Uploaded November 2026 | Updated November 2026, 15 hours ago
Databricks Lakebase autoscaling lets serverless Postgres size its own compute: you set a range, and Lakebase scales up for a rush, back down when it's quiet, and to zero when nobody is using it. In this video we test that live on a coffee shop database.
β° Chapters
00:00 - Lakebase Autoscaling Explained: Compute That Follows the Load
00:50 - Set the Autoscaling Range in Edit Compute
01:30 - Morning Rush: Latency Drops as Compute Scales Up
02:30 - 16 to 48 Clients: Hitting the 8 CU Maximum
03:19 - Scaling Down After the Rush and Scale to Zero Overnight
04:33 - Choosing Autoscaling Settings for Dev and Production
π How Lakebase autoscaling works
Lakebase separates compute from storage, so the compute can change size without moving your data. You set a minimum and a maximum in compute units (about 2 GB of memory each, with the CPU to match), up to 16 units apart, and Lakebase picks its size inside that range from CPU, memory and how much of your data is in use. With scale to zero on, the compute suspends after a timeout and wakes up on the next query, so an idle database costs nothing for compute.
π οΈ What you'll see in the demo
Using a coffee shop project with a range of 1 to 8 compute units:
β’ Setting the autoscaling range in Edit compute, and how the 16-unit spread moves the minimum when you raise the maximum
β’ A morning rush with pgbench: at 4 clients, latency falls from about 1.5 seconds to 0.4 seconds and throughput more than quadruples as compute is added
β’ 16 and 48 clients: throughput levels off at about 25 requests per second at the 8 CU maximum, with CPU maxed
β’ A probe that queried once a second through the whole rush: 165 out of 165 succeeded, no restarts and no dropped connections
β’ The Monitoring tab: allocated memory peaks at 17 GB, then returns to the 2 GB minimum after the rush
β’ Scale to zero with a 60-second timeout, and 5 timed wake-ups: the first query came back in 2.35 to 2.8 seconds, median 2.46 seconds
β’ How to choose settings for dev, test and production, including the always-on rate for a steady baseline
Whether you're a developer paying for an idle dev database, a platform engineer sizing Postgres for spiky traffic, or a team lead watching database costs, Lakebase autoscaling lets one database handle the rush, the quiet afternoon and the empty night without anyone changing a setting.
π Resources
π Lakebase autoscaling documentation: docs.databricks.com/aws/en/oltp/projects/autoscaling
π€ Scale to zero documentation: docs.databricks.com/aws/en/oltp/projects/scale-to-zero
π Databricks Lakebase: databricks.com/product/lakebase
π€ Stay connected
Website: databricks.com
Twitter: https://x.com/databricks/
LinkedIn: linkedin.com/company/databricks
Instagram: instagram.com/databricksinc
Facebook: facebook.com/databricksinc
Databricks Lakebase autoscaling lets serverless Postgres size its own compute: you set a range, and Lakebase scales up for a rush, back down when it's quiet, and to zero when nobody is using it. In this video we test that live on a coffee shop database.
β° Chapters
00:00 - Lakebase Autoscaling Explained: Compute That Follows the Load
00:50 - Set the Autoscaling Range in Edit Compute
01:30 - Morning Rush: Latency Drops as Compute Scales Up
02:30 - 16 to 48 Clients: Hitting the 8 CU Maximum
03:19 - Scaling Down After the Rush and Scale to Zero Overnight
04:33 - Choosing Autoscaling Settings for Dev and Production
π How Lakebase autoscaling works
Lakebase separates compute from storage, so the compute can change size without moving your data. You set a minimum and a maximum in compute units (about 2 GB of memory each, with the CPU to match), up to 16 units apart, and Lakebase picks its size inside that range from CPU, memory and how much of your data is in use. With scale to zero on, the compute suspends after a timeout and wakes up on the next query, so an idle database costs nothing for compute.
π οΈ What you'll see in the demo
Using a coffee shop project with a range of 1 to 8 compute units:
β’ Setting the autoscaling range in Edit compute, and how the 16-unit spread moves the minimum when you raise the maximum
β’ A morning rush with pgbench: at 4 clients, latency falls from about 1.5 seconds to 0.4 seconds and throughput more than quadruples as compute is added
β’ 16 and 48 clients: throughput levels off at about 25 requests per second at the 8 CU maximum, with CPU maxed
β’ A probe that queried once a second through the whole rush: 165 out of 165 succeeded, no restarts and no dropped connections
β’ The Monitoring tab: allocated memory peaks at 17 GB, then returns to the 2 GB minimum after the rush
β’ Scale to zero with a 60-second timeout, and 5 timed wake-ups: the first query came back in 2.35 to 2.8 seconds, median 2.46 seconds
β’ How to choose settings for dev, test and production, including the always-on rate for a steady baseline
Whether you're a developer paying for an idle dev database, a platform engineer sizing Postgres for spiky traffic, or a team lead watching database costs, Lakebase autoscaling lets one database handle the rush, the quiet afternoon and the empty night without anyone changing a setting.
π Resources
π Lakebase autoscaling documentation: docs.databricks.com/aws/en/oltp/projects/autoscaling
π€ Scale to zero documentation: docs.databricks.com/aws/en/oltp/projects/scale-to-zero
π Databricks Lakebase: databricks.com/product/lakebase
π€ Stay connected
Website: databricks.com
Twitter: https://x.com/databricks/
LinkedIn: linkedin.com/company/databricks
Instagram: instagram.com/databricksinc
Facebook: facebook.com/databricksinc










