[ElDji]

I successfully migrated from MinIO to Ceph, which I highly recommend. Along the way, I tested SeaweedFS, which looked promising. However, I ran into a strange bug, and after diagnosing it with the help of Claude, I realized the codebase was vibe-coded and riddled with a staggering number of structural errors. In my opinion, SeaweedFS should absolutely not be used for anything beyond testing — otherwise you're almost certain to lose data.

[MikeKusold]

Seaweed has been around for a long time. I think you just discovered what legacy codebases look like.

[g-mork]

Laughed reading this. We pretend Claude can't code because we don't like to acknowledge what code always turns out looking like, which is exactly what it's trained on

[chrislusf]

(I work on SeaweedFS.)

Haha, you used Claude to find the Clause code.

I used Claude to generate a lot of admin UI pages, saved a lot of time. The core storage engine part I dare not using AI, same as you.

[vimda]

Ceph is the OG. Every now and then different attempts to replace it pop up, work well for some use cases, and then realise how hard the actual problem they are trying to solve is. Ceph always wins in the end.

[Giefo6ah]

Ceph solves the distributed consistent block storage problem very well. But I hardly ever need that problem solved, it's way more often that I need a distributed highly available blob storage, and Ceph makes the wrong tradeoffs for this task.

[q3k]

Ceph is fundamentally an object storage system. RBD (block devices) are built on top of that layer.

[yencabulator]

Yeah. 100% this. Remember, Ceph's storage nodes are called OSD, as in Object Storage Daemon.

The biggest reasons to not use Ceph are:

- You plan on using <=30 disks on <=3 computers. Some of the petabyte scale stuff is just in the way at that scale. Even more so if you're looking at a single computer.

- You don't have at least fractional ops staff that will look at a dashboard regularly. In this world, you're better off using a cloud service.

- You're 100% satisfied with the S3 interface and will never want anything else than its write-once objects. Ceph's writable object support won't gain you anything in that world, and genuinely makes the distributed systems problems involved much harder. Ceph was architected as a distributed filesystem and excels as a networked block store for virtual machines. Ceph can do S3-compatibility, but another implementation can cut off a big chunk of the functionality and provide just S3-compat, and simpler can be better.

I doubt many software projects mentioned in this conversation have gone through the extensive stress testing with glitching nodes while trying to maintain performance that Ceph has; simple systems can be too simple. Ceph is quite well battle-hardened by now. Ceph's dedicated QA hardware pool is likely bigger than many competing projects have tried as a cluster size!

Disclaimer: ex-Ceph-developer.