Azure Machine Learning compute provided on demand whenever a job names no compute target. No cluster needs creating or managing, and jobs do not wait in a queue behind one another.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Serverless compute in context, with comparison tables and the common traps.
Terms in this definition
- Azure Machine Learning
Azure platform where you train, deploy, monitor and retrain models of your own, practising MLOps through tools such as pipelines, online endpoints, model monitoring and prompt flow.
- Job
A sequence of steps executed together on one agent or runner, or on the server for agentless work. While running, each occupies one of your parallel jobs.
Related terms
- Apache Spark cache
A cache you request explicitly, in memory or on disk, for a table or DataFrame using
df.cache(),df.persist()orCACHE TABLE, and release withunpersist. It differs from the automatic disk cache and cannot be used on serverless compute. - Autoscale Billing for Spark
An opt-in way of paying for Spark (Learn also calls it on-demand billing) where jobs use separate serverless compute rather than drawing on the Fabric capacity, and you pay only while jobs are active, capped by a maximum CU setting. Smoothing and bursting play no part, and at the cap new batch jobs wait in a queue while interactive ones are throttled.
- Azure Functions
Serverless compute for event-driven code, started by triggers such as HTTP, timer, Blob or Event Grid and hosted on a Consumption, Premium or Dedicated plan.
- Classic compute
All-purpose, jobs and pipeline compute running in your own Azure subscription, which you set up and manage yourself. Its counterpart is serverless compute, which Azure Databricks runs for you.
- Compute metrics
Hardware, Spark and GPU figures sampled once a minute and kept for 30 days, shown on the Metrics tab of classic all-purpose and jobs compute. Query insights replace it on serverless compute.
- Predictive optimization
Unity Catalog managed tables can have
OPTIMIZE,VACUUMandANALYZEscheduled and run for them on serverless compute without manual effort. Accounts created on or after 11 November 2024 get this turned on by default. - Query History
A page in Databricks SQL where you can filter through past queries from SQL warehouses and serverless compute and open the profile of any of them.
- RDD
Short for Resilient Distributed Dataset, the older low-level Spark interface for working with distributed collections. Serverless compute and standard access mode don't allow it, so code relying on RDDs needs dedicated compute.