Comment by nabla9 | Hacker News Reader

Comment by nabla9 | Hacker News Reader

Parent Full thread

First:

> zfs is block based, not extent based, whereas all other modern filesystems have been extent based for years: the reason they did this is that extents plus snapshots are really hard.

then:

>Snapshots: In progress. bcachefs's snapshot implementation is going to be significantly more capable than any competing implementations. When it's done, ...

Multiple ZFS challengers can get easily to the point where they have 80-90% of the cool features that ZFS has. Then everything slows down to crawl because getting the same performance and reliability + features is super hard. When you leave the essential but hard features 'for later' they are harder to add later without big rewrite (or losing performance).

Ticking all the boxes with good performance becomes exponentially harder as number of features grow. There comes time to order things: sacrifice performance in this to attain this.

bcachefs is specifically trying to avoid the btrfs garbage fire by building slowly on top of well-tested bcache. I am confident it won't end up like btrfs where RAID5/6 is still horribly broken yet people claim the filesystem is totally stable and fine for production use.

You can use btrfs on RAID5/6 without problems if you don't use the RAID features that btrfs provides.

the trouble with block based raid5/6 is the write hole - when you're partially updating a stripe, you can't avoid having the p/q blocks be momentarily inconsistent because there's no way of doing atomic writes to multiple drives simultaneously (short of full data journalling). That means data unrelated to a given write has inconsistent p/q while that write is in flight.

the appeal of doing raid5/6 within a COW filesystem is that it's potentially an opportunity to get rid of the write hole. Of course, the devil is in the details.

Kent's a really capable and experienced guy, he has a design in mind and I have every confidence he'll get snapshots working and fast.

But even if that's not true, bcachefs still has an excellent feature set already with great performance, especially in long-tail latency.

It's not about being excellent or not, or if he can implement something that technically works.

He or any other excellent guy don't have an idea how the implementation he designs works in multiple workloads and different machines and disk drives. There will be time period of several years of adjusting and tinkering and compromising.

Reply on news.ycombinator.com