Comment by simonjgreen | Hacker News Reader

Comment by simonjgreen | Hacker News Reader

Parent Full thread

Slight tangent from the point, however: friends don't let friends use RAID5. Putting this out as a PSA reminder as it's not often discussed any more given the prevalance of cloud storage.

RAID5 gives you really the worst of all worlds. You get poor performance (n/4) due to the number of operations required and you get the fragility of a RAID storm that comes from a cascade RAID failure when a drive dies. A "stripe of mirrors" (RAID10) gives you the safest and the highest performance array (at scales where these decisions are being made which most people would see anyway). The common argument against is the cost (usableCapacity = totalCapacity/2) but my counter argument is that RAID5 is a timebomb with a secret countdown. RAID5 is great if you don't need the data on the array.

Source: >20 years running complex large arrays

Or “software” raid 10 with ZFS mirror dev’s all slapped into one pool?

Precisely that

What's your opinion on RAID6? Reasonable middle ground between cost and redundancy or added complexity with the same performance problems as RAID5?

With Ceph, and I imagine other similar solutions, you can skip RAID altogether and manage redundancy at higher level over plain disks. RAID makes sense if you specifically want to have conventional filesystem on top of it, but for object storage its not necessary.

You're applying the same principles, just at the object level rather than the filesystem level. You still have to choose a redundancy pattern.

But I do agree with you, it does make the problem significantly easier/different.

ZFS and Ceph are marvelous creations.

RAID6 is passable for a low performance small scale bulk storage, but the performance hit is worse than RAID5 and the rebuild times are even longer.

RAID10 for 22TB disks with only two drives at "1" level? I would expect no less than 3 drives at such capacity. Otherwise there is a huge chance to lose the second drive while 1. changing the disc (=time, could be even days) and 2. replication.

What are your thoughts on implementing RAID 0 over four RAID 1 arrays, each consisting of three disks, resulting in a total of 12 drives?

Agreed, 22TB is a _long_ rebuild time. I'd either go 3 drives in the mirror in that scenario, or go for smaller spindles, depending on requirements. It's not always about the biggest spindles possible.

Reply on news.ycombinator.com