Global performance tuning
I see that there are several international datacenters available for deployment. If I have a single non-replicated, vertically scaling Postgres database machine. Does it make any sense to deploy front-end machines to the datacenters around the globe? Or will I get better user performance deploying FEs to only the same DC as the database? Each user request typically makes at least one, but maybe more round trips to the DB machine.
If it does make sense to deploy FEs globally, how would the client target the closest one?
If you are up for it you can also consider some kind of database replication. Some applications can use master-slave replication where you always write to a single master somewhere in the world but reads go from a local read slave. This is only a benefit if most page loads do no writes at all.
Finally, there are possibilities to set up master-master (also called multi-master) replication over WAN. This usually slows down the speed of commits since they need to lock while waiting for acknowledgements from other nodes. I'm not very familiar with the Postgre ecosystem, but I found these two projects that might be interesting if you want to go down this road:
If you are deploying a global (around the Earth sort of thing) system, then you definitely need to setup front-end facing servers around the globe in different data centres. But you will have to be very smart with data replication, caching and performance tuning, not to mention timing and conflict resolution.
I kept a large (30+) cloud of servers on the linode Frankfurt data centre and the whole data centre went down last week for half a day, so all my high availability systems became irrelevant. Oh well…
It's pretty concerning that an entire datacenter could go down, that would essentially require multi DC replication for any datastore. Of course, I'm just starting out a product, so dome downtimes will be understandable.
I've discussed with close friends, and we decided that, if the data is just in one location, there's really no point to distributing front-ends. Any global distribution would pretty much require data replication to provide any benefit.
I also host things with other providers, the Canadian data centre died once (for nearly a week) and the US data centre has died twice. Amazon has had similar downtime as well in recent history.
So I think linode is within the average death count, in the industry