AWS NVMe Storage - hosting a database

Question

Now that AWS are offering NVMe through the i3 range of servers, is there a best practice for hosting a database on the instance storage of one of these?

My understanding is that if the instance is stopped, the storage may be completely wiped. This doesn't appear to be the case if the server reboots, intentionally or unintentionally, but you are still one button press away from wiping important data so this is quite scary.

My understanding of the underlying infrastructure is that this is because the NVMe storage is directly attached to the physical host, and therefore if Amazon decide to move your VM to another host you would lose your data. Also it would be bad to store mission-critical data on a single hardware device AWS aside.

But given the performance benefits of NVMe over EBS (SAN?) storage, what would a recommended setup be? VM Replicas, transaction log backups to permanent storage, etc.

Answer 1

It is possible to turn the NVMe SSDs on i3 instances into persistent highly available storage.

Options:

1) Mirroring between NVMe SSDs on 2 or 3 instances

2) Mirroring between NVMe SSDs and EBS (EBS can be on a different instance) with reads primarily from NVMe SSDs.

While write performance will still be limited by network or EBS, you do get full read performance of NVMe. In most cases read bandwidth is what large databases really need for running heavy queries.

However, there are still questions about failing over the database between the instances and restoring redundancy after instance stop/start of failure.

Check this whitepaper and page 9 specifically for details about how it is done for Oracle database clusters: https://www.flashgrid.io/wp-content/sideuploads/resources/FlashGrid_OracleRAC_on_AWS.pdf

The paper is focused on Oracle RAC databases, but same solution works for single-instance Oracle and also for any other Linux-based database. Although, you would still need Oracle Clusterware (free).