README.md

Service Performance Monitoring (SPM) Development/Demo Environment

Service Performance Monitoring (SPM) is an opt-in feature introduced to Jaeger that provides Request, Error and Duration (RED) metrics grouped by service name and operation that are derived from span data. These metrics are programmatically available through an API exposed by jaeger-query along with a "Monitor" UI tab that visualizes these metrics as graphs.

For more details on this feature, please refer to the tracking Issue documenting the proposal and status.

The motivation for providing this environment is to allow developers to either test Jaeger UI or their own applications against jaeger-query's metrics query API, as well as a quick and simple way for users to bring up the entire stack required to visualize RED metrics from simulated traces or from their own application.

This environment consists the following backend components:

• MicroSim: a program to simulate traces.
• Jaeger All-in-one: the full Jaeger stack in a single container image.
• OpenTelemetry Collector: vendor agnostic integration layer for traces and metrics. Its main role in this particular development environment is to receive Jaeger spans, forward these spans untouched to Jaeger All-in-one while simultaneously aggregating metrics out of this span data. To learn more about span metrics aggregation, please refer to the spanmetrics connector documentation.
• Prometheus: a metrics collection and query engine, used to scrape metrics computed by OpenTelemetry Collector, and presents an API for Jaeger All-in-one to query these metrics.

The following diagram illustrates the relationship between these components:

flowchart LR
    SDK -->|traces| Receiver
    Receiver --> MG
    Receiver --> Batch
    MG --> ExpMetrics
    Batch --> ExpTraces
    ExpMetrics -->|metrics| Prometheus[(Prometheus)]
    ExpTraces -->|traces| Jaeger[Jaeger
     Collector]
    Prometheus -.-> JaegerUI
    Jaeger --> Storage[(Storage)]
    Storage -.-> JaegerUI[Jaeger
     Query
     and UI]

    style Prometheus fill:red,color:white
    style Jaeger fill:blue,color:white
    style JaegerUI fill:blue,color:white
    style Storage fill:gray,color:white

    subgraph Application
        SDK[OTel
         SDK]
    end

    subgraph OTEL[OTel Collector]
        Receiver
        Batch
        MG[Span
         Metrics
         Connector]
        ExpTraces[Traces
         Exporter]
        ExpMetrics[Metrics
         Exporter]
    end

Loading

Getting Started

Quickstart

This brings up the system necessary to use the SPM feature locally. It uses the latest image tags from both Jaeger and OpenTelemetry.

docker compose up

Jaeger v2

docker compose -f docker-compose-v2.yml up

Tips:

• Let the application run for a couple of minutes to ensure there is enough time series data to plot in the dashboard.
• Navigate to Jaeger UI at http://localhost:16686/ and inspect the Monitor tab. Select redis service from the dropdown to see more than one endpoint.
• To visualize the raw metrics stored on the Prometheus server (for debugging and local development use cases), use the built-in Prometheus UI at http://localhost:9090/query. For example, http://localhost:9090/query?g0.expr=traces_span_metrics_calls_total&g0.tab=0&g0.range_input=5m

Warning: The included docker-compose.yml file uses the latest version of Jaeger and other components. If your local Docker registry already contains older versions, which may still be tagged as latest, you may want to delete those images before running the full set, to ensure consistent behavior:

make clean-all

To use an official published image of Jaeger, specify the version via environment variable:

JAEGER_IMAGE_TAG=1.62.0 docker compose up

or for Jaeger v2:

JAEGER_IMAGE_TAG=2.0.0 docker compose -f docker-compose-v2.yml up

Development

These steps allow for running the system necessary for SPM, built from Jaeger's source.

The primary use case is for testing source code changes to the SPM feature locally.

Build jaeger-all-in-one docker image

make build

Bring up the dev environment

make dev

Sending traces

We will use tracegen to emit traces to the OpenTelemetry Collector which, in turn, will aggregate the trace data into metrics.

Start the local stack needed for SPM, if not already done:

docker compose up

Generate a specific number of traces with:

docker run --env OTEL_EXPORTER_OTLP_TRACES_ENDPOINT="http://otel_collector:4317" \
  --network monitor_backend \
  --rm \
  jaegertracing/jaeger-tracegen:1.49 \
    -trace-exporter otlp-grpc \
    -traces 1

Or, emit traces over a period of time with:

docker run --env OTEL_EXPORTER_OTLP_TRACES_ENDPOINT="http://otel_collector:4317" \
  --network monitor_backend \
  --rm \
  jaegertracing/jaeger-tracegen:1.49 \
    -trace-exporter otlp-grpc \
    -duration 5s

Navigate to Jaeger UI at http://localhost:16686/ and you should be able to see traces from this demo application under the tracegen service:

Then navigate to the Monitor tab at http://localhost:16686/monitor to view the RED metrics:

Querying the HTTP API

Example 1

Fetch call rates for both the driver and frontend services, grouped by operation, from now, looking back 1 second with a sliding rate-calculation window of 1m and step size of 1 millisecond

curl "http://localhost:16686/api/metrics/calls?service=driver&service=frontend&groupByOperation=true&endTs=$(date +%s)000&lookback=1000&step=100&ratePer=60000" | jq .

