retired Public Preview
Legacy MLflow Model Serving
Legacy MLflow Model Serving is retired, replaced by Model Serving (retired September 2025).
Legacy MLflow Model Serving Model Serving
The original 2020 Databricks model serving, which hosted MLflow models from the Workspace Model Registry as REST endpoints on a single-node cluster in your own account, redeploying automatically as model versions and stages changed.
- It was serverless in exactly no sense - every served model got its own single-node cluster in your account, billed as an all-purpose cluster, and it kept running as long as serving was enabled even with no active model version to serve.
- It never left Public Preview, and its docs set a target of 200 queries per second and 99.5% availability 'although no guarantee is made as to either'.
- Endpoints were addressed by Model Registry stage as well as by version, so a URL ending in /Production/invocations repointed itself automatically whenever someone promoted a new version.
Limitations: Python MLflow models only, with every dependency declared in the model's conda environment; a 16 MB payload limit per request; targets of 200 qps and 99.5% availability with no guarantee on either; global init scripts did not run on the serving cluster.
Open in REbricked →- Category
- AI / ML
- Also known as
- MLflow Model Serving, Legacy model serving, Model Serving (legacy)
- Verified
- 2026-09-11