I think separating the database id, be it a shard id or the entire db, apart from the execution node/context where that database lives, so that the databases themselves can be mobile (or even duplicated) across multiple execution nodes makes perfect architectural sense to me. Keeping a _peers database which lists which databases are at which nodes makes sense me.
It seems like each "database" being its own thing separate and apart from the node it executes on is a cleaner model all around. Great idea! Mike On Apr 29, 2016 7:55 PM, "Adam Kocoloski" <kocol...@apache.org> wrote: > Hi all, > > I’ve doing a bit of poking around the container orchestration space lately > and looking at how we might best deploy a CouchDB 2.0 cluster in a > container environment. In general I’ve been pretty impressed with the > design point of the Kubernetes project, and I wanted to see how hard it > would be to put together a proof of concept. > > As a preamble, I needed to put together a container image for 2.0 that > just runs a single Erlang VM instead of the container-local “dev cluster”. > You can find that work here: > > https://github.com/klaemo/docker-couchdb/pull/52 < > https://github.com/klaemo/docker-couchdb/pull/52> > > So far, so good - now for Kubernetes itself. My goal was to figure out how > to deploy a collection of “Pods” that could discover one another and > self-assemble into a cluster. Kubernetes differs from the traditional > Docker network model in that every Pod gets an IP address that is routable > from all other Pods in the cluster. As a result there’s no need for some of > the port gymnastics that one might encounter with other Docker environments > - each CouchDB pod can listen on 5984, 4369 and whatever distribution port > you like on its own IP. > > What you don’t get with Pods is a hostname that’s discoverable from other > Pods in the cluster. A “Service” (a replicated, load-balanced collection of > Pods) can optionally have a DNS name, but the Pods themselves do not. This > throws a wrench in the most common distributed Erlang setup, where each > node gets a name like “couchdb@FQDN” and the FQDNs are resolvable to IP > addresses via DNS. > > It is certainly possible to specify an Erlang node name like “ > couchdb@12.34.56.78 <mailto:couchdb@12.34.56.78>”, but we need to be a > bit careful here. CouchDB is currently forcing the Erlang node name to do > “double-duty”; it’s both the way that the nodes in a cluster figure out how > to route traffic to one another and it’s the identifier for nodes to claim > ownership over individual replicas of database shards in the shard map. > Speaking from experience it’s often quite useful operationally to remap a > given Erlang node name to a new server and have the new server be > automatically populated with the replicas it’s supposed to own. If we use > the Pod IP in Kubernetes for the node name we won’t have that luxury. > > I think the best path forward here would be to extend the “Node" concept > in a CouchDB cluster so that it has an identifier which is allowed to be > distinct fro the Erlang node name. The “CouchDB Node” is the one that owns > database shard replicas, and it can be remapped to different distributed > Erlang nodes over time via modification of an attribute in the _nodes DB. > > Hope you all found this useful — I’m quite interested in finding way to > make it easier for users to acquire a highly-available cluster configured > in the “right way”, and I think projects like Kubernetes have a lot of > promise in this regard. Cheers, > > Adam
