Cut Supabase + Vercel spend without changing user-facing behavior (#355) * Suppress yjs realtime broadcasts when no other peer is subscribed A solo editor streams doc updates every 50ms and cursor awareness every 80ms into an empty channel — every message billed by supabase, delivered to nobody. Solo sessions are the common case, so this was the bulk of ~7.9M billed realtime messages/month. Track channel membership via presence and drop doc-update/awareness broadcasts while no other peer is present. Late joiners converge through replicache exactly as they already did for updates sent before they subscribed. Recently received broadcasts also count as peer presence so pre-presence clients keep working across the deploy boundary. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Cut database and storage egress in the hot read paths Uncached supabase egress ran 839GB/mo against 53GB cached. Three contributors addressed here: - Replicache pulls return a full document snapshot every time, and every open tab polled one per minute. Stretch the idle poll to 5 minutes and pull explicitly when the poke channel resubscribes after a drop, so convergence still happens immediately after a missed-poke disconnect. Worst case for a poke lost on a live socket is now 5 minutes of staleness for a passive viewer instead of 1. - The bsky feed skeleton endpoints selected entire document rows (including full content jsonb, 25 at a time, on unbounded external traffic) to read only a post ref. Project just uri/sort_date/refs. - storage-js expects cacheControl in seconds; passing a full header string stored a malformed Cache-Control (max-age=public, ...) on every uploaded image, defeating CDN caching of user assets. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Resize images with sharp instead of Supabase image transforms Supabase bills $5 per 1,000 origin images transformed per month (19,275 last cycle). Replace all three transform pipelines with sharp, which the repo already uses for cover thumbs: - atproto_images: resize the cached PDS blob in-function rather than via storage.download({transform}); also drops the per-request list() existence probe and reuses fetched bytes instead of re-downloading. - link_previews: store the screenshot already downscaled to its 240x208 display size and serve the object directly, transform-free. - next/image loader: point at a new /api/resized_images sharp proxy (originals come off the storage CDN; responses are edge-cached for a year) instead of the render/image endpoint. Output dimensions and no-upscale behavior match imgproxy's; EXIF orientation is applied explicitly via rotate(). Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Serve image bytes from Supabase's CDN via redirect, not through Vercel The Vercel invoice shows fast data transfer already ~500GB past the included 1TB, so every image byte streamed through these routes bills at $0.15/GB — worse than the Supabase egress the sharp rewrite was saving, and cover/doc images are pure byte volume. Store each resized variant (and the cached original) in storage once and 302 to its public URL: bytes then leave via Supabase's CDN as cached egress ($0.03/GB with free headroom) and Vercel serves only a year-long-edge-cached redirect. ensureOriginalCached also repairs legacy objects stored with the malformed 'max-age=public, ...' Cache-Control by re-uploading them once with a well-formed value, since redirecting to an uncacheable object would forfeit the CDN saving. Streaming from the PDS remains the last resort when nothing could be cached. Tradeoff: one extra redirect hop per image on a cold browser cache. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Cut Vercel Fluid compute burned on screenshots, favicons, and push logging Fluid bills provisioned memory for full instance wall time, so the multi-second remote-browser renders behind og-images were the main burn: every unfurl bot re-fetch after the 60s revalidate paid a fresh render. OG cards are effectively immutable post-publish — raise revalidate to a day and let the CDN serve them (stale by at most a day after an edit). The share-card screenshot route had no caching at all; edge-cache it per url. Also: the per-leaflet favicon was force-dynamic + no-cache, running two DB queries and a satori render on nearly every page load — cache it, picking up theme edits within the hour. Gate the per-push performance log summary behind LOG_PUSH_PERFORMANCE. Stop runtime-cache writes for arbitrary user-pasted URLs in resolveBlueskyPostUrl (near-zero hit rate, one billed cache write per paste). Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Reduce Vercel request volume and reader-page bundle weight - Serve the hot feed from a CDN-cacheable GET route. It's identical for every user and already 5 minutes stale from redis, but each reader's load paid an uncacheable POST invocation returning 50 full posts. Worst-case staleness goes from ~5 to ~10 minutes. - Widen replicache pushDelay 500ms -> 1s: halves push invocations during sustained typing. Live co-editing latency rides the yjs channel; replicache-only changes (e.g. block moves) reach collaborators ~0.5s later, and unsent mutations survive tab close in local storage. - Load the comment editor and bsky share composer on interaction instead of bundling them: both drag prosemirror into every public post page for readers who never comment or share. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Drop sub-1%-impact interventions Ranking the branch's changes by predicted monthly savings, three fall under 1% of the total (~$310/mo) because their frequency drivers are human actions, not machine traffic: quote-screenshot caching (share-modal opens), the push perf-log gate (sub-ms CPU per push), and the no-store paste-URL resolution (a runtime-cache write per paste). Revert them to keep the diff to changes that pay for their risk. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Answer conditional feed requests with 304s The rss route gets polled harder than any page on the site (156K requests in the last transfer window), and transfer out is billed even on CDN cache hits, so every poll paid for the same unchanged bytes. Feed readers nearly all send If-None-Match; derive an ETag from the body — generateFeed already pins the updated stamp so an unchanged feed is byte-stable — and return an empty 304 on match. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_018MbcWeXrpeshUJLFDJKzek * Snap image-resize dimensions to a fixed width ladder (#358) The resized_images and atproto_images routes accept caller-controlled dimensions, and every novel value mints a permanent stored variant on cache miss — an unauthenticated caller could mint ~2000 objects per image (~4M per cid for w×h) just by walking the width space. Snap requests to a 4-entry ladder shared from supabase/imageSizes.js so the keyspace per image is bounded: 360/800 are the app's own cover-thumbnail widths, 1200 covers document-body images at retina density, 2000 is the existing cap. Snapping up (never rejecting) keeps URLs already living in edge and browser caches working across the deploy, and off-ladder requests collapse onto the same stored objects because the snapped value keys both the sharp resize and the variant path. atproto_images height is snapped too rather than dropped, keeping old cached URLs valid while bounding the w×h grid. Align the clients so real traffic is on-ladder before it reaches the edge cache: the next/image loader snaps in the browser, and images.deviceSizes is set to the ladder (imageSizes emptied — its entries must sit below the smallest deviceSize and every icon-sized use already collapses to 360 via the loader) so srcset candidates come out of the same four widths instead of Next's 16-entry default. The share composer's 1000px cover prefetch moves to the 1200 rung it would be snapped to anyway. Claude-Session: https://claude.ai/code/session_01MQBgS6Lb1BrTgg2pu1Bfm4 Co-authored-by: Claude --------- Co-authored-by: Claude