grafana/contribute/backend/instrumentation.md

Instrumenting Grafana

Guidance, conventions and best practices for instrumenting Grafana using logs, metrics and traces.

Logs

Logs are files that record events, warnings and errors as they occur within a software environment. Most logs include contextual information, such as the time an event occurred and which user or endpoint was associated with it.

Usage

Use the pkg/infra/log package to create a named structured logger. Example:

import (
  "fmt"

  "github.com/grafana/grafana/pkg/infra/log"
)

logger := log.New("my-logger")
logger.Debug("Debug msg")
logger.Info("Info msg")
logger.Warning("Warning msg")
logger.Error("Error msg", "error", fmt.Errorf("BOOM"))

Naming conventions

Name the logger using lowercase characters, e.g. log.New("my-logger") using snake_case or kebab-case styling.

Prefix the logger name with an area name when using different loggers across a feature or related packages, e.g. log.New("plugin.loader") and log.New("plugin.client").

Start the log message with a capital letter, e.g. logger.Info("Hello world") instead of logger.Info("hello world"). The log message should be an identifier for the log entry, avoid parameterization in favor of key-value pairs for additional data.

Prefer using camelCase style when naming log keys, e.g. remoteAddr, to be consistent with Go identifiers.

Use the key error when logging Go errors, e.g. logger.Error("Something failed", "error", fmt.Errorf("BOOM")).

Validate and sanitize input coming from user input

If log messages or key/value pairs originates from user input they should be validated and sanitized.

Be careful to not expose any sensitive information in log messages e.g. secrets, credentials etc. It's especially easy to do by mistake when including a struct as value.

Log levels

When to use which log level?

• Debug: Informational messages of high frequency and/or less-important messages during normal operations.
• Info: Informational messages of low frequency and/or important messages.
• Warning: Should in normal cases not be used/needed. If used should be actionable.
• Error: Error messages indicating some operation failed (with an error) and the program didn't have a way of handle the error.

Contextual logging

Use a contextual logger to include additional key/value pairs attached to context.Context, e.g. traceID, to allow correlating logs with traces and/or correlate logs with a common identifier.

You must Enable tracing in Grafana to get a traceID

Example:

import (
  "context"
  "fmt"

  "github.com/grafana/grafana/pkg/infra/log"
)

var logger = log.New("my-logger")

func doSomething(ctx context.Context) {
  ctxLogger := logger.FromContext(ctx)
  ctxLogger.Debug("Debug msg")
  ctxLogger.Info("Info msg")
  ctxLogger.Warning("Warning msg")
  ctxLogger.Error("Error msg", "error", fmt.Errorf("BOOM"))
}

Enable certain log levels for certain loggers

During development, it's convenient to enable certain log level, e.g. debug, for certain loggers to minimize the generated log output and make it easier to find things. See [log.filters] for information how to configure this.

It's also possible to configure multiple loggers:

[log]
filters = rendering:debug \
          ; alerting.notifier:debug \
          oauth.generic_oauth:debug \
          ; oauth.okta:debug \
          ; tsdb.postgres:debug \
          ; tsdb.mssql:debug \
          ; provisioning.plugins:debug \
          ; provisioning:debug \
          ; provisioning.dashboard:debug \
          ; provisioning.datasources:debug \
          datasources:debug \
          data-proxy-log:debug

Metrics

Metrics are quantifiable measurements that reflect the health and performance of applications or infrastructure.

Consider using metrics to provide real-time insight into the state of resources. If you want to know how responsive your application is or identify anomalies that could be early signs of a performance issue, metrics are a key source of visibility.

Metric types

See Prometheus metric types for a list and description of the different metric types you can use and when to use them.

There are many possible types of metrics that can be tracked. One popular method for defining metrics is the RED method.

Naming conventions

Use the namespace grafana as that would prefix any defined metric names with grafana_. This will make it clear for operators that any metric named grafana_* belongs to Grafana.

Use snakecase style when naming metrics, e.g. _http_request_duration_seconds instead of httpRequestDurationSeconds.

Use snakecase style when naming labels, e.g. _status_code instead of statusCode.

If metric type is a counter, name it with a _total suffix, e.g. http_requests_total.

If metric type is a histogram and you're measuring duration, name it with a _<unit> suffix, e.g. http_request_duration_seconds.

If metric type is a gauge, name it to denote it's a value that can increase and decrease , e.g. http_request_in_flight.

Label values and high cardinality

Be careful with what label values you add/accept. Using/allowing too many label values could result in high cardinality problems.

If label values originates from user input they should be validated. Use metricutil.SanitizeLabelName(<label value>) from pkg/infra/metrics/metricutil package to sanitize label names. Very important to only allow a pre-defined set of labels to minimize the risk of high cardinality problems.

Be careful to not expose any sensitive information in label values, e.g. secrets, credentials etc.

