Neon and Databricks Lakebase are working toward zero-downtime database patching. The first technique is prewarming: before a planned restart, a new Postgres compute node is spun up in the background, its cache is populated with the current primary's page list from shared storage, and it subscribes to the WAL to stay current. When ready, the old primary is shut down and the new node is promoted without a full restart. Benchmarks with a 10 GB pgbench workload show throughput recovers nearly instantly with prewarming, versus a ~70% TPS drop during cold-cache warmup without it. This is now enabled by default for all read/write endpoints at no extra cost, leveraging the lakebase architecture's separation of stateless compute and disaggregated storage.
Sort: