Exporting Application Metrics
Exposing useful metrics is critical to understanding what is happening with your software in production. Without this quantifiable data, it is almost impossible to manage and develop your application intelligently. This guide covers how to expose metrics from your app for collection by Prometheus.
What Makes a Good Metric
A good metric provides quantifiable measurements on a time series that helps you understand:
- Application Performance
- Resource Consumption
This category is often expressed as “user experience” and encompasses measurements that indicate if users or client apps are getting what they should reasonably expect from the application. This also includes metrics that can affect the user experience indirectly and help identify the root cause of problems. Examples include:
- Apdex score
- Total request count
- Jobs processed (for batch-type workloads)
- Queue length and wait times (for batch-type workloads)
- Request durations
- Count on number 2xx responses sent back to clients
- Count on number 5xx responses sent back to clients
- DNS query time
- Garbage collection frequency and duration
- Payloads processed and their size
Resource consumption is important for two reasons: Capacity planning and cost management. How utilized is my application? In other words, how much more traffic can my application handle without scaling vertically or horizontally? How much infrastructure will need to be provisioned for increases in traffic or availability? How much does it cost the business to run a given workload? These are questions that can be answered, at least in part, by metrics such as:
- CPU usage
- Memory consumption
- Disk I/O usage
- Network bandwidth usage
- Pod replica counts
Implementing an Exporter
There are two methods to expose metrics from the software you run:
- Third-Party Exporters: These are distinct workloads that collect metrics from your running application and expose them to Prometheus. This is a good solution when leveraging open-source software such as nginx or redis as a part of your solution. These can often be run as sidecar containers.
- Instrumented Metrics: This option involves instrumenting your apps using the Prometheus client libraries or utilities so that exposing metrics is natively supported. This is a good solution if you are developing apps that are designed for and intended to run on Kubernetes.
The Prometheus docs offer an excellent source for finding third-party exporters as well as software that natively exposes Prometheus metrics.
Implementation in Java
Prometheus officially maintains a JMX exporter which can be run as a Java Agent for JVM-based applications.
Implementation in Python
If using the popular Django framework, consider the unofficial django-prometheus python library.
Implementation in Go
For Go applications, the official client_golang library is highly recommended. It’s important to note that this project includes both an instrumentation library as well as a client library for applications that need to query metrics from a Prometheus server, which is a different concern. The repo includes examples for instrumenting but other excellent examples exist in etcd and in Kubernetes
Implementation in Node.js
There is not an official client library for node.js, however there is an unofficial third-party prom-client project that has attracted considerable community participation. It supports each of the metric types addressed below, includes clear examples and good documentation.
Types of Metrics
Prometheus client libraries generally use four metric types. It is helpful to understand how each is used when instrumenting an application to expose Prometheus metrics.
A counter is a metric that can only increase. It is useful for things like total number of requests, error counts or any accumulating events.
A gauge is used for numerical values that can increase and decrease. Current resource usage and pod replica counts are examples of where you would use gauges.
Summaries track the size and number of events. An example would be DNS queries where the size would be the duration of the queries and the number would be the count of DNS queries. Another example would be for garbage collection where size is the duration of garbage collection events and number would be the number of times garbage collection occurred.
Histograms track the size and number of events and organize them into buckets. An Apdex score, for example, could be based on a histogram using buckets of request durations. These request durations should be defined according to the service level objectives (SLO) for the application. Define what is desired and tolerable, then use histograms to readily determine if your objectives are being met. It also provides convenient mechanisms upon which to alert if values violate tolerable thresholds.
Scraping Metrics With Prometheus
As an application developer, you will preferably have a Prometheus monitoring system available as a part of the platform. If so, when you deploy your application, you will have to tell Prometheus where to scrape your app’s metrics from.
Prometheus will need a scrape config to find your app’s metrics. Given this config Prometheus would scrape metrics from https://samplehost:8000/metrics:
# The job name assigned to scraped metrics by default. job_name: sample-app # List of statically configured targets where metrics will be scraped. static_configs: targets: - "samplehost:8000" # The HTTP resource path on which to fetch metrics from targets. metrics_path: /metrics # Default to scraping over https. If required, just disable this or change to `http`. scheme: https # Configures the scrape request's TLS settings. tls_config: # CA certificate to validate API server certificate with. ca_file: /path/to/ca.crt # Certificate and key files for client cert authentication to the server. cert_file: /path/to/sample-app.crt key_file: /path/to/sample-app.key
While it’s helpful to understand what the scrape config for Prometheus consists of, in a Kubernetes environment, it is highly recommended that you use the Prometheus Operator to deploy the Prometheus instances and manage the scrape configs by way of the ServiceMonitor resource.
You may not be responsible for deploying the Prometheus server, but it’s helpful
to understand how it references the ServiceMonitor resource which will be used
by the operator to configure Prometheus on your behalf. In the Prometheus
Operator’s Prometheus resource, the
serviceMonitorSelector is used to
associate Service Monitors with a Prometheus instance.
Consider this example of a Prometheus resource used to deploy a Prometheus server:
apiVersion: monitoring.coreos.com/v1 kind: Prometheus metadata: name: prometheus spec: serviceAccountName: prometheus serviceMonitorSelector: matchLabels: team: samples # references ServiceMonitor's label
Here is an example of a ServiceMonitor resource manifest you could deploy along with the other resources for your app to get the metrics scraped by the Prometheus server. The Service Monitor uses a label selector to identify Services and the associated Endpoint objects.
apiVersion: monitoring.coreos.com/v1 kind: ServiceMonitor metadata: name: sample-app labels: team: samples # used by Prometheus resource to associate this ServiceMonitor spec: namespeceSelector: matchNames: - sample-namespace selector: matchLabels: app: sample-app # this label must be on sample-app's service endpoints: - port: web # the named port on the Service from which to scrape path: /metrics