[Bug]: Dataloss when exceptions received from BQ while using Storage API Write in batch mode

[johnjcasey]

What happened?

This affects versions 2.35 - 2.47, though it is most severe in 2.44-2.47 due to #26520 causing guaranteed exceptions.

When we get an exception from BQ when attempting to append to a stream, we will retry the current append on a new stream. However, in doing so, we abandon the existing stream. This results in messages on that stream that have not yet been committed not making it to BQ, resulting in data consistency issues.

public static void main(String[] args) {
 BigQueryOptions options = PipelineOptionsFactory.fromArgs(args).create().as(BigQueryOptions.class);
 options.setStorageApiAppendThresholdRecordCount(5);
 Pipeline p= Pipeline.create(options);

 p.apply("ReadLines", TextIO.read().from("gs://apache-beam-samples/shakespeare/kinglear.txt"))
     .apply("Save Events To BigQuery", BigQueryIO.<String>write()
         .to("google.com:clouddfe:reprodataset.reprotable")
         .withFormatFunction(s -> new TableRow().set("words", s))
         .withMethod(Write.Method.STORAGE_WRITE_API)
         .withCreateDisposition(CreateDisposition.CREATE_NEVER)
         .withWriteDisposition(BigQueryIO.Write.WriteDisposition.WRITE_APPEND));

 p.run();
}

repros the issue on versions 2.44-2.47

Issue Priority

Priority: 1 (data loss / total loss of function)

Issue Components

• Component: Python SDK
• Component: Java SDK
• Component: Go SDK
• Component: Typescript SDK
• Component: IO connector
• Component: Beam examples
• Component: Beam playground
• Component: Beam katas
• Component: Website
• Component: Spark Runner
• Component: Flink Runner
• Component: Samza Runner
• Component: Twister2 Runner
• Component: Hazelcast Jet Runner
• Component: Google Cloud Dataflow Runner

[johnjcasey]

Fixed by #26503

[rszper]

As a workaround until the fix is available, use the File Loads method instead of the Storage Write API.

Issue symptom that customers might see:

• Data written to BigQuery is incomplete even though Dataflow shows the correct output write number.

[Dhaamodharan]

Can someone help me how to fix/update the SDK version? Currently i am using SDK 2.39.0, I dont find any solution to update the SDK on runtime, or Is there any other way to do without restarting the job?