Network Cost Configuration


The network costs daemonset is an optional utility that gives Kubecost more detail to attribute costs to the correct pods.
When networkCost is enabled, Kubecost gathers pod-level network traffic metrics to allocate network transfer costs to the pod responsible for the traffic.
See this doc for more detail on network cost allocation methodology.
The network-costs metrics are collected using a daemonset (one pod per node) that uses source and destination detail to determine egress and ingress data transfers by pod and are classified as internet, cross-region and cross-zone.


With the network-costs daemonset enabled, the Network column on the Allocations page will reflect the portion of network transfer costs based on the chart-level aggregation.
When using Kubecost version 1.99 and above: Greater detail can be accessed through Allocations UI only-when aggregating by namespace and selecting the link on that namespace. This opens the namespace detail page where there is a card at the bottom.

Grafana Dashboard

To view the raw network transfer data, Grafana dashboard that is almost ready to publish (just needs QA testing to verify accuracy):

Enabling network costs

To enable this feature, set the following parameter in values.yaml during Helm installation:
enabled: true

Cloud Provider Service Tagging

Service tagging allows kubecost to identify network activity between the pods and various cloud services (e.g. AWS S3, EC2, RDS, Azure Storage, Google Cloud Storage).
To enable this, set the following Helm values:
# google-cloud-services: when set to true, enables labeling traffic metrics with google cloud
# service endpoints
google-cloud-services: false
# amazon-web-services: when set to true, enables labeling traffic metrics with amazon web service
# endpoints.
amazon-web-services: false
# azure-cloud-services: when set to true, enables labeling traffic metrics with azure cloud service
# endpoints
azure-cloud-services: false
# user defined services provide a way to define custom service endpoints which will label traffic metrics
# falling within the defined address range.
# - service: "test-service-1"
# ips:
# - ""
# - service: "test-service-2"
# ips:
# - ""
# - ""

Additional configuration

You can view a list of common config options here.
If using the included Prometheus instance, the scrape is automatically configured.
If you are integrating with an existing Prometheus, you can set networkCosts.prometheusScrape=true and the network costs service should be auto-discovered.
Alternatively a serviceMonitor is also available.
Note: Network cost, which is disabled by default, needs to be run as a privileged pod to access the relevant networking kernel module on the host machine.

Resource limiting

In order to reduce resource usage, Kubecost recommends setting a CPU limit on the network-costs daemonset. This will cause a few seconds delay during peak usage and does not effect overall accuracy. This is done by default in Kubecost 1.99+.
For existing deployments, these are the recommended values:
cpu: 500m
cpu: 50m
memory: 20Mi

Benchmarking metrics

The network-simulator was used to real-time simulate updating conntrack entries while simultaneously running a cluster simulated network-costs instance. To profile the heap, after a warmup of roughly five minutes, a heap profile of 1,000,000 conntrack entries was gathered and examined.
Each conntrack entry is equivalent to two transport directions, so every conntrack entry is two map entries (connections).
After modifications were made to the network-costs to parallelize the delta and dispatch, large map comparisons were significantly lighter in memory. The same tests were performed against simulated data with the following footprint results.
Benchmarking metrics

Kubernetes network traffic metrics

The primary source of network metrics is a DaemonSet Pod hosted on each of the nodes in a cluster. Each daemonset pod uses hostNetwork: true such that it can leverage an underlying kernel module to capture network data. Network traffic data is gathered and the destination of any outbound networking is labeled as:
  • Internet Egress: Network target destination was not identified within the cluster.
  • Cross Region Egress: Network target destination was identified, but not in the same provider region.
  • Cross Zone Egress: Network target destination was identified, and was part of the same region but not the same zone.
These classifications are important because they correlate with network costing models for most cloud providers. To see more detail on these metric classifications, you can view pod logs with the following command:
kubectl logs kubecost-network-costs-<pod-identifier> -n kubecost
This will show you the top source and destination IP addresses and bytes transferred on the node where this Pod is running. To disable logs, you can set the helm value networkCosts.trafficLogging to false.

Overriding traffic classifications

For traffic routed to addresses outside of your cluster but inside your VPC, Kubecost supports the ability to directly classify network traffic to a particular IP address or CIDR block. This feature can be configured in values.yaml under networkCosts.config. Classifications are defined as follows:
  • In-zone: A list of destination addresses/ranges that will be classified as an in-zone traffic, which is free for most providers.
  • In-region: A list of addresses/ranges that will be classified as the same region between source and destinations but different zones.
  • Cross-region: A list of addresses/ranges that will be classified as the different region from the source regions
# In Zone contains a list of address/range that will be
# classified as in zone.
# Loopback Addresses in "IANA IPv4 Special-Purpose Address Registry"
- ""
# IPv4 Link Local Address Space
- ""
# Private Address Ranges in RFC-1918
- "" # Remove this entry if using Multi-AZ Kubernetes
- ""
- ""
# In Region contains a list of address/range that will be
# classified as in region. This is synonymous with cross
# zone traffic, where the regions between source and destinations
# are the same, but the zone is different.
in-region: []
# Cross Region contains a list of address/range that will be
# classified as non-internet egress from one region to another.
cross-region: []
# Direct Classification specifically maps an ip address or range
# to a region (required) and/or zone (optional). This classification
# takes priority over in-zone, in-region, and cross-region configurations.
direct-classification: []
# - region: "us-east1"
# zone: "us-east1-c"
# ips:
# - ""


To verify this feature is functioning properly, you can complete the following steps:
  1. 1.
    Confirm the kubecost-network-costs Pods are Running. If these Pods are not in a Running state, kubectl describe them and/or view their logs for errors.
  2. 2.
    Ensure kubecost-networking target is Up in your Prometheus Targets list. View any visible errors if this target is not Up. You can further verify data is being scrapped by the presence of the kubecost_pod_network_egress_bytes_total metric in Prometheus.
  3. 3.
    Verify Network Costs are available in your Kubecost Allocation view. View your browser's Developer Console on this page for any access/permissions errors if costs are not shown.

Common issues

  • Failed to locate network pods: Error message displayed when the Kubecost app is unable to locate the network pods, which we search for by a label that includes our release name. In particular, we depend on the label app=<release-name>-network-costs to locate the pods. If the app has a blank release name this issue may happen.
  • Resource usage is a function of unique src and dest IP/port combinations. Most deployments use a small fraction of a CPU and it is also ok to have this Pod CPU throttled. Throttling should increase parse times but should not have other impacts. The following Prometheus metrics are available in v15.3 for determining the scale and the impact of throttling:
kubecost_network_costs_parsed_entries is the last number of conntrack entries parsed kubecost_network_costs_parse_time is the last recorded parse time

Feature limitations

  • Today this feature is supported on Unix-based images with conntrack
  • Actively tested against GCP, AWS, and Azure
  • Pods that use hostNetwork share the host IP address