Too many shards in GCS

[damccorm]

We have observed case where the data was spread very thinly over automatically computed number of shards.

This caused wait for the buffers to fill before sending the data over to gcs causing upload timeout as we did not upload any data for while waiting.

However, by setting an explicit number of shards (1000 in my case) solved this problem potentially because all the shards had enough data to fill the buffer write avoiding timeout.

 

We can improve the sharding logic so that we don't create too many shards.

Alternatively, we can improve connection handling so that the connection does not timeout.

Imported from Jira BEAM-9752. Original Jira may contain additional context.
Reported by: angoenka.