Building blobd: single-machine object store with sub-ms reads and 15 GB/s upload (blog.wilsonl.in)

Similar systems include Facebook's Haystack and its open source equivalent, SeaweedFS.

Interesting project but lack of S3 protocol compatibility and fact it seems to YOLO your data means it's not acceptable for many.

That’s a lot of work creating a whole system that stores data on a raw block device. It would be nice to see this compared to… a filesystem. XFS, ZFS and btrfs are pretty popular.

> Despite serving from same-region datacenters 2 ms from the user, S3 would take 30-200 ms to respond to each request.

200ms seems fairly reasonable to me once we factor in all of the other aspects of S3. A lot of machines would have to die at Amazon for your data to become at risk.

> Direct I/O means no more fsync: no more complexity via background flushes and optimal scheduling of syncs. There's no kernel overhead from copying and coalescing. It essentially provides the performance, control, and simplicity of issuing raw 1:1 I/O requests.

Not true, you still need fsync in direct I/O to ensure durability in power loss situations. Some drives have write caches that means acknowledged writes live in non-volatile memory. So maybe the perf is wildly better because you’re sacrificing durability?

  /// Even when using direct I/O, `fsync` is still necessary, as it ensures the device itself has flushed any internal caches.
  async fn sync(&self) {
    let (fut, fut_ctl) = SignalFuture::new();
    self.sender.send(Request::Sync { res: fut_ctl }).unwrap();
    fut.await
  }

S3's whole selling point is 11 9s of durability across the whole region which is probably why it's slow to begin with.

When you have a service and really care about shoving of S3 latencies in the millisecond range, then you propably have enough users that all the tiny images are cached @ edge anyways.