Zero-Downtime Patching in Lakebase Part 1: Prewarming

This post explains how Lakebase eliminates performance degradation during planned database maintenance using a prewarming technique.

• •Traditional PostgreSQL restarts lose in-memory buffer caches, causing ~70% reduction in pgbench TPS during cache warmup after restart
• •Lakebase uses stateless, elastic compute nodes with disaggregated shared storage, enabling new nodes to be spun up without affecting the primary
• •Before a scheduled restart, a new compute node is started in the background, receives the current primary's cache page list, and loads those pages from shared storage
• •The new compute subscribes to WAL updates via Safekeepers to stay current without adding load to the primary
• •After prewarming completes, the old primary is shut down and the new compute is promoted via pg_promote at no additional replica cost