GeezerVison

Error telemetry copied the raw server error message and the JS stack into the emitted telemetry.error event and into the exported error_info.stack_trace. Server messages quote the failing SQL's identifiers (for example The table or view `sensitive_table` cannot be found), so table and column names left the customer process. Separately, statement.complete never reported resultFormat or compressed. The format was read from facade metadata that is only set when the application calls getMetadata(), and compressed was not passed at all.

Error telemetry now carries only the error class name and the server SQLSTATE. This matches the kernel's sanitized error_info (error_name + SQLSTATE in stack_trace, from databricks-sql-kernel#272). The emitter, aggregator and exporter each drop message and stack, so in-process listeners and the export see the same descriptor. On Thrift, statement.complete now takes resultFormat and compressed from result metadata the backend already holds (direct results, or the fetch the result handler already made) through a new optional IOperationBackend.getCachedResultMetadata(). This adds no RPC. compressed mirrors the server-declared lz4Compressed.

Verification: new unit tests cover the emitted and exported payloads. On origin/main, 15 new or updated tests fail; on this branch the full unit suite passes (1545), and lint, type-check and build are clean. Live Thrift run against a warehouse, with the exporter stubbed to a local capture so nothing was sent:

• SELECT 1: INLINE_ARROW / false.
• catalog_returns: EXTERNAL_LINKS / false, matching wire lz4Compressed=false.
• Inline result with useCloudFetch: false: INLINE_ARROW / true, matching wire lz4Compressed=true.
• Failing SELECT secret_token FROM sensitive_table, sync and async: OperationStateError + 42P01, and the canaries appear in neither the emitted events nor the export.

origin/main on the same run left the delivery fields unset and exported Error: [TABLE_OR_VIEW_NOT_FOUND] The table or view `sensitive_table`…. The Thrift e2e specs (operation_error_codes, arrow, cloudfetch, batched_fetch, data_types, iterators, query_parameters, timeouts) pass with telemetry captured locally, and every captured stack_trace is a SQLSTATE or empty.

Performance: the change only runs once per statement at close, and once per error, when telemetry is on. In a micro-benchmark (5 alternating fresh-process runs, median), close() went from 4.54 to 4.48 µs/op, which is noise, and error emission went from 10.58 to 0.85 µs/op because the regex redaction of message and stack is gone.

Not addressed here: the [sea] legs of TELEMETRY-001/002/005. Since #506 the driver does not initialize JS telemetry when useKernel is set. The Rust kernel owns statement telemetry there, and the pinned KERNEL_REV already includes databricks-sql-kernel#272, the reference fix. A live kernel run confirmed the JS layer emits nothing.

Addresses #508

This pull request and its description were written by Isaac.