Prefect

On GKE Autopilot, mapped child tasks sometimes start several times, even after the original has succeeded, leaving flow runs in inconsistent states. Hours-long tasks make it costly.

Affects: teams running flows on GKE Autopilot

On 3.4.20 a flow whose tasks all finished ends as Crashed after a lease renewal failure terminates execution.

After the 0.14.7 release, starting the local Prefect server fails because docker-compose rejects the depends_on entries.

Affects: users starting the server locally

A Helm-deployed Prefect server on EKS with RDS Postgres started hitting frequent deadlock errors in the message consume loop after the upgrade.

Affects: teams running Prefect server on EKS with RDS

A user who otherwise likes Prefect's design finds that large task inputs and outputs stay in memory even after deleting them, unless result persistence is disabled on the task.

Affects: data engineers passing large data frames

Multi-infrastructure workflows force a separate deployment per flow plus run_deployment, causing deployment sprawl and brittle coupling to server-side names. The author wants direct submission to remote infrastructure.

Affects: teams running workflows across several infrastructures

An automation that reruns crashed flows loops forever on a deterministic crash, and the flow retries setting does not apply to crashes. The user wants retries respected on Crashed state with a maximum count.

Deploying flows fails unless all dependencies of every flow are present, which is a problem with per-flow virtual environments and makes CI/CD duplicative of a Docker build.

Affects: teams deploying multiple flows from CI/CD

When an agent goes down and restarts, all missed late runs are picked up immediately. Users want to filter runs by state, tag or queue and bulk cancel or delete them.

Affects: users with hourly scheduled deployments

A team that bakes flow code into its container image wants module-path storage like in Prefect 1.0, rather than needing external S3 or Git storage that adds failure points.

Affects: teams packaging flows in Docker images on Kubernetes

Deployments sometimes get two scheduled runs with identical start times, apparently after maintenance or redeployments. The team runs a cleanup script that toggles schedules to fix it.

Affects: teams on Kubernetes with Helm

Each Prefect flow starts in its own ECS task, so slow container startup on Fargate added up. The team moved to an EC2 autoscaling group with a custom AMI that has images pre-baked, cutting startup from about 3 minutes to 30 seconds.

Affects: teams running flows on AWS ECS

The server was seen running out of memory and restarting on Kubernetes. Doubling CPU and memory limits eased the problem.

Affects: team running Prefect server on Kubernetes

Launching two concurrent calls to the same subflow raises an error that the task runner is already started. A staff member suggested using a task instead.

Affects: users running parallel subflows

A team that built a large business on Temporal says it feels much better than the orchestrators they used before, Prefect and Airflow, and they would adopt it again.

Affects: engineering team building business-critical applications