Databricks ML Professional Practice Question 20: Scale to Zero Cold Start
Databricks Certified Machine Learning Professional · Test 2 · question 10 of 10
An internal endpoint serves a few requests an hour during business hours and none overnight, and it was created with scale to zero enabled. Users report that the first request each morning takes roughly 15 seconds, after which every request returns in well under a second. The business has now said the first request of the day must also be fast.
What explains the delay, and what should the team change?
Show answer and explanation
Answer:
Scale to zero releases an endpoint's compute after about 30 minutes without traffic, which is why it costs nothing overnight. The first request afterwards has to wait for the endpoint to scale back up, typically 10 to 20 seconds and occasionally longer; the 15-second morning delay followed by fast responses is the textbook signature. Because capacity is not guaranteed while scaled to zero, Databricks advises against the setting for endpoints with a latency requirement. Turning it off keeps compute provisioned and removes the cold start, at the cost of paying for idle time overnight, which is the trade the business has just chosen.
Option B is contradicted by the evidence. The same model scores every subsequent request in under a second, so neither its size nor its feature count explains a delay that occurs exactly once.
Option C addresses throughput rather than start-up. Workload size sets provisioned concurrency, which matters when many requests arrive at once; a few requests an hour never saturates Small, and a larger size still starts from zero.
Option D describes real but negligible overhead. A TLS handshake costs milliseconds, not fifteen seconds, and it would recur whenever a connection was re-established rather than once each morning.
Reference: Create custom model serving endpoints