Model Serving Endpoint
Does ZopNight manage Model Serving Endpoint?
Provisioned model serving endpoints bill for their compute the entire time they are ready, whether they receive 1,000 requests or none. Scale-to-zero protects an endpoint only when every served model enables it, since 1 pinned model keeps the whole endpoint warm. ZopNight flags provisioned endpoints that cannot release compute when idle.
Rules that fire on Model Serving Endpoint
A model serving endpoint hosts models for real-time inference on provisioned compute, and provisioned is the operative word: the endpoint bills while it stands ready, whether it serves heavy traffic or none. Serving is where “we might need it later” translates most directly into a standing bill, because an endpoint nobody invokes produces no signal that it exists.
Billing follows readiness, not requests
Once provisioned, the endpoint’s compute is held continuously on your behalf. The workload size sets how much is held; request volume decides nothing about whether the meter runs. An endpoint that served its last prediction in March bills the same as one serving production traffic today.
Scale-to-zero is all-or-nothing
An endpoint can release compute when idle only if every served model behind it enables scale-to-zero. A single pinned model keeps the entire endpoint warm, so discovery reports scale-to-zero as enabled only when all served models allow it. That is the cost-relevant reading, since one exception makes the endpoint effectively always-on regardless of what the other models are configured to do.
What inventory includes and skips
Only endpoints serving custom models on dedicated compute appear. Foundation Model APIs and external-model proxies are pay-per-token system endpoints: they hold no provisioned compute, cannot be started or stopped, and generate no recommendation, so listing them would be noise. Endpoint readiness maps to inventory status (READY as running, NOT_READY as stopped), and the type is discovery-only: ZopNight never schedules serving endpoints.
The experiment graveyard
The fastest-growing category of idle ML spend is endpoints provisioned for a demo or an evaluation and never torn down. With no invocations there is no error, no alert, no log line. Nothing draws attention. An inventory listing every provisioned endpoint per workspace, with its workload size and scale-to-zero posture, is what makes the graveyard visible.
Audit scale-to-zero across endpoints
databricks serving-endpoints list --output jsonFor each endpoint, inspect the served models’ scale_to_zero_enabled flags. Any endpoint where even one flag is false holds warm compute 24 hours a day. Either enable scale-to-zero on every served model or delete what the experiment left behind.