Advertised Address
A Kafka client connects in "two hops": first to any broker in bootstrap.servers to fetch metadata, then directly to the target broker (partition leader / coordinator) using the address the metadata provides. So the address a broker tells the client must be actually reachable from that client — this is what the "advertised address" solves.
Bind address vs advertised address
| Concept | Meaning |
|---|---|
| Bind address | What the socket actually listens on. Often 0.0.0.0:9092 in containers. |
| Advertised address | The host:port told to clients via Metadata / FindCoordinator. |
The two may differ. Clients never use the bind address — only the advertised one.
How RobustMQ registers and advertises
This is a shared mechanism across protocols: at startup a node writes each protocol's external address into the meta layer's node registry (NodeExtend). Kafka's entry is kafka.tcp_addr (MQTT and NATS each have their own fields, going through the same registration / consumption path).
- Advertised address =
broker_ip+kafka_runtime.tcp_port. - When
broker_ipis unset, the auto-detected local IP is used. - Both
Metadata(broker list, partition leaders) andFindCoordinator(coordinator) readkafka.tcp_addrfrom the registry to tell clients which node to connect to directly.
Per-environment guidance
| Scenario | Recommendation |
|---|---|
| Local / dev | The default broker_ip = "127.0.0.1" is fine. |
| Multi-host cluster in one subnet | Set broker_ip to the node's cluster-reachable IP. |
| Docker | If clients are on the host, set broker_ip to a host-reachable address, not the container's internal IP. |
| Kubernetes | Set it to an address reachable by clients for that Service / Pod; each broker needs its own stable, routable address. |
| NAT / public | Set it to the public address reachable from the client side. |
Common failure
If the advertised address is set to something unreachable from the client (e.g. a container-internal IP, or 127.0.0.1 connected from a remote host), the symptom is usually: bootstrap connects and metadata is fetched, but the subsequent direct connect to the partition leader / coordinator times out or is refused. When debugging, first verify the host:port returned by Metadata is ping/telnet-reachable from the client.
Related
- The advertised port comes from
kafka_runtime.tcp_port, see Broker Configuration. - How coordinators use the registry to resolve, see Cluster & Controller.
- Address setup during deployment, see Deployment.