bazel test //cli:index_run_e2e_test has been failing on a clean tree for a while, and every release since has recorded it as "the BigQuery credentials fixture". That framing is wrong in a way worth correcting: the credentials are present and they work — //tests/integration:bigquery.spec passes against live BigQuery with the same file.
The actual failure:
sqlanvil encountered an error: Unexpected property "projectId", or property value type of "string" is incorrect.
ReferenceError: Unexpected property "projectId", or property value type of "string" is incorrect.
test_credentials/bigquery.json is the {projectId, location, credentials} wrapper documented in contributing.md. The e2e test hands it to the CLI as a .df-credentials.json, where the strict config validation added in 1.25.1 rejects projectId as an unknown property. The integration spec doesn't hit this because it unwraps the fixture before constructing the adapter.
So it is a test-harness bug, not a credentials problem — either the e2e test should unwrap the fixture the way the spec does, or it should carry its own correctly-shaped credentials file.
Worth fixing rather than living with, for a reason beyond tidiness: a permanently-red test is a permanently-ignored one. Every release currently reports "63 of 64, the known one failed", which is exactly the shape of report that hides a 64th failure the day it appears.
bazel test //cli:index_run_e2e_testhas been failing on a clean tree for a while, and every release since has recorded it as "the BigQuery credentials fixture". That framing is wrong in a way worth correcting: the credentials are present and they work —//tests/integration:bigquery.specpasses against live BigQuery with the same file.The actual failure:
test_credentials/bigquery.jsonis the{projectId, location, credentials}wrapper documented in contributing.md. The e2e test hands it to the CLI as a.df-credentials.json, where the strict config validation added in 1.25.1 rejectsprojectIdas an unknown property. The integration spec doesn't hit this because it unwraps the fixture before constructing the adapter.So it is a test-harness bug, not a credentials problem — either the e2e test should unwrap the fixture the way the spec does, or it should carry its own correctly-shaped credentials file.
Worth fixing rather than living with, for a reason beyond tidiness: a permanently-red test is a permanently-ignored one. Every release currently reports "63 of 64, the known one failed", which is exactly the shape of report that hides a 64th failure the day it appears.