Best Practices for Lakebase
Overview
| Experience | In Person |
|---|---|
| Track | Lakebase |
| Industry | Enterprise Technology |
| Technologies | Lakebase |
| Skill Level | Advanced |
This deep-dive session covers architectural and operational best practices for building production-grade applications on Lakebase. You'll learn how to design schemas and connection patterns for performance, how to leverage autoscaling and branching for cost-efficient development workflows, and how to apply governance controls at scale using Unity Catalog. Whether you're migrating an existing PostgreSQL workload or building something new, you'll leave with actionable guidance for running Lakebase reliably in production.
Session Speakers
Som Natarajan
/Solution Architect
Databricks
Yanic Motte
/Senior Specialist Solutions Architect
Databricks
Full Summary
Lakebase in production: rethinking Postgres habits for a disaggregated world
Running Postgres at scale creates muscle memory, from avoiding truncates to obsessing over vacuum and connections. Lakebase, Databricks' managed Postgres, changes those instincts because it separates stateless compute from durable, multi‑AZ storage. Compute can autoscale and even scale to zero, while shared storage enables instant branching, snapshots, and point‑in‑time recovery. Architects and platform teams moving beyond POCs on the Databricks Platform need to optimize for that reality.
FAQ
No. Branches are intended as short‑lived environments for testing, bug fixes, and risky operations. Promote changes through your normal deployment pipeline, and use point‑in‑time recovery or snapshots when you need to restore or compare states.
Default to snapshot sync for most workloads because it copies in parallel and swaps partitions atomically. Choose continuous sync only when near‑real‑time updates have clear business value, since each continuous pipeline typically costs about 18 to 40 dollars per day and expenses grow with table width and count.
Per‑request connection patterns quickly exhaust limits on small autoscaling instances. Route traffic through the built‑in PgBouncer pooled endpoint or implement app‑side pooling, and remember that observability tools and system processes also consume connections.
No. TimescaleDB is not included, and it is unlikely to be added. For time‑series use cases, use native partitioning with TTL and keep cold partitions in Delta Lake while serving hot data from Lakebase.
Not today. Unity Catalog governs sync table metadata, but roles and grants inside Lakebase are managed separately. Plan to administer entitlements in both systems until unified governance is available.