Skip to content
Moonin Documentation

Moonin Node Agent

Moonin Node Agent is a Linux Kubernetes DaemonSet that runs on each selected worker node. It observes runtime activity with eBPF and application-runtime probes, associates it with Kubernetes workloads, and sends the resulting telemetry to Moonin.

Unlike Discovery Agent, which maintains Kubernetes inventory, Node Agent observes how workloads communicate and behave while they run. It does not modify Deployments, HPAs, or application traffic.

What it captures

Section titled “What it captures”

Node Agent enriches the following signals with available cluster, namespace, workload, pod, and container context:

Signal Examples of captured context
HTTP, HTTPS, HTTP/2 and gRPC Method, path, protocol, status, latency, peer and workload context
Database activity Connections and queries for supported PostgreSQL, MySQL, MongoDB and Redis clients
Google Cloud activity Supported Pub/Sub, Firestore, Secret Manager, BigQuery, Cloud Storage, Cloud Tasks, Eventarc, reCAPTCHA and Vertex AI operations
Distributed traces Server and downstream spans derived from observed requests and supported trace exports
Runtime errors OOM kills, fatal signals, termination signals and non-zero process exits associated with workloads
Network activity TCP state transitions and aggregated UDP activity
Workload metrics CPU, memory, throttling, network and restart context collected from cgroups and /proc

!!! note Container-log upload is disabled in the default chart. The agent can send runtime error signals, but this does not mean that all container logs are collected.

How data reaches Moonin

Section titled “How data reaches Moonin”
flowchart LR
    A[Workloads on Linux nodes] --> B[Moonin Node Agent]
    C[Kubernetes API] --> B
    B --> D[Moonin Ingestion API]
    D --> E[app.moonin.app]

The agent batches event data before sending it. By default, it aggregates events for up to 180 seconds and dispatches batches every 60 seconds. Metrics are sent every 60 seconds. Data may therefore take a few minutes to appear after activity occurs.

The platform controls collection through a remote ingestion policy. Until Node Agent obtains a valid policy, it does not send telemetry. The policy can enable or disable collection, limit it to namespaces or services, exclude scopes, and apply sampling.

View captured data in app.moonin.app

Section titled “View captured data in app.moonin.app”

Select one cluster in the top-bar cluster selector before using the observability explorers. Availability also depends on the permissions assigned to the signed-in user.

Where What to use it for
/events Inspect raw HTTP, database, cache, and supported Google Cloud event families. Filter by time range, cluster, and available attributes, then expand an event for its captured context.
/traces Find a request by time, service, namespace, operation, duration, or error state. Open a trace to inspect its timeline, spans, workload context, and downstream calls.
/dependencies Explore observed outbound resources and identify services that consume a resource, URL, topic, bucket, or other dependency.
Service 360 Open a service from the service catalog to review its Events, Dependencies, and Metrics tabs in the workload context.

Typical permissions are observability.events.view, observability.traces.view, observability.dependencies.view, and observability.metrics.view.

!!! warning Network events are sent for platform processing but do not currently appear as a selectable family in the Events explorer. Error signals can be correlated with Moonin error workflows; they are not a substitute for a full container-log search.

Deployment and host requirements

Section titled “Deployment and host requirements”

Node Agent is enabled by default in the Moonin Agent chart and schedules on Linux worker nodes. Control-plane nodes are excluded by default.

It requires elevated host access to attach eBPF probes and attribute activity to workloads:

  • a privileged Linux container with hostPID and hostNetwork
  • host mounts for /proc, cgroups, kubelet pod metadata, kernel modules, BPF and tracing filesystems
  • read access to pods, services, endpoints, EndpointSlices, Jobs, and ReplicaSets for Kubernetes enrichment
  • cluster credentials (CLUSTER_ID and CLUSTER_TOKEN) and outbound HTTPS access to the Moonin ingestion API

The chart includes a privileged init container that can set kernel.perf_event_paranoid to -1 on the host when needed for eBPF performance events. Review these requirements with the cluster security owner before installation.

Privacy and collection boundaries

Section titled “Privacy and collection boundaries”
  • The agent excludes its own namespace and any namespaces excluded by its runtime configuration or ingestion policy.
  • HTTP headers and captured request or response text are redacted for common credentials and sensitive identifiers before transmission.
  • Event and metric collection is telemetry, not a traffic proxy: the agent observes supported runtime activity and does not route or modify application requests.
  • Trace delivery uses a bounded local spool for retries. Other event batches are memory-buffered and can be dropped if delivery cannot be completed.
Section titled “Related documentation”