Careful!
You are browsing documentation for a version of Kuma that is not the latest release.
Service Health Probes
Kuma is able to track the status of the Envoy proxy and the underlying app.
Compared to HealthCheck
policies, Health Probes have the following advantages:
- knowledge about health is propagated back to
kuma-cp
and is visible in both in Kuma GUI and Kuma CLI - scalable with thousands of data plane proxies
Unlike HealthCheck
policies, Health Probes:
- only updates when
kuma-cp
is up and running and the proxy can connect to it. - doesn’t check connectivity between data plane proxies.
Every inbound
in the Dataplane
model has a health
section:
type: Dataplane
mesh: default
name: web-01
networking:
address: 127.0.0.1
inbound:
- port: 11011
servicePort: 11012
health:
ready: true
tags:
kuma.io/service: backend
kuma.io/protocol: http
This health.ready
status is intended to show the status of the endpoint itself.
It is set differently depending on the environment (Kubernetes or Universal).
It’s treated the same way regardless of the environment:
- if
health.ready
is true orhealth
section is missing - Kuma considers the inbound as healthy and includes it into load balancing set. - if
health.ready
is false - Kuma doesn’t include the inbound into the load balancing set.
Also, health.ready
is used to compute the status of the data plane proxy and service. You can see these statuses both in Kuma GUI and Kuma CLI:
Data plane proxy health follows these rules:
- if proxy status is
Offline
, then data plane proxy isOffline
. - if proxy status is
Online
:- if all inbounds have
health.ready=true
or nohealth
then data plane proxy isOnline
- if all inbounds have
health.ready=false
then data plane proxy isOffline
- if at least one of the inbounds has
health.ready=false
then data plane proxy isPartially degraded
- if all inbounds have
Service health is computed by aggregating the health of all data plane proxies that have one inbound with a given kuma.io/service
and is set this way:
- if all data plane proxies are
Online
then the service isOnline
- if no data plane proxy is
Online
then the service isOffline
- if at least one of the data plane proxies is
Offline
arePartially degraded
then the service isPartially degraded
Kubernetes
Kuma leverages container statuses within the pod to evaluate dataplane.inbound.health
and will follow these rules:
- If the sidecar container is not ready then all inbound will have
health=false
. - If the side container is ready then the health of the inbound will be whatever is the readiness of the container which exposes the port used by the inbound.
Kubernetes Probes
Native probe functionalities are not supported until 2.9 release. HTTPGet
and gRPC
probes always fail when mTLS enabled and TCPSocket
probes always succeed regardless of mTLS.
Virtual Probes
For better lifecycle management Kubernetes has readiness, liveness and startup probes.
When using mTLS the container ports are not available outside the pod and Kubernetes will fail to check the probes. To work around this Kuma generates a special non-mTLs listener and overrides the probe definitions in the Pod to proxy them through this sidecar listener. These are called Virtual probes.
This feature currently only supports httpGet
probes.
For example, if we specify the following probe:
livenessProbe:
httpGet:
path: /metrics
port: 3001
initialDelaySeconds: 3
periodSeconds: 3
Kuma will replace it with:
livenessProbe:
httpGet:
path: /3001/metrics
port: 9000
initialDelaySeconds: 3
periodSeconds: 3
Where 9000
is a default virtual probe port, this is configurable:
With the config yaml:
runtime:
kubernetes:
injector:
virtualProbesPort: 19001
or environment variable: KUMA_RUNTIME_KUBERNETES_VIRTUAL_PROBES_PORT=19001
You can also disable Kuma’s probe virtualization:
With the config yaml:
runtime:
kubernetes:
injector:
virtualProbesEnabled: false
or environment variable: KUMA_RUNTIME_KUBERNETES_VIRTUAL_PROBES_ENABLED=false
Universal probes
On Universal there is no single standard for probing the service.
For health checking, the Dataplane
status on Universal Kuma uses Envoy’s Health Discovery Service (HDS).
Envoy does health checks and reports the status back to the Kuma Control Plane.
In order to configure health checks for your Dataplane
you have to update inbound
config with serviceProbe
:
type: Dataplane
mesh: default
name: web-01
networking:
address: 127.0.0.1
inbound:
- port: 11011
servicePort: 11012
serviceProbe:
timeout: 2s # optional (default value is taken from KUMA_DP_SERVER_HDS_CHECK_TIMEOUT)
interval: 1s # optional (default value is taken from KUMA_DP_SERVER_HDS_CHECK_INTERVAL)
healthyThreshold: 1 # optional (default value is taken from KUMA_DP_SERVER_HDS_CHECK_HEALTHY_THRESHOLD)
unhealthyThreshold: 1 # optional (default value is taken from KUMA_DP_SERVER_HDS_CHECK_UNHEALTHY_THRESHOLD)
tcp: {}
tags:
kuma.io/service: backend
kuma.io/protocol: http
If there is a serviceProbe
configured for the inbound, Kuma will automatically fill the inbound.health
section and update it
with the interval equal to KUMA_DP_SERVER_HDS_REFRESH_INTERVAL
.
Alternatively, it’s possible to omit a serviceProbe
section and develop custom automation that periodically updates the health
of the inbound.
If the grpc stream with Envoy is disconnected then Kuma considers this proxy offline.
It will, however, still advertise inbounds using the final update on their health before disconnection.
This is to avoid connection issues between kuma-cp
and kuma-dp
blocking data plane traffic.
Additionally, when serviceProbe
is defined, probes takes into account a health of Envoy.
When kuma-dp receives the first shutdown signal, it goes into draining state and all inbounds are considered unhealthy.