perf(ama): drain durable operations inline, relax cron cadence

[CaliCastle]

Why

Neon compute usage has been ~183 CU-hrs/month — the endpoint never autosuspends. The */5 * * * * AMA work cron issues an unconditional hold-release UPDATE plus a claimDue query on every run, resetting Neon's 5-minute autosuspend timer exactly at the threshold, 288 times a day. Page visits are not a factor (public routes are fully static under Cache Components and never touch the database).

What changed

Inline drain on mutation. New kickAmaOperations() in lib/ama/booking/server.ts runs runner.run() via waitUntil in the background of the current invocation. It is called after a successful response in the five routes that enqueue durable work:

• Stripe webhook — confirmation email and Finalizing Booking recovery start the moment payment lands (previously up to 5 minutes later)
• Guest cancel / reschedule — refunds, artifact updates, and emails go out immediately
• Admin booking action / operation retry — admin "retry" executes instantly

Each of these routes now sets maxDuration = 60 so the runner's 45-second time budget always fits.

Sparser crons. AMA work drops from */5 to */30; media reconcile drops from */15 to hourly on the hour, aligned so wakeups consolidate. The cron is now purely the clock fallback for reminders, retry backoff, and expired Slot Hold bookkeeping.

Reviewer notes

• No double-execution risk: the operations runner claims work under leases (claimDue + lease token), so a concurrent inline drain and scheduled sweep cannot run the same operation twice. An interrupted drain's lease expires and the next run reclaims it.
• Correctness never depended on sweep cadence: availability filters expired-but-unreleased holds by their own expiry, and hold→booking conversion transactionally requires expiresAt > now. releaseExpiredHolds is bookkeeping only.
• Accepted trade-off: the 24h/1h session reminders and failed-operation retry backoff can now land up to 30 minutes later than scheduled. Everything guest-facing gets faster.
• Expected Neon impact: 48 short wakeups/day instead of always-on — roughly ~30 CU-hrs/month (~85% reduction).
• Cron schedule changes take effect on the next production deploy.

Verification

• npx tsc --noEmit clean
• pnpm test:ama — 668 tests + 9 migration tests pass
• pnpm test:unit — 1245 tests pass

Greptile Summary

The PR starts AMA durable-operation drains from successful enqueueing mutations and reduces cron frequency to allow the database to autosuspend.

• Adds background, lease-safe draining after booking creation, guest management mutations, and relevant admin actions.
• Repeats full operation batches within a bounded inline budget.
• Avoids draining for Stripe webhook outcomes that enqueue no work.
• Changes AMA work cadence to every 30 minutes and media reconciliation to hourly.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
lib/ama/booking/server.ts	Adds a bounded inline drain that repeatedly runs full operation batches under the existing lease-based runner.
lib/ama/booking/http.ts	Exposes processed webhook outcomes to route wiring without changing signature verification or durable booking semantics.
app/api/ama/stripe/webhook/route.ts	Starts the inline drain only when webhook processing creates a booking and enqueues durable work.
app/api/ama/manage/[token]/cancel/route.ts	Starts durable-operation processing after successful guest cancellation responses.
app/api/ama/manage/[token]/reschedule/route.ts	Starts durable-operation processing after successful guest rescheduling responses.
app/api/admin/ama/bookings/[bookingId]/route.ts	Starts durable-operation processing after successful owner booking actions.
app/api/admin/ama/operations/[operationId]/route.ts	Starts durable-operation processing after successful operation actions such as retry.
vercel.json	Reduces AMA and media cron frequency to consolidate database wakeups.

Sequence Diagram

sequenceDiagram
  participant Mutation as AMA mutation/webhook
  participant Queue as Durable operations
  participant Drain as Inline runner
  participant Provider as Email/Calendar/Stripe
  participant Cron as 30-minute cron
  Mutation->>Queue: Enqueue durable work
  Mutation->>Drain: kickAmaOperations via waitUntil
  loop While batch is full and budget remains
    Drain->>Queue: Claim due work under leases
    Drain->>Provider: Execute side effects
    Drain->>Queue: Complete or schedule retry
  end
  Cron->>Queue: Fallback sweep

Loading

_{Reviews (3): Last reviewed commit: "fix(ama): give the inline drain room to ..." | Re-trigger Greptile}

Context used (3)

• Knowledge Base — AMA booking flow
• Knowledge Base — AMA Admin Operations
• Knowledge Base — AMA Stripe Payments

[github-actions]

Preview deployment ready

Environment	Deployment	Commit
Preview	Visit Preview	26ce340

Branch: perf/ama-inline-operations-drain · View deployment logs