Anyone know how they handle failover? I'd imagine they won't be physically copying 500GB databases between machines. But then the size limitation indicates they're constrained to what a single machine can hold.
We offer two different ways of working with Cloud SQL: synchronous and asynchronous writes. With the synchronous writes, you'll get an OK back when you update the database, which lets you know that the write completed successfully and is replicated. If something bad happens, you'll get an exception back and can keep the data in memory until the database is live and accepting connections again.
A faster method is asynchronous, which performs the writes every second or so, so you're not waiting for replication to complete. If something dies during that second, you could lose data during that period and not know it.
Does this help?
Disclaimer: I'm a developer on Cloud SQL.
Latency is the same as if you were running MySQL on any other platform AFAIK. If the database is cold, yes, the DB is spun down. An incoming connection will cause the DB to spin up, and in most languages the common MySQL connector will block until it's ready, so it doesn't require special coding around. Databases almost always come up in a matter of seconds.