Why does my Apache Beam Dataflow pipeline not write to BigQuery?

I'm working on an Apache Beam pipeline that processes data and writes it to BigQuery. The pipeline works perfectly when using the DirectRunner, but when I switch to the DataflowRunner, it completes without errors or warnings but doesn't insert any rows into BigQuery. Additionally, I see large leftover files in the temporary directory of my Cloud Storage bucket (gs://my-bucket/temp/bq_load/...), and no data appears in the target table.

Here’s the pipeline structure:

worker_options.sdk_container_image = '...'

with beam.Pipeline(options=pipeline_options) as p:
    processed_data = (
        p
        | "ReadFiles" >> beam.Create(FILE_LIST)
        | "ProcessFiles" >> beam.ParDo(ProcessAvroFileDoFn())
        | "WriteToBigQuery" >> beam.io.WriteToBigQuery(
            table=f"{PROJECT_ID}:{DATASET_ID}.{TABLE_ID}",
            schema=BQ_SCHEMA,
            write_disposition=beam.io.BigQueryDisposition.WRITE_APPEND,
            create_disposition=beam.io.BigQueryDisposition.CREATE_IF_NEEDED
        )
    )

Key Observations:

• The pipeline succeeds with the DirectRunner, writing data to BigQuery without any issues.
• With the DataflowRunner, the pipeline completes without errors or warnings, but: No rows are written to BigQuery & large temporary files remain in the bucket (e.g., bq_load/...).
• The data being processed is valid NDJSON.
• The BigQuery schema matches the data structure.

What I’ve Tried:

• Inspecting the leftover temp files, I downloaded the temp file and verified that it contains valid NDJSON rows. Manually uploading this file to BigQuery using the bq load command works fine.

• Testing with other datasets: I tried many different inputs, but the issue persists.

• Checking Dataflow logs: I looked at the logs in the Dataflow Monitoring Console but found no errors or warnings.

• Different service account: a service account with insufficient dataflow permissions throws an error. It therefore seems unlikely that the issue is in permissions for the workers.

I saw one other thread about this (Can't make apache beam write outputs to bigquery when using DataflowRunner) but nothing got resolved there.

The problem turned out to be that I defined the google Storage Client object globally in my script.

When I refactored the code to be modular, and put the storage client & bucket initialization in the setup function of my DoFn, it started working.

I don't know why it failed silently like this and left no trace, this was such a pain to debug.