Model Deployment – Databricks ML Professional Study Guide
This domain accounts for 12% of the exam — the smallest slice, but a focused one. It covers the deployment paradigms available on Databricks (batch, streaming, real-time serving), custom model serving, and controlled rollout strategies. The core skill is matching a paradigm to the workload's latency/throughput/cost needs and knowing how Model Serving supports safe rollouts.
Deployment Paradigms
| Paradigm | Use for | How |
|---|---|---|
| Batch scoring | Large volumes, no latency requirement (e.g., nightly) | A scheduled Job applies the model and writes to Delta |
| Streaming inference | Continuous events scored as they arrive | Structured Streaming applies the model to a streaming source |
| Real-time serving | Sub-second, per-request predictions | Databricks Model Serving REST endpoint |
Choosing is driven by required latency, throughput/volume, and cost: overnight bulk → batch; continuous event streams → streaming; sub-second per request → real-time serving.
Databricks Model Serving
- Exposes a registered model as a low-latency REST endpoint for real-time predictions.
- Serve by alias (e.g.,
@champion) so promoting a new version repoints the alias and updates what the endpoint serves — no code or endpoint recreation. - Deployments are governed and traceable to a registry version/alias.
Unlock with a paid plan
3 more sections (about 216 words) continue below. Every paid plan unlocks all Machine Learning Professional study guides, the full question bank and mock exams, from $4.97/month.