I found this article very unsatisfying due to the lack of key details. This reads like a blog post tailored for an audience of folks with pointy hair.

As far as I can tell, they went from Hadoop mappers pulling logical backups from the DBs using python and mysqldump to ... Hadoop mappers pulling logical backups from S3, which were pushed there by scripts running on the DBs, probably still using mysqldump. Although I have no idea, since there are no details. Are these backups physical? Logical? How is the 12 hour big table problem solved by this approach? Why was there a limitation on the number of mappers usable by the old approach?

And what of the old system's DB failover problem? Nothing that can't be solved with a script that is failover aware! Nice. No reason to have dedicated ingestion slaves that aren't, you know, master candidates, when you have scripts that can restart themselves.