Writing to BigQuery from Dataflow - JSON files are

问题:

One of our Dataflow jobs writes its output to BigQuery. My understanding of how this is implemented under-the-hood, is that Dataflow actually writes the results (sharded) in JSON format to GCS, and then kicks off a BigQuery load job to import that data.

However, we've noticed that some JSON files are not deleted after the job regardless of whether it succeeds or fails. There is no warning or suggestion in the error message that the files will not be deleted. When we noticed this, we had a look at our bucket and it had hundreds of large JSON files from failed jobs (mostly during development).

I would have thought that Dataflow should handle any cleanup, even if the job fails, and when it succeeds those files should definitely be deleted Leaving these files around after the job has finished incurs significant storage costs!

Is this a bug?

Example job id of a job that "succeeded" but left hundreds of large files in GCS: 2015-05-27_18_21_21-8377993823053896089

回答1:

Because this is still happening we decided that we'll clean up ourselves after the pipeline has finished executing. We run the following command to delete everything that is not a JAR or ZIP:

gsutil ls -p <project_id> gs://<bucket> | grep -v '[zip|jar]$' | xargs -n 1 gsutil -m rm -r

回答2:

Another possible cause of left over files is cancelled jobs. Currently dataflow does not delete files from cancelled jobs. In other cases files should be cleaned up.

Also the error listed on the first post "Unable to delete temporary files" is the result of a logging issue on our side, and should be resolved within a week or two. Until then, feel free to ignore these errors as they do not indicate left over files.

回答3:

This was a bug where the Dataflow service would sometimes fail to delete the temporary JSON files after a BigQuery import job completes. We have fixed the issue internally and rolled out a release with the fix.