Conventional wisdom suggests Postgres-backed queues don't scale, necessitating dedicated systems like RabbitMQ or Redis. However, with specific optimizations, Postgres can handle demanding queue workloads. This article details how one team achieved 30K workflow executions per second.

Lesson 1: Implement `FOR UPDATE SKIP LOCKED`. This clause prevents contention by allowing workers to lock rows, ensuring only one worker processes a workflow and skipping already locked rows. Without this, scaling is limited to around 100 workflows per second.

Lesson 2: Adjust Transaction Isolation Levels. `REPEATABLE READ` can cause serialization failures at high concurrency. Switching to `READ COMMITTED` for queues without global flow control eliminates these failures, significantly improving throughput.

Lesson 3: Optimize Indexes. Inefficient indexes increase CPU usage. Creating a more selective, partial index for the dequeue query and conditional indexes for observability drastically reduces CPU load and maintenance overhead. These combined optimizations enabled scaling to 30K workflows per second.