paulgorman.org

< ^ txt

Tue Oct 6 09:18:05 EDT 2015 Went to bed around 11:30, slept well, and woke up around 7:30. Cloudy today with a high of sixty-six. I'm more or less decided now that I'll get a new car this weekend. Leaning toward leasing, but we'll see what they have used. Goals: Work: - Formulate an informed opinion about the necessity of ecc for zfs (I guess zfs stuff is now semi-work-related) Conclusions: - First, if your processor and motherboard supports ECC RAM, just buy ECC RAM! - If your data is business-critical --- if the box can't be down while you restore from backup --- buy ECC RAM! - For less sensitive stuff, these are the considerations: - The more writes you're doing, the more vulnerable you are if you don't have ECC RAM. - The magnitude of the potential problem depends in part on your storage configuration. - If you're building a household fileserver, ECC RAM is nice to have but not critical. (You should still have backups!) - Of course, having backups is only helpful if know you have bad data, and your backups go back far enough to still have good data. - Running memtest86 to detect a bad DIMM is a good idea, although it's not a significant mitigation. Reasoning: - Data might be corrupted in RAM (cosmic ray bit flip, bad memory module, etc.). - ECC can prevent or detect many of these errors. - ZFS checksums data written to disk. - Without ECC, you might write bad data to disk believing it's good, or good data might get a get flagged with a checksum mismatch. - ZFS isn't significantly more susceptible to this problem than any other filesystem, but... - Certain disaster scenarios posit a bad RAM checksum-fail/zfs-redundancy-fix-fail cascade scenario that could lead to the death of a vdev. - Because ZFS stripes data across every vdev in a pool, the loss of a single vdev kills the whole pool. - I don't find this disaster scenario totally convincing, but... - ZFS doesn't include any data recovery tools; its design assumes ECC. - Furthermore, because of the way zfs holds writes in memory as transaction groups, unwritten data has more opportunity to get bit flipped that with some other file systems. - (This is the only clear and convincing argument I've heard about why ZFS might be somewhat more vulnerable than other file systems.) - Ask Jason to study vlan tutorial Done. - Investigate mailing list issue No. Forty minute walk at lunch. Perhaps a quarter of the trees are turning, a few orange and red. Home: - Read more of Gilgamesh Done. - Order new shirts Done. Watched a couple of episodes of The Newroom. A little stagey, but good. Larry Wall released Perl 6.0 today. Neat. Need to check it out.

< ^ txt