Healthchecks are essential to signal whether a service is up and running normally. In a microservices-based architecture, health checks are used to discover healthy service instances. It is vital to have an accurate health check indicator for the reliability of our applications.
A client-library, load-balancer, or service discovery co-ordinator (such as consul) takes care of periodically checking each instance’s health. When connecting to the service, the library or load-balancer checks the instances’ status and binds only to healthy service instances. A monitoring service may also rely on the health indicator to fire-off alerts.
How to design meaningful application health checks
Your application health check should be a boolean function — whether healthy or not. If your application X depends on A, B, Y, Z for each and every operation, where each of them may be databases or other microservices, then the health of X H(X) should be the aggregate of the health of A, B, Y, and Z. More formally,
H(X) = H(A) & H(B) & H(Y) & H(Z)
Partial functioning and optional dependencies
In a health check, only include the mandatory dependencies. For example, if X uses A only for a few functions and can provide functionality if A is down, then we can exclude it from being a mandatory part of the health check.
H(X) = H(A) | H(B) & H(Y) & H(Z)
Ideally, you should have a fallback strategy if A is unavailable.
You can define a health check endpoint, say (
/healthcheck) for a rest service. Simply blindly returning 200 is a reachability indicator and not a health check. The success should be an indicator of whether the service instance can process requests.
- Returns a 200 response when the service is healthy and can serve traffic.
- Return 503 is service is unavailable deliberately. This is useful for scenarios like graceful shutdown and planned restart.
- Return 500 if the health check fails due to a dependent health check.
Determining the health check status
- Determine the status of each of the dependent services and external APIs. If the dependent service is another microservice, you can call it’s
/healthcheckHowever, this may cause an explosion of health check traffic. You can query a service discovery co-ordinator (such as consul) or the application load-balancer (if it uses one) to determine the overall health of the microservice. You can also consider a TTL, for example, if B's health check passed within the last 5 minutes, assume B to be healthy.
- For SQL databases, you can use a ping query such as
SELECT 1. This helps you to determine reachability. You can also use database-specific readiness SQL queries as well.
- Redis and other databases also offer similar
Using health checks
You can configure the health check endpoints in load-balancers and service discover co-ordinator to discover healthy instances and route traffic.
If you are using Kubernetes, you can configure liveliness and readiness probes. This signals Kubernetes to not send traffic until the pods have started successfully. If the probe fails, Kubernetes can restart the application.
Many frameworks come with libraries that can assist you in implementing them. If you are using Spring, they provide a library called Actuator that automatically configures health checks for associated resources and allows you to add custom checks too.