On Thu, 23 Jun 2011 09:28:49 PDT ron minnich <[email protected]> wrote: > > The main point I took from the talk they gave was that failure was > most strongly related to the number of writes in FLASH. If your > striping strategy is to duplicate writes to each drive, you faced the > happy prospect of doing a write and having both drives fail at the > same time. Hard drives have a different way of failing. We've seen > weirdness like this here, with drives in a bunch of nodes that all > seem to fail simultaneously, well within rated lifetime. Not cheap > drives either. Of course that was a little while ago and things seem > to have gotten better, but it's worth a warning.
All they are saying is to age SSDs at different rate to avoid correlated failures. Disk drives have a similar problem in that disks from the same batch seem to die at a similar age. One issue is that N years later it is not cost effective to get a replacement disk of the same size. Now I think this (dying at the same age) is actually a good thing! The key is to not wait to replace until they die; just replace them all when you decide to replace *any*! zfs helps since it will automatically grow the space (So for instance, on my home system originally I used a mirror of 2 250GB used IDE disks and another mirror of 2 300GB sata disks, striped together. I first replaced both IDE disks with bigger ATA disks. Later I replaced the 300GB sata disks with 1TB disks and now I have a lot more space to play with). @work I used ZFS raidz2 on 2TBx6 drives and a 2x80GB SSD mirror for root + the write intent log (this is a server for backing up N machines, so write performance is more critical). Due to a mixup we are using MLC SSDs instead of SLC SSDs (to be replaced at some point). Not ideal but works well enough.
