How AWS S3 serves 1 petabyte per second on top of slow HDDs

I realize you're making a general point about space/IO ratios and the below is orthogonal, no contradiction.

It's actually a lot less user-facing per disk IO capacity that you will be able to "sell" in a large distributed storage system. There's constant maintenance churn to keep data available: - local hardware failure - planned larger scale maintenance - transient, unplanned larger scale failures (etc)

In general, you can fall back to using reconstruction from the erasure codes for serving during degradation. But that's a) enormously expensive in IO and CPU and b) you carry higher availability and/or durability risk because you lost redundancy.

Additionally, it may make sense to rebalance where data lives for optimal read throughput (and other performance reasons).

So in practice, there's constant rebalancing going on in a sophisticated distributed storage system that takes a good chunk of your HDD IOPS.

This + garbage collection also makes tape really unattractive for all but very static archives.

Ceph's design is to avoid a single bottleneck or single point of failure, with many nodes that can all ingest data in parallel (high bandwidth across the whole cluster) and be redundant/fault tolerant in the face of disk/host/rack/power/room/site failures. In exchange it trades away some of: low latency, efficient disk space use, simple design, some kinds of flexibility. If you have a "small" use case then you will have a much easier life with a SAN or a bunch of disks in a Linux server with LVM, and probably get better performance.

How does it work with no single front end[1] and no centralised lookup table of data placement (because that could be a bottleneck)? All the storage nodes use the same deterministic algorithm for data placement known as CRUSH, guided by placement rules which the admin has written into the CRUSH map, things like:

- these storage servers are grouped together by some label (e.g. same rack, same power feed, same data centre, same site).

- I want N copies of data blocks, separated over different redundancy / failure boundaries like different racks or different sites.

There's a monitoring daemon which shares the CRUSH map out to each node. They get some data coming in over the network, work through the CRUSH algorithm, and then send the data internally to the target node. The algorithm is probabalistic and not perfectly balanced so some nodes end up with more data than others, and because there's no central data placement table this design is "as full as the fullest disk" - one full disk anywhere in the cluster will put the entire cluster into read-only mode until you fix it. Ceph doesn't easily run well with random cheap different size disks for that reason, the smallest disk or host will be a crunch point. It runs best with raw storage below 2/3rds full. It also doesn't have a single head which can have a fast RAM cache like a RAID controller can have.[2] Nothing about this is designed for the small business or home use case, it's all designed to spread out over a lot of nodes[3].

It’s got a design where the units of storage are OSDs (Object Storage Devices) which correspond roughly to disks/partitions/LVM volumes, each one has a daemon controlling it. Those are pulled together as RADOS (Reliable Autonomic Distributed Object Store) where Ceph internally keeps data, and on top of that the admin can layer user-visible storage such as the CephFS filesystem, Amazon S3 compatible object storage, or a layer that presents as a block device which can be formatted with XFS/etc.

It makes a distributed system that can ingest a lot of data in parallel streams using every node’s network bandwidth, but quite a lot of internal shuffling of data around between nodes and layers adding latency, and there are monitor daemons and management daemons overseeing the whole cluster to keep track of failed storage units and make the CRUSH map available to all nodes, and those ought to be duplicated and redundant as well. It's a bit of a "build it yourself storage cluster kit" which is pretty nicely designed and flexible but complex and layered and non-trivial.

There are some talks on YouTube by people who managed and upgraded it at CERN as targets of particle accelerators data which are quite interesting. I can only recommend searching for "Ceph at Cern" and there are many hours of talks, I can't remember which ones I've seen. Titles like: "Ceph at CERN: A Year in the Life of a Petabyte-Scale Block Storage Service", "Ceph and the CERN HPC Infrastructure", "Ceph Days NYC 2023: Ceph at CERN: A Ten-Year Retrospective", "Ceph Operations at CERN: Where Do We Go From Here?".

[1] If you are not writing your own software that speaks to Ceph's internal object storage APIs, then you are fronting its with something like a Linux machine running an XFS filesystem or the S3-compatible gateway, and that machine becomes your single point of failure and bottleneck. Then you front one Ceph cluster with many separate Linux machines as targets, and have your users point their software to different front ends, and in that case why use Ceph at all? You may as well have had many Linux machines with their own separate internal storage and rsync, and no Ceph. Or two SANs with data replication between them. Do you need (or want) what Ceph does, specifically?

[2] I have only worked on HDD based clusters, with some SSDs for storing metadata to speed up performance. These clusters were not well specced and the metadata overflowed onto the HDDs which didn't help anything.

[3] There are ways to adjust the balance of data on each node to work with different size or nearly full disks, but if you get to read-only mode you end up waiting for it to internally rebalance while everything is down. This isn't so different to other storage like SANs, it's just that if you are going for Ceph you probably have a big cluster with a lot of things using it so a lot of things offline. You still have to consider running multiple Ceph clusters to limit blast radius of failures, if you are thinking "I don't want to bother with multiple storage targets I want one Ceph" you still need to plan that maybe you don't just want one Ceph.

RAID doesn’t exactly make writes faster, it can actually be slower. Depends on if you are using RAID for mirroring or sharding. When you mirror, writes are slower since you have to write to all disks.

Yeah unfortunately mirrors is no go due to efficiency requirements, but luckily read performance is not that important if I manage to completely offload FS/S3 metadata and small files to flash storage (separate zpool for Garage metadata, separate special VDEV for metadata/small files).

I think I'm going to go with 8x RAIDz2 VDEVs 10x HDDs each, so that the 20 drives in the internal drive enclosure could be 2 separate VDEVs and not mix with the 60 in the external enclosure.

What you wrote only applies to rotational latency, not seek latency. The seek latency is the time it takes for the head to reach the target. Heads only rotate within the small range like [ 0 .. 25° ], they are designed for rapid movements in either direction.