Event-Driven Notification Platform · View 20 of 26 · 5 · Operations

Autoscaling, Capacity and Backpressure

What each tier scales on, where scaling stops, and what the platform does once it has stopped.

Editable source SVG draw.io All views

Capacity targets

5 000 events/s normal, 25 000 peak, sustained for 30 minutes
Fan-out 1.8, so 45 000 notifications/s at peak and about 750 M a day
Maximum event size 64 KB hard, 4 KB at p99 — events carrying more than that are carrying data that belongs behind a reference

Where scaling stops

Partition count is the concurrency ceiling, which is why partitions are provisioned for peak rather than for today
Provider TPS is contracted, not elastic — beyond it, queueing is the only option
Database connections are bounded by PgBouncer, deliberately, so a scale-out event cannot exhaust PostgreSQL

Backpressure order

Shed bulk traffic first — marketing pauses before anything transactional degrades
Then 429 with Retry-After at the gateway, per tenant, so the producer learns rather than times out
Park to the DLQ only as a last resort, because everything parked has to be replayed and deduplicated later

◀ Delivery Pipeline and Environments All views Observability and Operations ▶