Scott Peterson

Some of our Intel Ceph team (@GuifengTang, @qiaoweix, @optimistyzy, @majianpeng, @CongMinYin) is starting to look into the details of building this. This issue seems like a good place to hold...

The general consensus seems to be that the discovery service could/should be written in Python, because that might make it simpler to implement and maintain. After some investigation it seems...

> I think one additional requirement that I don't see explicitly mentioned is the ability to have one discovery service span multiple Ceph clusters, e.g. think about migrating a volume...

> SPDK has what we need here, but not in the form we need it in. Actually SPDK might already export a lot of what we need (NVMe/TCP and some...

> When the config in OMAP changes, the discovery service should update its contents to match, right? When it does that, hosts will be updated via the standard NVMe-oF discovery...

> A gateway may be down. Will it be ok if the discovery service advertises resources that are down? It's not ideal for a DS to advertise ports that remain...

> > SPDK has what we need here, but not in the form we need it in. > > Actually SPDK might already export a lot of what we need...

> > When the config in OMAP changes, the discovery service should update its contents to match, right? When it does that, hosts will be updated via the standard NVMe-oF...

Yes, Even if we don't make it an official feature, a Prometheus exporter is a valuable tool for benchmarking and debugging. I've done this before on small scales with a...

The metrics provided by this are at least part of what's needed to meet the requirement in #116. #116 probably requires some metrics from the discovery services as well, to...