DataFrame caching entered Public Preview on Databricks serverless compute in October 2026

The clearest product change for serverless in October 2026 was DataFrame caching. Community posts reported that df.cache() and df.persist() now work on serverless compute and entered Public Preview [2]. Those calls had previously not worked there. The same week, Python asynchronous TransformWithState for streaming was reported as generally available [5].

Networking and platform direction also drew attention. A community post titled "Simplifying networking for serverless" appeared mid-month [10]. A video roundup of recent updates listed serverless version 6, a private network gateway, Lakeflow Jobs changes, task value loops, Databricks Apps and Unity Catalog skills [8]. The summary of that video gave no further detail on the serverless items beyond naming them, so the specifics should be checked against official release notes. A separate video covered Databricks express setup, which lets people launch a managed workspace with a time and credit limited trial and no cloud credentials, with serverless compute and sample data available immediately [11]. On the tooling side, Databricks CLI release v1.22.0 added support for the `features` resource type, let `bundle deployment bind` and `unbind` accept `type.key` references, and removed the 50-workspace limit in `auth login` [6].

Most community discussion concerned cost and availability. Two posts raised billing for managed ingestion: one reported a serverless compute charge on a managed ingestion pipeline although the 100 DBUs per day threshold had not been reached [3], and another asked how to cost a Lakeflow Connect integrated CDC pipeline [4]. Another user reported that serverless compute was not available in their account or workspaces [13]. Other posts covered tag propagation [9], vacuumable files with little recovered space [12], compute management frustration [7] and early impressions of Databricks Apps [1].