On Sat, Mar 30, 2013 at 3:10 PM, Roy Smith <r...@panix.com> wrote: > In article <mailman.3986.1364615879.2939.python-l...@python.org>, > Chris Angelico <ros...@gmail.com> wrote: > >> Side point: You mentioned SSDs. Are you aware of the fundamental risks >> associated with them? Only a handful of SSD models are actually >> trustworthy for databasing. > > We haven't decided if we're going that route yet, but if we do, we will > probably use do RAID SSD for added reliability. We also have all our > database servers in failover clusters, so we get added reliability that > way too.
But will you know if you have corruption? Normally, transactional integrity means: 1) If a transaction, from begin to commit, is not completely applied, then it is completely not-applied; and 2) If a transaction that was affected by this one has been applied, then so has this one. SSDs that lie about fsync (and some hard disks lie too, as do some operating systems and some file system drivers, but - under Linux at least - it's possible to guarantee the OS and FS parts) can violate both halves. Anything might have been written, anything might have been missed. I did some tests with PostgreSQL on an SSD, and the results were seriously scary. > But, we have some runway left with more conventional technologies, so we > don't need to decide for a while. Ultimately, however, as reliability > goes up and cost comes down, it's hard to imagine the SSD isn't going to > be a big part of our lives at some point. Yes. I hope that by then, the manufacturers will realize that TPS isn't the only thing that matters. I'm sure SSDs will mature to the point where we can trust all brands equally (or at least, most brands - maybe it'll be like "server SSDs" and "desktop SSDs"?), but until then, there aren't many options. ChrisA -- http://mail.python.org/mailman/listinfo/python-list
