Avi Kivity

Tablets can't migrate away from a node that is down. Replace should be equivalent to adding a new node and rebuilding tablets on the missing node, so indeed it should...

Looks good except that I don't like the new coordinator column.

> > That would make the get_replicas() API inconvenient. Instead of returning a > > vector it would return vector, and > > everyone using it would have to handle...

> this change is marked as a draft. as it would fail the tests before we are able to test using newer linux kernels. We can't depend on the kernel...

Didn't we leak these members? Will layered file implementations break? I see it in scylladb-enterprise mirrored_file_impl.

We might want to just expose IORING_OP_URING_CMD and let the user be responsible for everything else. This enables everything that IORING_OP_URING_CMD can offer (not just NVMe ZNS), and when Linux...

> yeah, that'd be much more flexible. but i think we could probably also offer some tested and ready-to-use popular commands built on top of this API. they can be...

Sorry about the missed review. Please ping earlier, or it will drag forever. I think it's too intrusive (the io_result changes). This is a bypass, so it should be implemented...

abort_when_xx sounds like a function, not a class. I don't know how to feel about this. It looks like a maze of small features, for something that should be easy...

> > abort_when_xx sounds like a function, not a class. > > I don't know how to feel about this. It looks like a maze of small features, for something...