Describe the current behavior
Running Prefect flows on Nebius Serverless Jobs currently requires custom integration code to submit containers, monitor jobs, collect logs and propagate cancellation. Users must also handle ambiguous submission responses and reconnect to jobs after a worker restart.
There is no Nebius-specific work-pool worker in the Prefect repository.
Describe the proposed behavior
Provide a prefect-nebius integration with a nebius-serverless worker. Users would deploy a containerized flow to a work pool configured with their Nebius project, subnet, GPU platform, preset and image.
The worker would submit the job, monitor its lifecycle, forward logs and handle cancellation. It would preserve submission identity across restarts and reconcile uncertain submissions without automatically creating another job. The Prefect engine would remain responsible for reporting the flow’s application outcome.
This follows Prefect’s existing infrastructure-worker model, similar to the Google Cloud Run V2 integration, without requiring changes to Prefect core.
Example Use
Schedules a nightly GPU batch-inference flow in Prefect. The Nebius worker launches its container on the selected GPU, forwards logs and handles cancellation. After a worker restart, it reconnects to the existing job instead of submitting a duplicate.
Additional context
No response
Describe the current behavior
Running Prefect flows on Nebius Serverless Jobs currently requires custom integration code to submit containers, monitor jobs, collect logs and propagate cancellation. Users must also handle ambiguous submission responses and reconnect to jobs after a worker restart.
There is no Nebius-specific work-pool worker in the Prefect repository.
Describe the proposed behavior
Provide a prefect-nebius integration with a nebius-serverless worker. Users would deploy a containerized flow to a work pool configured with their Nebius project, subnet, GPU platform, preset and image.
The worker would submit the job, monitor its lifecycle, forward logs and handle cancellation. It would preserve submission identity across restarts and reconcile uncertain submissions without automatically creating another job. The Prefect engine would remain responsible for reporting the flow’s application outcome.
This follows Prefect’s existing infrastructure-worker model, similar to the Google Cloud Run V2 integration, without requiring changes to Prefect core.
Example Use
Schedules a nightly GPU batch-inference flow in Prefect. The Nebius worker launches its container on the selected GPU, forwards logs and handles cancellation. After a worker restart, it reconnects to the existing job instead of submitting a duplicate.
Additional context
No response