Skip to main content
A queued job runs work asynchronously instead of keeping an inference HTTP request open until all work finishes. Tensorfuse separates deploying the worker from submitting individual job payloads.

​
Job lifecycle

  1. Deploy a job worker with the GPU, CPU, memory, secrets, and maximum scale needed by the application.
  2. Queue a job using the deployed job name, a job ID, and a payload string.
  3. Read the queued payload inside the container with the Tensorfuse SDK. If the payload contains JSON, parse it before using its fields.
  4. Poll the job status using the same job name and job ID.
Keep the submitted job ID in your application so status checks can identify the execution. Design any writes to your own storage to tolerate retries; a queue alone does not establish exactly-once side effects. The job queues guide documents submission, payload access, and polling. Use the job CLI reference for options. For synchronous HTTP inference, start with deployments.
Tensorfuse jobs architecture diagram