Super-cluster with Gateways
Gateways enable connecting one or more clusters together into a full mesh; they allow the formation of superclusters from smaller clusters. Cluster and Gateway protocols listen on different ports. Clustering is used for adjacent servers; gateways are for joining clusters together.
Gateway configuration is similar to clustering:
- gateways have a dedicated port where they listen for gateway requests
- gateways gossip gateway nodes and remote discovered gateways
Unlike clusters, gateways:
- have names, specifying the cluster they are in
- don't form a full mesh between gateway nodes, but form a full mesh between cluster instead
- are bound by uni-directional connections
- don't gossip gateway nodes to clients
Gateways exist to:
- reduce the number of connections required between servers
- optimize the interest graph propagation
If gateways are to be used in a cluster, all servers of this cluster need to have a gateway configuration with the same name. Furthermore, every gateway node needs to be able to connect to any other gateway node and vice versa. Everything else is considered a misconfiguration.
A nats-server in a gateway role will specify a port where it will accept gateway connections. If the configuration specifies other external
gateways, the gateway will create one outbound gateway connection for each gateway in its configuration. It will also gossip other gateways it knows or discovers. Fewer external
gatewaysmean less configuration. Yet, the ability to discover more gateways and gateway nodes depends on these servers running. This is similar to seed server in cluster. It is recommended to have all seed server of a cluster listed in the
If the local cluster has three gateway nodes, this means there will be three outbound connections from the local cluster to each external gateway cluster.
In the example above cluster A has configured gateway connections for B (solid lines). B has discovered gateway connections to A (dotted lines). Note that the number of outgoing connections always matches the number of gateways with the same name.
Gateway Discovered Gateways
In this second example, again configured connections are shown with solid lines and discovered gateway connections are shown using dotted lines. Gateways A and C were both discovered via gossiping; B discovered A and A discovered C.
A key point in the description above is that each node in the cluster will make a connection to a single node in every remote cluster — a difference from the clustering protocol, where every node is directly connected to all other nodes.
For those mathematically inclined, cluster connections are
N(N-1)/2where N is the number of nodes in the cluster. On gateway configurations, outbound connections are the summation of
Ni(M-1)where Ni is the number of nodes in a gateway i, and M is the total number of gateways. Inbound connections are the summation of
U-Niwhere U is the sum of all gateway nodes in all gateways, and N is the number of nodes in a gateway i. It works out that both inbound and outbound connection counts are the same.
The number of connections required to join clusters using clustering vs. gateways is apparent very quickly. For 3 clusters, with N nodes:
A cluster section is not needed for gateways, they work with single server as well. Yet, they start to be useful when participating cluster consist of more than one server and they reduce the number of connections.
Messages from clients directly connected to a gateway node will be sent along outgoing gateway connections according to the following three interest propagation mechanisms:
- Optimistic Mode
- Interest-only Mode
- Queue Subscriptions
Local interest permitting, the receiving gateway node sends the messages directly to its subscribing clients as well as server within the cluster.
When a publisher in A publishes "foo", the A gateway will check if cluster B has registered no interest in "foo". If not, it forwards "foo" to B. If upon receiving "foo", B has no subscribers on "foo", B will send a gateway protocol message to A expressing that it has no interest on "foo", preventing future messages on "foo" from being forwarded.
Should a subscriber on B create a subscription to "foo", B knowing that it had previously rejected interest on foo, will send a gateway protocol message to cancel its previous no interest on "foo" in A.
When a gateway on A sends many messages on various subjects for which B has no interest. B sends a gateway protocol message for A to stop sending optimistically, and instead send if there's known interest in the subject. As subscriptions come and go on B, B will update its subject interest with A.
When a queue subscriber creates a new subscription, the gateway propagates the subscription interest to other gateways. The subscription interest is only propagated once per Account and subject. When the last queue subscriber is gone, the cluster interest is removed.
Queue subscriptions work on Interest-only Mode to honor NATS' queue semantics across the Super Cluster. For each queue group, a message is only delivered to a single queue subscriber. If the same queue group exists in multiple clusters, the server will pick one member from the queue group in its cluster, only sending to a different cluster if there is no interest in its cluster. In other words, a server will always try to serve local queue subscribers first and only failover when a local queue subscriber is not found. The server will pick the cluster with the lowest RTT.