fix(cubejs): BigQuery rollups read-only; failed rollup downloads no longer crash the refresh worker

[acmeguy]

Why

The refresh worker restarted 33 times in under four days. Each crash stopped rollup builds for every team. The cause was one tenant's BigQuery rollups (Abler):

• abler-web (BQ QA): Cube stages every rollup partition as a table in the customer's BigQuery project before loading it into Cube Store. Its service account may not create that dataset (bigquery.datasets.create), so none of its rollups ever built.
• abler-web-prod (BQ Prod): 2,045 rollups are in Cube Store, but some builds lost their staging table mid-load. Cube 1.7.30 pipes BigQuery's stream through two pipe() calls that never forward errors, so a "Not found: Table" became an uncaughtException, and the process guard exits on those.

What

1. BigQuery datasources are read-only (prepareDbParams). Cube runs the rollup query and loads the result straight into Cube Store, with no staging tables in the customer's project.
   • Rollup table versions don't depend on readOnly, so existing Cube Store tables stay current and nothing mass-rebuilds.
   • Each partition's result is held in worker memory. Abler's largest is 54,796 rows.
2. Stream errors are forwarded (postinstall patches, alongside the existing YAML and ClickHouse ones). One patch covers the BigQuery driver's stream(), the other the loader's LargeStreamWarning wrap at both of its sites. A failed download now fails only that build.
3. Docs: cube principles on SQL wrapping, ClickHouse map keys, refresh cadence and the Cube Store Nullable(Bool) limit.

Tests

• Loader path (cube17Regression): the test runs the real BigQuery driver through PreAggregationLoader.getTableDataWithTempTable. Before the patches it failed with uncaughtException; with them it is a handled rejection.
• Drift guards: both new patches refuse to apply if Cube's source changes.
• Read-only (prepareDbParams): a BigQuery driver built from the prepared settings reports readOnly().
• Full suite, locally: the only failures are the 2 dependency-hardening checks against a stale local install. This PR's image build runs the full suite from the lockfile.

Rollout

After merge, restart synmetrix-cubejs and synmetrix-cubejs-refresh-worker together, so the API and the worker hash the BigQuery datasource the same way. Then watch:

• the worker's restart count;
• Abler's rollup tables in Cube Store;
• other tenants' build errors.

🤖 Generated with Claude Code