Configure modified retry for uploading partition yamls to GCS

[karcot1]

When uploading a large number of yamls to GCS (10,000) we are encountering a 503 error. This indicates the connection to GCS was interrupted, and is commonly encountered when uploading a large number of files via python.

The exact error message: 'Request failed with status code', 503, 'Expected one of', <HTTPStatus.OK: 200>

The recommendation with 5xx errors is to retry, and we have some sample code as a reference: https://github.com/googleapis/python-storage/blob/main/samples/snippets/storage_configure_retries.py#L55

The request is to please modify _write_gcs_file in gcs_helper.py to use retries.

Suggestion (based on above sample code):

from google.cloud import storage
from google.cloud.storage.retry import DEFAULT_RETRY

def _write_gcs_file(file_path: str, data: str):
    gcs_bucket = get_gcs_bucket(file_path)
    blob = gcs_bucket.blob(_get_gcs_file_path(file_path))

    modified_retry = DEFAULT_RETRY.with_deadline(500.0)
    modified_retry = modified_retry.with_delay(initial=1.5, multiplier=1.2, maximum=45.0)

    blob.upload_from_string(data, **retry=modified_retry**)

Ideally, the values (deadline, initial, multiplier, maximum) can be parameterized, so that end users can modify the values to get optimal performance.

[sundar-mudupalli-work]

Hi,

Review of the customer usage showed that the customer had upto 8 separate generate-partitions executions running in parallel. Given how long it takes to write to GCS (about 4 hrs for 10,000 objects). Each DVT execution is single threaded, so unclear how it raised 503 errors. DVT creates a client for each file as outlined below - which could exacerbate the situation.

When writing a large number of GCS files for partitions, it can take about 1.25 seconds to write each file to GCS. This is awfully slow. The problem is in get_gcs_bucket function - which creates a http client and a connection to the bucket every time. So if you have 10,000 files, this can take very long. However if the return value from get_gcs_bucket is cached, then writes to GCS for small files improve dramatically (> 20x).

It may make sense to restructure the gcs_helper package as a class, perhaps rename it to filsys_helper - which allows DVT to interact with Cloud Storage or the file system as specified by the user. If the customer specifies a GCS destination (or source), we can save the storage_client.get_bucket('bucket_name') result in the class so that it can be reused. This will significantly improve performance.

Sundar Mudupalli