Databricks variant shredding will lock Runtime 15.4 out of new tables

Databricks variant shredding will lock Runtime 15.4 out of new tables

12 October 2026: Databricks will start adding the variantShredding table feature automatically on 18 November to every new table with a VARIANT column created on Databricks Runtime 17.3 or later, and clusters still on Runtime 15.4 cannot read from or write to those tables at all. Databricks lists it as a breaking change and tells customers to move affected clusters to Runtime 16.4 or later first.

Not yet independently verified. Single source: Databricks' own What's coming page, which it edits in place, so the dates can move. The OpenSharing URL transition date is not yet set. We will update this when it can be confirmed, and remove this note.

The notice sits on the company's "What's coming" page, last updated on 12 October, alongside three other changes with fixed dates that need someone in a Databricks account to act.

Why one runtime version breaks

Variant shredding stores the fields a query reads most often in a typed, columnar layout instead of inside the semi-structured VARIANT blob, which Databricks says speeds those queries up. The catch is that the table now carries a feature older readers do not understand. Runtime 15.4 is a long-term support release, so it tends to linger on scheduled jobs, shared clusters and the compute that external tools connect through.

The change applies only to tables created after 18 November on 17.3 or above. Existing tables are not rewritten. The risk is therefore a mixed estate: a pipeline on a newer runtime creates a table, and a downstream job or dashboard still pinned to 15.4 fails when it tries to read it. Finding every 15.4 cluster that reads tables written by newer runtimes is the work to do before the date.

The other dates on the page

• 1 November 2026: the account and workspace toggle for partner-powered AI features is removed. Each workspace keeps whatever value it has, enabled or disabled, but after that date changing it means contacting the Databricks account team. A team that wants these features off should set that now.
• 2 December 2026: CREATE OR REPLACE VIEW starts updating views in place rather than dropping and recreating them, so grants, tags and comments survive. Dropping a column that carries a governed tag through the statement will then fail with a CANNOT_DROP_TAGGED_COLUMN error, to protect attribute-based access policies built on the tag. The NOCOPY clause drops grants, tags and comments on purpose.
• 8 December 2026: OpenSharing bearer tokens created before the move to the new recipient-specific URL format expire one year from creation or on 8 December 2026, whichever is later. Data providers have to rotate tokens to move recipients onto the new format, and Databricks suggests setting an expiry on the current token during rotation to give recipients a downtime window.

Smaller changes for compliance workspaces

Workspaces with the compliance security profile enabled get several features switched on by default over the next two months: workspace-level package repositories for Databricks Apps from 13 November, and Feature Views, JDBC Unity Catalog connections and the remote_query function in November. The READ METADATA privilege arrives in those workspaces in mid-October; it shows grants, row filters, column masks and ABAC policies, so it is a privilege to hand out deliberately.

Separately, Databricks Runtime 20 stops bundling the BigQuery JDBC driver. Spark JDBC jobs using a jdbc:bigquery: URL will fail with a driver-not-found error unless a driver is installed before the upgrade.

Where Databricks sits in our catalogue

We list Databricks among the 29 products in our data warehouse software category, next to Databricks SQL and Snowflake. The OpenSharing token date is the one most likely to reach outside the company, because the recipients holding old tokens are often partners and customers rather than internal teams.

Related on TrustList:

• Power BI Pro and Premium Per User licences will include Fabric apps and Fabric Database at no extra cost, from a preview in the coming weeks

Sources

• Databricks: What's coming (page last updated 12 October 2026), 12 October 2026