> I have come across some ideas on the Internet
> where people hinted at running production PostgreSQL workloads
> on AWS ephemeral SSD storage.
I think people were more serious about that before AWS introduced
PIOPS. I wouldn't do this unless I had streaming replication to a
standby, plus regular backups (e.g. with WAL-E). Actually that's what
I use even with regular EBS volumes, and it's not hard to set up. The
standby gives you fast failover/recovery, and WAL-E gives you an extra
layer of insurance.
Another worry about ephemeral storage is that you can't add more, and
the amount you get depends on the instance type. Also it seems like
modern instances don't come with as much ephemeral storage as they
used to, although maybe that impression is mistaken.
I agree that PIOPS is still expensive though. Some people get around
that by running a RAID0 array of gp2 EBS volumes. The conversation at
https://news.ycombinator.com/item?id=13842044 has some details. I've
set it up for one client, and it wasn't too bad. It's been running
fine for 6 months or so.
Paul