Database integration sample: How to handle transient database errors

Creating this issue to track figuring out how to handle resuming training when a transient error (for example any of the errors described at https://docs.microsoft.com/azure/sql-database/sql-database-connectivity-issues) occurs. The simple strategy would be to throw away all the progress made, but for a very large data set, this may mean the operation never finishes. If we can come up with a reasonably general pattern to resume from the last correctly read row (probably something that reasons about primary keys), that would be great.

cc @CESARDELATORRE