Testing on production data with data branches

Unit tests run your pipeline on DuckDB against fixture data. Sometimes you need more: validating a new transform on the real production volume, or checking that a schema change does not break a MERGE on last month's data. A data branch does that. Set SL_DATA_BRANCH and Starlake redirects every write of the job into a sandbox built from zero-copy clones of your production tables. Production tables are never modified.

note

This page covers branches on BigQuery and Snowflake, driven by SL_DATA_BRANCH. DuckLake databases served by Quack-on-Demand have their own branching, with row-level diffs and reviewed merges: see QoD branching.

Supported engines

Engine	How the branch is built	Where the results land
BigQuery (native transforms and native loads)	Source and target tables are copied with CREATE TABLE ... CLONE, and the task SQL is rewritten to use the copies	Dataset <branch> in the target table's project
Snowflake (transforms and native loads)	The Starlake JDBC proxy driver clones each table with CREATE TABLE ... CLONE the first time a statement modifies it	Schema <branch>
Any other engine	Not supported	Production: the setting is ignored

warning

On PostgreSQL, DuckDB, Redshift and every other engine, SL_DATA_BRANCH has no effect and the job writes to production as usual. Do not rely on a branch there.

Enabling a branch

The branch name is resolved in this order, first match wins:

1. SL_DATA_BRANCH (or sl_data_branch) in the active environment: your shell, metadata/env.sl.yml or metadata/env.<ENV>.sl.yml
2. sl_data_branch (or SL_DATA_BRANCH) in the options of the sink connection. BigQuery only: the Snowflake path reads the environment alone.

An empty value or None disables branching.

From the command line:

SL_DATA_BRANCH=test_order_summary starlake transform --name kpi.order_summary
SL_DATA_BRANCH=test_order_summary starlake load --domains sales

Or for a whole environment, in metadata/env.BRANCH.sl.yml, activated with SL_ENV=BRANCH:

version: 1
env:
  SL_DATA_BRANCH: "test_order_summary"

The branch name becomes a BigQuery dataset or a Snowflake schema, so use letters, digits and underscores only.

BigQuery

When a native BigQuery transform runs on a branch, Starlake:

1. Creates the dataset <branch> in the target table's project, in the location of the sink connection. location is required: the job fails without it.
2. Parses the task SQL and clones every source table it references into <branch>, unless a table with that name already exists there.
3. Clones the existing target table into <branch>, so MERGE and upsert write strategies start from the production rows.
4. Rewrites the presql, main SQL and postsql to read from and write to the clones. The result lands in <branch>.<table>.

A native load on a branch clones the target table and loads the files into <branch>.<table>.

Expectations of a branched transform or native load check the branch copy, the table the job actually wrote. Their reports record the branch name as the domain, so the expectations audit table tells branch runs apart from production runs.

# metadata/application.sl.yml
application:
  connectionRef: "bigquery"
  connections:
    bigquery:
      type: "bigquery"
      options:
        location: "europe-west1"   # required for branching
        authType: "APPLICATION_DEFAULT"

What this means in practice:

• The snapshot is taken on first use. A source table is cloned the first time a task on the branch reads it, and later runs reuse that clone. To pick up fresh production data, use a new branch name or drop the branch dataset.
• Chained tasks see each other's output. When a downstream task reads a table that an upstream task already wrote on the branch, it finds that table in <branch> and reads the branch version, not production. Running a whole DAG on the same branch therefore tests the chain end to end.
• All domains share one dataset. sales.orders and kpi.orders would both map to <branch>.orders. Inside a single task the second one gets a _1 suffix, but across tasks the first table created under that name wins. Keep table names unique across the domains you run on a branch.
• Spark-engine tasks are only half branched. A task that runs on Spark and writes to BigQuery writes to <branch>.<table>, but reads its sources from production: only native BigQuery SQL is rewritten.

Snowflake

On a branch, Starlake routes Snowflake connections through the Starlake JDBC proxy driver, which ships with Starlake: the URL jdbc:snowflake:... becomes jdbc:starlake:snowflake:.... On every new connection, Starlake opens the branch with:

BRANCH START <branch> WITH PERSIST = TRUE, KEEP = TRUE

From then on:

• Copy on write. The first statement that modifies a table clones it into schema <branch>. That statement and all the later ones are rewritten to use the copy.
• Reads of untouched tables hit production. A table the branch never modified is not copied, so reads see live production data, not a frozen snapshot.
• The branch outlives the job. PERSIST keeps the history of the statements run on the branch, and KEEP keeps the copies after the connection closes. A later run with the same branch name picks up where the previous one stopped.

Inspecting the results

On BigQuery, query the branch dataset and compare it with production:

-- rows the branch produces that production does not have
SELECT * FROM `my-project.test_order_summary.order_summary`
EXCEPT DISTINCT
SELECT * FROM `my-project.kpi.order_summary`;

On Snowflake, query <branch>.<table> the same way. Through a connection opened with the Starlake JDBC driver, you can also run:

BRANCH DESCRIBE test_order_summary;  -- statements recorded on the branch, with their timestamps
BRANCH STATUS test_order_summary;

Promoting or discarding a branch

There is no starlake command to apply or drop a branch yet. Do it directly on the warehouse:

Engine	Discard	Promote
BigQuery	Delete the dataset: bq rm -r -f -d my-project:test_order_summary	No apply step: run the job again without SL_DATA_BRANCH
Snowflake	BRANCH DROP test_order_summary through the Starlake JDBC driver: drops the copies and the history	BRANCH APPLY test_order_summary replays the recorded statements on production. Or run the job again without SL_DATA_BRANCH

To use the Starlake JDBC driver from a SQL client, set the driver class to ai.starlake.jdbc.StarlakeDriver and prefix your usual Snowflake URL: jdbc:starlake:snowflake://<account>.snowflakecomputing.com/....

Limitations

• Only BigQuery and Snowflake are supported. Every other engine ignores the setting.
• On BigQuery, branch tables live in a single dataset, so table names must be unique across the domains you run on the branch.
• The connection-level sl_data_branch option is read on BigQuery only.
• Branches are never cleaned up automatically: drop them when you are done, or they keep accruing storage for the rows they change.