Hacker News Reader

Comment by feld | Hacker News Reader

when the RAID is handled by a different layer you lose some very important integrity features

IIRC, doesn't BTRFS also allow you to do cool things like change RAID levels dynamically? (E.g. You can be running a 2-disk RAID-1 array, pop in another disk and tell BTRFS to make it a RAID-5 array instead, then a year later pop in 2 more disks and switch to RAID6, all with no downtime.) I imagine that wouldn't be possible if you were doing RAID at a different layer.

Linux's built-in software RAID (implemented at the block level) has supported online RAID level changes for many years. Check out the "grow mode" section of the mdadm(8) man page sometime; it goes into great detail about which operations are supported.

Yeah, shout out to mdadm and the md driver. I've been using it for years and years and it's rock solid. Being able to grow arrays (online) and convert plain drives into mirror sets is great. I feel that md and lvm are underappreciated by a large number of Linux users…

That's why all big data in big corporations uses exotic-open-source-filesystem-based RAID. That's where the action is when it comes to integrity.

None of that virtual block device or driver-level software junk, let alone hardware RAID controller solutions.

Well, there are other reasons; you want to write code that operates on the data, and neither the code nor the data fits on a single machine - you have to target an abstraction which spans machines. Block storage is too low-level an abstraction.

That isn't to say that using high performance block storage isn't still a win even when the redundancy is multiplied at a higher level. The higher level redundancy is also about colocating more data with the code - i.e. it's not just redundant for integrity, but to increase the probability it's close to the code.

Block storage can be network-abstracted.

Even virtual memory for that matter. Now ancient concept:

https://en.wikipedia.org/wiki/Distributed_shared_memory

Of course. Most production monoliths are deployed on networked block storage - aka SAN - and NUMA is already structurally distributed memory, even on a single box. But it's not the right paradigm to scale well, no more than chatty RPC that pretends the network doesn't exist is the right way to design a distributed system.

I think the high profile ones (Google, Amazon etc) use relatively dumb OS drivers and do the fancy distributed FS abstraction stuff in userspace. Certainly stuff like Ceph and Gluster don't have very good reputations and are mostly sold to relatively clueless "enterprise" customers.

RAID 5/6 in btrfs doesn't have data integrity features so you're back to square 1 on this regard.

Reply on news.ycombinator.com