CQRS said the worker owns decisions and FastAPI only reads cache. Users still waited seconds on /dashboard. The compute boundary was fine. The read path was not.

This post is the engineering story behind ADR-056 and the ADR-036 OHLCV amendment.

What was actually slow

Three different clocks stacked:

1. Fly contention. One shared vCPU runs uvicorn and the market-cache worker on one SQLite file. During a refresh, public GETs for /api/v2/dashboard often took 3–12 seconds. Logs showed ohlcv_written≈67266 every cycle — full Yahoo windows rewritten when history had not moved.
2. Serial console UX. Client SignedInOnly waited for session before mounting pages that only need a public cache snapshot. Then cache: "no-store" plus a Bearer on unscoped reads skipped the CDN middleware on purpose.
3. Stale HTML + heavy JS. Golden Pit kept a 1h ISR seed (all D) after grades were already B/B- on the API. Dashboard shipped ~3700 lines of client UI in one chunk.

Decision math was not running on the request path. Waiting for Fly during a delete-rewrite of K-lines was.

The pack

Edge cache for identity-free GETs

PublicCdnCacheMiddleware allowlists worker-backed routes and sets CDN-Cache-Control + Vercel-Cache-Tag. Vercel rewrite caching is on for the same paths. Unscoped dashboard / today / golden-pit client fetches no longer send Bearer or no-store. Sign-in stays a product gate, not an API auth requirement for the shared market snapshot.

After ship: CDN HIT on dashboard about 0.15s.

Parallel auth + cache seed

SSR runs auth() and the cache fetch with Promise.all (lib/console-session.ts). Guests still see the sign-out gate. Signed-in users do not wait login → then data.

Stop rewriting unchanged history

OHLCV cycles now full-refresh only when the store is empty, the last full pull is older than 24h, or a closed bar’s price moved (split / re-base). Otherwise fetch ~5 days and upsert. Forced production run on 2026-09-30: written=3915, full=0, incremental=69 in ~18s.

Delivery details

Dashboard entry is a thin dynamic() shell; body is a separate chunk. Golden Pit revalidate=60 and always revalidates client-side after paint.

What we did not do

• Recompute engines inside FastAPI when the cache “felt stale.”
• CDN-cache responses that vary by user identity.
• Pretend a bigger VM alone fixes write amplification.

Numbers (honest)

Check	Before (typical)	After (2026-09-30)
Dashboard via invest CDN	often multi-second MISS to Fly	~0.15s HIT
OHLCV rows per cycle	~67k	~4k incremental when warm
Console HTML shells	mixed	mostly 0.2–0.5s TTFB

Browser hydrate is still not free. /method and /plans 404s are routing gaps, not latency.

Links

• Repo: https://github.com/xingaiapp/xingai-invest-ai
• ADR-056: https://github.com/xingaiapp/xingai-invest-ai/blob/main/docs/adr/056-console-read-path-cdn-parallel-auth.md
• ADR-036: https://github.com/xingaiapp/xingai-invest-ai/blob/main/docs/adr/036-worker-ohlcv-bars-cache.md
• Design: https://github.com/xingaiapp/xingai-enterprise-ai-design/blob/main/articles/2026-09-30-shared-cpu-cqrs-edge-cache-vs-write-amplification.md
• Product: https://invest.xingai.app

Disclaimer: informational engineering notes. Not investment, legal, or operational advice for your deployment.