Asynchronous Tasks

Celery tasks for the llmhost app.

• launch_llmhost(): launch an LLMHost, and create the DNS record of its Ingress.

• destroy_llmhost(): destroy an LLMHost’s Kubernetes resources, and its DNS record.

• refresh_llmhost_status(): check the status of one, or every deployed, LLMHost. Celery Beat runs it every few minutes, so that status, events and cost stay current.

• reconcile_llmhost_compute(): give a compute’s node group the nodes that its LLMHosts need. It runs again every 30 seconds until the node group is settled, i.e. while a node starts or is removed, and checks the status of the compute’s LLMHosts each time.

• charge_llmhost_computes(): charge an hour of each compute’s running nodes, so that budgets attached to a compute, its owner, or their account see the cost of the infrastructure.

• reconcile_llmhost_computes(): reconcile every compute. Celery Beat runs it every few minutes, so that a node that is no longer needed is removed even if a reconcile failed.

The tasks are thin: LLMHostService does the work.

Every task except charge_llmhost_computes() manages Kubernetes resources, which can take minutes, so they run in the infrastructure queue, and never block operational tasks. The charge is operational: it runs in the queue of the other operational tasks, such as charges and budgets.