Short answer: with defaults, a Prefect flow run typically starts within a few seconds to low tens of seconds after you trigger it. The biggest factors are your worker’s poll interval and your infrastructure’s cold-start time. What determines startup latency - Prefect orchestration: near-instant (milliseconds to sub-second to schedule/assign). - Worker polling: by default workers poll for new work every ~10 seconds. That means average wait ~5s, worst-case ~10s before the worker grabs the run. - Infrastructure spin-up: - Process worker (same machine): usually sub-second to a couple seconds if the environment is warm. - Docker worker: 2–5s if the image is already on the host; if the image must be pulled, startup can jump to tens of seconds or minutes (depends on image size/registry). - Kubernetes worker: 5–20s on a warm cluster with cached images; 30–120s+ if the image needs pulling or if the cluster/nodes must scale up. Very large images or autoscaling can push this further. - Code and environment setup: cloning source, importing heavy libraries, and any runtime initialization can add seconds. Relevant defaults in Prefect 3.x - Worker poll interval: 10 seconds (controls how often workers query for new runs). - Prefetch window: 10 seconds (workers will pick up scheduled runs a bit early to help them start on time). You can see these in the Worker settings: - Settings reference - Health/worker behavior: Worker healthchecks How to reduce cold start time - Lower the poll interval: - Set

PREFECT_WORKER_QUERY_SECONDS=1

or

2

to cut pick-up delay to ~1–2s. - Increase prefetch for scheduled runs: -

PREFECT_WORKER_PREFETCH_SECONDS=30

so workers grab runs early and start right on time. - Keep infra warm: - Process worker for fastest starts when running on the same machine. - Docker/Kubernetes: pre-pull images, keep nodes warm, avoid scale-to-zero if you need fast response. - Use small images and local/nearby registries; pin to digests for caching. - In Kubernetes, use

IfNotPresent

image pull policy and consider a DaemonSet to pre-pull. - Trim runtime setup: - Avoid heavy Git clones or do shallow clones; minimize expensive imports at module import time; cache credentials and connections where appropriate. Example: start a worker tuned for fast pickup

PREFECT_WORKER_QUERY_SECONDS=1 PREFECT_WORKER_PREFETCH_SECONDS=30 prefect worker start -p my-pool

If you share which worker type you’re using (process, Docker, or Kubernetes) and what you’re seeing now, I can suggest a tighter plan to get startup as close to “instant” as your environment allows.