Service Discovery

The service registry is the central database or directory that tracks the network locations and metadata of all available agents or services. It is the authoritative source for discovery queries. Agents register upon startup and deregister upon shutdown. Common implementations include etcd (used by Kubernetes), Consul, and Apache ZooKeeper. The registry must be highly available and partition-tolerant to prevent system-wide outages.

This is the two-part process that keeps the service registry accurate.

• Dynamic Registration: Agents automatically register their network endpoint (IP and port) and capability advertisements upon startup.
• Health Maintenance: A heartbeat mechanism or periodic health check confirms an agent is alive. This is often managed via a lease mechanism; if an agent fails to renew its lease (e.g., due to a crash), it is automatically deregistered after a timeout, preventing traffic from being sent to failed instances.

There are two primary architectural patterns for how a client uses the registry:

• Client-Side Discovery: The service consumer (client) queries the registry directly to obtain a list of available instances and is responsible for load balancing requests among them. This offers more client control but couples clients to the registry library.
• Server-Side Discovery: The client sends a request to a stable intermediary (like an API Gateway or load balancer). This intermediary queries the registry and handles routing. This decouples the client but introduces a central routing component.

A service mesh (e.g., Istio, Linkerd) abstracts service discovery and other networking concerns into a dedicated infrastructure layer. It uses the sidecar pattern, deploying a proxy (like Envoy Proxy) alongside each service instance. The sidecar handles all communication, automatically discovering services via the mesh's control plane. This provides uniform observability, security, and traffic management without requiring changes to application code.

This approach leverages the Domain Name System (DNS) for discovery, providing a familiar and standardized interface.

• DNS-SD (DNS-Based Service Discovery): Uses standard DNS record types (SRV, TXT) to advertise a service's location, port, and metadata. Clients perform DNS queries to discover services.
• mDNS (Multicast DNS): Used in local networks without a dedicated DNS server. Agents broadcast their presence via multicast, enabling zero-configuration discovery. This is common in IoT and local device networks.

Beyond simple location lookup, advanced discovery involves finding agents based on their functional attributes. A capability query allows a client to search the registry for agents that match specific interfaces, supported protocols, or performance characteristics (advertised as part of a Service-Level Agreement (SLA)). This is critical in multi-agent systems where agents are heterogeneous specialists, and a workflow engine needs to find an agent that can perform a very specific task.

A health check is a periodic probe sent to an agent to verify its operational status and availability for receiving requests. It is critical for maintaining registry accuracy.

• Prevents routing traffic to failed or degraded agents.
• Typically involves an HTTP endpoint, TCP connection attempt, or custom script.
• Failures trigger automatic deregistration from the registry.

These are two fundamental patterns for where the discovery logic resides.

• Client-Side Discovery: The service consumer (client) queries the registry directly and selects an instance. This adds logic to the client but reduces hops. Example: Netflix Eureka client.
• Server-Side Discovery: An intermediary (like a load balancer or API gateway) queries the registry. The client sends requests to the intermediary, which handles routing. This simplifies clients but introduces a central point.

Protocols that leverage the Domain Name System for zero-configuration discovery in local networks.

• DNS-SD (DNS-Based Service Discovery): Uses standard DNS SRV and TXT records to advertise and discover services. Defined in RFC 6763.
• mDNS (Multicast DNS): Resolves hostnames to IP addresses within small networks without a dedicated DNS server. The basis for Apple's Bonjour. Enables 'plug-and-play' networking for agents.

Coupled mechanisms that ensure registry entries are current and not stale.

• Lease: A time-bound grant of registration. An agent's entry is valid only for the lease duration.
• Heartbeat: A periodic signal (often a renewal request) sent by the agent to the registry to refresh its lease.
• If heartbeats stop, the lease expires and the agent is automatically deregistered, providing failure detection.