Checkpointing every step sounds expensive. If it were, nobody would wrap small calls, and the benefit would vanish.
Where the time goes
A step start involves a state write, a scheduling decision and a network hop to your code. We batch the first two and keep a warm connection for the third.
Region-local state
Step state lives in the same region as the executor. Since v4.0.3, median start latency in eu-west is 41 ms and p99 overhead is under 50 ms in every region.
What we measure
We publish our latency dashboards on the status page, broken down by region and step type.
