Dependencies Maps
A Dependencies map draws which Kubernetes workloads, or which Linux hosts, talk to each other and how much traffic flows between them. It covers any TCP or UDP traffic, including databases, caches and other services that are not traced. The data comes from the eBPF network flow metric obi_network_flow_bytes.
Reading the map
- Cards are the ends of a connection: workloads on the Kubernetes map, hosts and named networks on the hosts map.
- Lines are connections. A thicker line carries more bytes per second.
- Arrows point from the client (the side that opened the connection) to the server. A grey line means the direction is unknown, or both sides open connections to each other.
- Hover or select a line to see its rate. Select a line to open a side panel with the rate in each direction and the split between request, response and unknown direction.
- The Connections list gives keyboard and screen-reader access to the same connections.
Rates are approximate bytes per second, averaged over the selected time range up to its end. The averaging window is at least 5 minutes and at most 24 hours. Connections below 1 B/s are hidden until you turn on Low-traffic connections.
The map draws at most 80 cards and 200 connections, busiest first. When it drops some, it says how many of the total it shows.
Kubernetes workloads
Go to Infra > Kubernetes, and in the resource tree select Cluster > Dependencies. The map covers the selected cluster.
- One card per workload, such as a Deployment, StatefulSet, DaemonSet or CronJob. A card's colour is its namespace.
- Node cards stand for host processes and host-network pods on a node.
- External or node-local stands for addresses outside the cluster, or traffic inside a node such as kubelet probes.
- Click a workload card to open its details panel.
Filter the map with:
- The namespace filter and Search workloads.
- System namespaces: shows
kube-system,kube-node-lease,openobserve-collector,cert-managerandopentelemetry-operator-system. Off by default. - Node-to-node traffic: shows traffic between nodes. Off by default.
- Low-traffic connections: shows connections below 1 B/s. Off by default.
- Grouping: Auto groups workloads by namespace once the map has more than 40 endpoints, Collapsed always groups and Expanded never does. Click a namespace card to show its workloads. Grouping is off while a namespace filter or search is active.
The switches are kept in the URL, so you can share the view.
Linux hosts
Go to Infra > Hosts and select the Dependencies tab next to Table.
Each card is a name from the agent's --network-flows list, not a raw address:
- Host: a name that matches a host name. Click it to open the host's details.
- Network: a range of addresses named in the
--network-flowslist. - Unnamed addresses: addresses that no
--network-flowsentry covers.
Turn on Low-traffic connections to include connections below 1 B/s. The selected tab and the switch are kept in the URL.
For every host to show up, install the agent on every host with the same --network-flows list, and give each host's own address (/32) an entry named after its host name. A name that covers several hosts may show only the traffic of the busiest of them.
Set up network flows
When no flow data is found, the map shows No network flow data or No network flows in this time range. Click Set up network flows to open the matching data source setup.
Kubernetes
On the Data sources page, open the Kubernetes setup and follow the optional step See which workloads talk to each other: eBPF network flows. It upgrades the OpenObserve collector to chart 0.8.0 or later with eBPF network flows turned on:
helm repo update openobserve
helm --namespace openobserve-collector upgrade o2c openobserve/openobserve-collector \
--version '>=0.8.0' --reset-then-reuse-values \
--set ebpf.enabled=true \
--set ebpf.network.enabled=trueRequirements: Helm 3.14+ and OpenTelemetry operator v0.142.0+. The openobserve-collector namespace must allow privileged pods. Nodes must run Linux on amd64 or arm64 with kernel 5.8+ (or RHEL-family 4.18 with eBPF backports) and BTF. Port 8888 on 127.0.0.1 of each node must be free. Re-running the collector install command turns flows off again, so run the upgrade again afterwards.
Linux
On the Data sources page, open the Linux setup and follow the step See which hosts talk to each other: eBPF network flows. Enter the networks in Networks to label as comma-separated CIDR or CIDR=NAME entries, for example 10.0.1.5/32=web-1,10.0.1.6/32=db-1. The install command then includes --network-flows '<list>'.
Requirements: Linux 5.8+ (or RHEL-family 4.18 with eBPF backports) with BTF. The agent then runs as root. Each run of the installer replaces the agent's configuration, so add your tracing flags again if you use eBPF tracing.
The same bytes can be reported more than once: traffic between nodes or hosts is seen by both ends. The Kubernetes map counts each node pair once. If your collector does not send the source and destination node labels (k8s_src_node_name, k8s_dst_node_name), a workload pair spread over several nodes can be undercounted.