Spring Metrics

Early work on dimensional Spring Metrics for application monitoring.

Reference Documentation

You can view an early preview of the reference documentation here. Lots more to come.

Goals

In addition to the features already present in Spring Boot Actuator Metrics, these are some goals for inclusion in an initial release:

• Adopt a dimensional system rather than hierarchical. Many metrics backends are dimensional in nature. A hierarchical name can be derived consistently from a series of dimensions for those that are not (example of this in spectator-reg-metrics3).
• Reduce instrumentation cost (in terms of CPU/memory overhead).
• Support timers, gauges, counters, and distribution summaries as our meter primitives.
• Configurable clock source for better test support.
• Support tiering metrics by priority to allow for different push intervals (the relatively fewer critical metrics are published at a shorter interval than less critical metrics).
• Support dynamically determined dimension values.
  • Similarly, allow dynamically determined, high-cardinality dimensions to be turned off application wide by property.
• Logging appenders to track the number of log messages and stack traces being emitted from the app.
• Fold in buffer pool and memory pool metrics reported by the JDK via JMX.
• Fold in information about GC causes.

Installing

Pre-release artifacts are being published frequently, but are NOT intended for production use.

In Gradle:

compile 'org.springframework.metrics:spring-metrics:latest.release'

Or in Maven:

<dependency>
  <groupId>org.springframework.metrics</groupId>
  <artifactId>spring-metrics</artifactId>
  <version>${metrics.version}</version>
</dependency>

Building Locally

Run ./gradlew pTML to publish a snapshot to your Maven local repo. To consume:

repositories {
    mavenLocal()
}

dependencies {
    compile 'org.springframework.metrics:spring-metrics:latest.integration'
}

Instrumentation

Defined: The API used to create and interact with meter instances, generally through a meter registry. An individual meter holds one or more metrics depending on its type.

Existing API: Metric, ½ of MetricWriter, CounterService, GaugeService

Support for Prometheus and Spectator collectors is in development.

A TCK verifying collector implementation correctness is provided in org.springframework.metrics.instrument.

Support targets:

• Spectator
• Prometheus Java client
• Dropwizard Metrics (hierarchical)
• StatsD and Datadog StatsD

Exporters

Defined: Exporters identify export-worthy metrics from a meter registry, transform (e.g. convert monotonically increasing counters to a rate), add global dimensions, and push metrics to a metrics backend on a periodic interval(s).

Existing API: Exporter, @ExportMetricWriter, @ExportMetricReader, MetricCopyExporter, MetricReader, AggregateMetricReader, ½ of MetricWriter

• servo-atlas
• Dropwizard Metrics "Reporters"
• Datadog StatsD (also acts as a collector)
• Direct to Datadog HTTP API
• JMX
• PCF Metrics

Backends

Defined: A metrics backend is a time-series database. It may be a plain TSDB or contain optimizations for metrics specifically like:

• Tiered reduction in granularity of older metrics to save on storage
• Tiered persistence of metrics, prioritizing more recent metrics for fastest lookup and older metrics for cost
• Act like a circular buffer for time-series data to more strictly bound resource consumption

Metrics backends typically contain some metrics-specific query interface (e.g. PromQL, Atlas stack language). In some cases the metrics backend also happens to serve dashboards and alerting configuration that we are calling a metrics frontend. In other cases, these pieces are distinct.

Potential support targets:

• Atlas
• Prometheus (Pushgateway)
• Graphite
• InfluxDB
• Datadog
• OpenTSDB
• RRDtool?
• Ganglia
• Some supported statsd backends
• PCF Metrics
• JMX
• Message channel -- not a backend in itself of course, but could be linked to one indirectly?
• AWS CloudWatch
• Elasticsearch -- understand why projects like Orestes exist first

Frontends

Defined: Frontends provide some combination of individual metrics graphs, whole dashboards, and alerting configuration.

Support targets:

• Grafana
• Graphite
• Prometheus
• RRDtool
• Atlas -- Requires prying the Atlas frontend out of closed source at Netflix. They also have two more closed source pieces cleverly named:
• Alerts
• Netflix Dashboards