fan-out to the AT Protocol ecosystem regardless of how recently someone else hit the same route. No request was ever cheap.

what changed

maxDuration config exports. Added export const config: Config = { maxDuration: 60 } to each route file using the Vercel adapter — the correct per-route mechanism for SvelteKit. Also set a global "": { "maxDuration": 60 } in vercel.json as a backstop.

CDN Cache-Control headers. setHeaders with s-maxage in all page load functions; Cache-Control on all API responses. Blog posts get an hour — they don't change once published. The home page gets sixty seconds. Subscriptions and recommendations get five minutes. Vercel's edge now caches these; the cold function only runs once per TTL window per edge node rather than on every request.

AbortController timeouts. Added a withTimeout wrapper applied at the entry point of fetchSubscriptions, fetchRecommendations, and fetchComments. Every individual external fetch — PDS record lookups, profile fetches — now aborts after 8 seconds. Previously a single unresponsive PDS could stall the entire Promise.allSettled fan-out indefinitely. Now it fails that one record and the settled results still return.

DID document cache. Module-level Map in resolveDid with a one-hour TTL. Within a single request resolving multiple records from the same author — which is the common case for subscriptions and recommendations — this skips redundant plc.directory round-trips. Also compounds on warm Vercel instances.

about/+page.server.ts.** Was calling every ATProto function without threading fetch through — falling back to bare global fetch. Fixed, and added setHeaders for CDN caching while I was in there.

The .catch() fallbacks that were already on every external call ensure graceful degradation if the AT Protocol network is genuinely unavailable. The timeout guards just make sure "unavailable" resolves in 8 seconds rather than never.