Example 2

Fetch P95 latencies for both the driver and frontend services from now, looking back 1 second with a sliding rate-calculation window of 1m and step size of 1 millisecond, where the span kind is either "server" or "client".

curl "http://localhost:16686/api/metrics/latencies?service=driver&service=frontend&quantile=0.95&endTs=$(date +%s)000&lookback=1000&step=100&ratePer=60000&spanKind=server&spanKind=client" | jq .

Example 3

Fetch error rates for both driver and frontend services using default parameters.

curl "http://localhost:16686/api/metrics/errors?service=driver&service=frontend" | jq .

Example 4

Fetch the minimum step size supported by the underlying metrics store.

curl "http://localhost:16686/api/metrics/minstep" | jq .

HTTP API Specification

Query Metrics

/api/metrics/{metric_type}?{query}

Where (Backus-Naur form):

metric_type = 'latencies' | 'calls' | 'errors'

query = services , [ '&' optionalParams ]

optionalParams = param | param '&' optionalParams

param =  groupByOperation | quantile | endTs | lookback | step | ratePer | spanKinds

services = service | service '&' services
service = 'service=' strValue
  - The list of services to include in the metrics selection filter, which are logically 'OR'ed.
  - Mandatory.

quantile = 'quantile=' floatValue
  - The quantile to compute the latency 'P' value. Valid range (0,1].
  - Mandatory for 'latencies' type.

groupByOperation = 'groupByOperation=' boolValue
boolValue = '1' | 't' | 'T' | 'true' | 'TRUE' | 'True' | 0 | 'f' | 'F' | 'false' | 'FALSE' | 'False'
  - A boolean value which will determine if the metrics query will also group by operation.
  - Optional with default: false

endTs = 'endTs=' intValue
  - The posix milliseconds timestamp of the end time range of the metrics query.
  - Optional with default: now

lookback = 'lookback=' intValue
  - The duration, in milliseconds, from endTs to look back on for metrics data points.
  - For example, if set to `3600000` (1 hour), the query would span from `endTs - 1 hour` to `endTs`.
  - Optional with default: 3600000 (1 hour).

step = 'step=' intValue
  - The duration, in milliseconds, between data points of the query results.
  - For example, if set to 5s, the results would produce a data point every 5 seconds from the `endTs - lookback` to `endTs`.
  - Optional with default: 5000 (5 seconds).

ratePer = 'ratePer=' intValue
  - The duration, in milliseconds, in which the per-second rate of change is calculated for a cumulative counter metric.
  - Optional with default: 600000 (10 minutes).

spanKinds = spanKind | spanKind '&' spanKinds
spanKind = 'spanKind=' spanKindType
spanKindType = 'unspecified' | 'internal' | 'server' | 'client' | 'producer' | 'consumer'
  - The list of spanKinds to include in the metrics selection filter, which are logically 'OR'ed.
  - Optional with default: 'server'

Min Step

/api/metrics/minstep

Gets the min time resolution supported by the backing metrics store, in milliseconds, that can be used in the step parameter. e.g. a min step of 1 means the backend can only return data points that are at least 1ms apart, not closer.

Responses

The response data model is based on MetricsFamily.

For example:

{
  "name": "service_call_rate",
  "type": "GAUGE",
  "help": "calls/sec, grouped by service",
  "metrics": [
    {
      "labels": [
        {
          "name": "service_name",
          "value": "driver"
        }
      ],
      "metricPoints": [
        {
          "gaugeValue": {
            "doubleValue": 0.005846808321083344
          },
          "timestamp": "2021-06-03T09:12:06Z"
        },
        {
          "gaugeValue": {
            "doubleValue": 0.006960443672323934
          },
          "timestamp": "2021-06-03T09:12:11Z"
        },
      ]
...
    }
...
  ]
...
}

If the groupByOperation=true parameter is set, the response will include the operation name in the labels like so:

      "labels": [
        {
          "name": "operation",
          "value": "/FindNearest"
        },
        {
          "name": "service_name",
          "value": "driver"
        }
      ],

Disabling Metrics Querying

As this is feature is opt-in only, disabling metrics querying simply involves omitting the METRICS_STORAGE_TYPE environment variable when starting-up jaeger-query or jaeger all-in-one.

For example, try removing the METRICS_STORAGE_TYPE=prometheus environment variable from the docker-compose.yml file.

Then querying any metrics endpoints results in an error message:

$ curl http://localhost:16686/api/metrics/minstep | jq .
{
  "data": null,
  "total": 0,
  "limit": 0,
  "offset": 0,
  "errors": [
    {
      "code": 405,
      "msg": "metrics querying is currently disabled"
    }
  ]
}