[BEAM-13132] WriteToBigQuery submits a duplicate BQ load job if a 503 error code is returned from googleapi

• Type: Bug
• Status: Open
• Priority: P1
• Resolution: Unresolved
• Affects Version/s: 2.24.0
• Component/s: io-py-gcp
• Labels: GCP
• Environment: Apache Beam Python 3.7 SDK 2.24.0

Language

Python

• Assignee: Unassigned
• Reporter: James Prillaman
• Created: 27/Oct/21 19:44
• Updated: 04/Jun/22 22:30

Description

When running a WriteToBigQuery beam step, a 503 error code is returned from `https://www.googleapis.com/resumable/upload/storage/v1/b/<our_tmp_dataflow_location>`. This is causing duplicated data as the BQ load job is still successfully submitted but the workitem returns "Finished processing workitem with errors". This causes dataflow to resubmit an identical job and thus insert duplicate data into our BigQuery tables.

Problem you have encountered:

1.) WriteToBigQuery step starts and triggers a BQ load job.

"Triggering job beam_bq_job_LOAD_AUTOMATIC_JOB_NAME_LOAD_NAME_STEP_650_f2f7eb5ec442aa057357302eb9cb0263_9704d08e74d74e2b9cc743ef8a40c524"

2.) An error occurs in the step, but apparently after the load job was already submitted.

"Error in _start_upload while inserting file gs://<censored_bucket_location>.avro: Traceback (most recent call last):
 File "/usr/local/lib/python3.7/site-packages/apache_beam/io/gcp/gcsio.py", line 644, in _start_upload
 self._client.objects.Insert(self._insert_request, upload=self._upload)
 File "/usr/local/lib/python3.7/site-packages/apache_beam/io/gcp/internal/clients/storage/storage_v1_client.py", line 1156, in Insert
 upload=upload, upload_config=upload_config)
 File "/usr/local/lib/python3.7/site-packages/apitools/base/py/base_api.py", line 731, in _RunMethod
 return self.ProcessHttpResponse(method_config, http_response, request)
 File "/usr/local/lib/python3.7/site-packages/apitools/base/py/base_api.py", line 737, in ProcessHttpResponse
 self.__ProcessHttpResponse(method_config, http_response, request))
 File "/usr/local/lib/python3.7/site-packages/apitools/base/py/base_api.py", line 604, in __ProcessHttpResponse
 http_response, method_config=method_config, request=request)
apitools.base.py.exceptions.HttpError: HttpError accessing <https://www.googleapis.com/resumable/upload/storage/v1/b/bqflow_dataflow_tmp/o?alt=json&name=tmp%2F<censored_bucket_location>.avro&uploadType=resumable&upload_id=ADPycdtKO3HR5PjM_lE6lBin-QqIRuTBeiaCe3dPx9gUKAIPI5fzpfuTs4J5XEF9XiayNvMrhGsGe0XP1CJv90xsuBUrZy6mpw>: response: <\{'content-type': 'text/plain; charset=utf-8', 'x-guploader-uploadid': 'ADPycdtKO3HR5PjM_lE6lBin-QqIRuTBeiaCe3dPx9gUKAIPI5fzpfuTs4J5XEF9XiayNvMrhGsGe0XP1CJv90xsuBUrZy6mpw', 'content-length': '0', 'date': 'Tue, 05 Oct 2021 18:01:51 GMT', 'server': 'UploadServer', 'status': '503'}>, content <>
"

3.) workitem finishes with errors

 Finished processing workitem X with errors. Reporting status to Dataflow service.

4.) Beam re-runs the workitem which spawns another identical BQ load job.

Triggering job beam_bq_job_LOAD_AUTOMATIC_JOB_NAME_LOAD_NAME_STEP_650_f2f7eb5ec442aa057357302eb9cb0263_1247e55bd00041d8b8bd4de491cd7063

This causes a single WriteToBigQuery beam step to spawn two identical BQ load jobs. This creates duplicated data in our tables.

What you expected to happen:

I would expect the HTTP call to be retried before returning an error. Otherwise, if this did fail, I would expect the same BQ load job to not be successfully submitted twice without cancellation of the first job. A third option would be to implement something similar to "insert_retry_strategy" but for batch files that can allow us to not create another bq load job when a failure occurs.