DynamoDB table purge 400M records from a 1Billion item table & batchwrite throttling

With parallel scan, each thread scans a continuous segment in the table. If the items to be deleted are evenly distributed across the table space, you would encounter table level throttling instead of partition level throttling. The fact that you encountered partition level throttling without exceeding the provisioned WCU indicated that the items to be deleted might reside in a few partitions only. The result is, almost all items in a particular BatchWriteItem API call belong to the same partition, and you might be deleting from only a few partitions at a time.

One way to improve the performance would be perform a shuffling before doing the deletes. That is, the parallel scan threads pushes records to be deleted into a list. After the parallel scan finishes, perform a random ordering on items in the list. After that, the delete threads retrieve items from the list for deletion. With this approach, you increase the possibility that items in a BatchWriteItem API call are distributed in multiple partitions, taking advantage of the write capacity in multiple partitions.