Guarantee the existence of metrics

If you want to guarantee the existence of metrics before any observations has happened there's a couple of helper methods available in the pkg/infra/metrics/metricutil package.

How to collect and visualize metrics locally

1. Ensure you have Docker installed and running on your machine

2. Start Prometheus

   make devenv sources=prometheus
   

3. Run Grafana, and create a Prometheus datasource if you do not have one yet. Set the server URL to http://localhost:9090, enable basic auth, and type in the same auth you have for local Grafana

4. Use Grafana Explore or dashboards to query any exported Grafana metrics. You can also view them at http://localhost:3000/metrics

Traces

A distributed trace is data that tracks an application request as it flows through the various parts of an application. The trace records how long it takes each application component to process the request and pass the result to the next component. Traces can also identify which parts of the application trigger an error.

Usage

Grafana uses OpenTelemetry for distributed tracing. There's an interface Tracer in the pkg/infra/tracing package that implements the OpenTelemetry Tracer interface, which you can use to create traces and spans. To get a hold of a Tracer you would need to get it injected as dependency into your service, see Services for more details. For more information, see https://opentelemetry.io/docs/instrumentation/go/manual/.

Example:

import (
   "fmt"

   "github.com/grafana/grafana/pkg/infra/tracing"
   "go.opentelemetry.io/otel/attribute"
   "go.opentelemetry.io/otel/trace"
)

type MyService struct {
   tracer tracing.Tracer
}

func ProvideService(tracer tracing.Tracer) *MyService {
   return &MyService{
      tracer: tracer,
   }
}

func (s *MyService) Hello(ctx context.Context, name string) (string, error) {
   ctx, span := s.tracer.Start(ctx, "MyService.Hello", trace.WithAttributes(
      attribute.String("my_attribute", "val"),
   ))
   // this make sure the span is marked as finished when this
   // method ends to allow the span to be flushed and sent to
   // storage backend.
   defer span.End()

   // Add some event to show Events usage
   span.AddEvent("checking name...")

   if name == "" {
      err := fmt.Errorf("name cannot be empty")

      // Use the helper functions tracing.Errorf or tracing.Error
      // to set the span’s status to Error to make
      // the span tracking a failed operation as an error span and
      // record error as an exception span event for the provided span.
      return "", tracing.Errorf(span, "failed to check name: %w", err)
   }

   // Add some other event to show Events usage
   span.AddEvent("name checked")

   // Add attribute to show Attributes usage
   span.SetAttributes(
      attribute.String("my_service.name", name),
      attribute.Int64("my_service.some_other", int64(1337)),
   )

   return fmt.Sprintf("Hello %s", name), nil
}

Naming conventions

Span names should follow the guidelines from OpenTelemetry.

Span Name	Guidance
get	Too general
get_account/42	Too specific
get_account	Good, and account_id=42 would make a nice Span attribute
get_account/{accountId}	Also good (using the “HTTP route”)

Span attribute and span event attributes should follow the Attribute naming specification from OpenTelemetry. Good attribute key examples:

• service.version
• http.status_code

See Trace semantic conventions from OpenTelemetry for additional conventions regarding well-known protocols and operations.

Span names and high cardinality

Be careful with what span names you add/accept. Using/allowing too many span names could result in high cardinality problems.

Validate and sanitize input coming from user input

If span names, attribute or event values originates from user input they should be validated and sanitized. It's very important to only allow a pre-defined set of span names to minimize the risk of high cardinality problems.

Be careful to not expose any sensitive information in span names, attribute or event values, e.g. secrets, credentials etc.

Span attributes

Consider using attributes.<Type>("<key>", <value>) in favor of attributes.Key("<key>").<Type>(<value>) since it requires less characters and thereby reads easier.

Example:

attribute.String("datasource_name", proxy.ds.Name)
// vs
attribute.Key("datasource_name").String(proxy.ds.Name)

attribute.Int64("org_id", proxy.ctx.SignedInUser.OrgID)
// vs
attribute.Key("org_id").Int64(proxy.ctx.SignedInUser.OrgID)

How to collect, visualize and query traces (and correlate logs with traces) locally

1. Start Jaeger

make devenv sources=jaeger

2. Enable tracing in Grafana

To enable tracing in Grafana, you must set the address in your config.ini file

[tracing.opentelemetry.jaeger]
address = http://localhost:14268/api/traces

3. Search/browse collected logs and traces in Grafana Explore

You need provisioned gdev-jaeger and gdev-loki datasources, see developer dashboard and data sources for setup instructions.

Open Grafana explore and select gdev-loki datasource and use the query {filename="/var/log/grafana/grafana.log"} | logfmt.

You can then inspect any log message that includes a traceID and from there click on gdev-jaeger to split view and inspect the trace in question.

4. Search/browse collected traces in Jaeger UI

You can open http://localhost:16686 to use the Jaeger UI for browsing and searching traces.