grafana/contribute/backend/instrumentation.md
Raiko Wielk 3d9ae801cb
Docs: Fix broken links in contribute/**/*.md (#92182)
Co-authored-by: Dave Henderson <dhenderson@gmail.com>
Co-authored-by: Jack Baldry <jack.baldry@grafana.com>
2024-09-12 15:51:30 +03:00

12 KiB
Raw Blame History

Instrumenting Grafana

This guide provides 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, for example, 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; for example, log.New("plugin.loader") and log.New("plugin.client").

Start the log message with a capital letter, for example, 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.

To be consistent with Go identifiers, prefer using camelCase style when naming log keys; for example, remoteAddr.

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

Validate and sanitize input coming from user input

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

Be careful not to expose any sensitive information in log messages; for example, secrets and credentials. It's easy to do this by mistake if you include a struct as a value.

Log levels

When should you use each log level?

  • Debug: Informational messages of high frequency, less-important messages during normal operations, or both.
  • Info: Informational messages of low frequency, important messages, or both.
  • Warning: Use warning messages sparingly. If used, messages should be actionable.
  • Error: Error messages indicating some operation failed (with an error) and the program didn't have a way to handle the error.

Contextual logging

Use a contextual logger to include additional key/value pairs attached to context.Context. For example, a traceID, used to allow correlating logs with traces, correlate logs with a common identifier, either or both.

You must Enable tracing in Grafana to get a traceID.

For 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

You can enable certain log levels during development to make logging easier. For example, you can enable debug to allow certain loggers to minimize the generated log output and makes it easier to find things. Refer to [log.filters] for information on how to to set different levels for specific loggers.

You can also configure multiple loggers. For example:

[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 to prefix any defined metric names with grafana_. This prefix makes it clear for operators that any metric named grafana_* belongs to Grafana.

Use snake_case style when naming metrics; for example, http_request_duration_seconds instead of httpRequestDurationSeconds.

Use snake_case style when naming labels; for example, status_code instead of statusCode.

If a metric type is a counter, name it with a _total suffix; for example, http_requests_total.

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

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

Label values and high cardinality

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

If label values originate from user input they should be validated. Use metricutil.SanitizeLabelName(<label value>) from the pkg/infra/metrics/metricutil package to sanitize label names.

Important: Only allow a pre-defined set of labels to minimize the risk of high cardinality problems. Be careful not to expose any sensitive information in label values such as secrets and credentials.

Guarantee the existence of metrics

To guarantee the existence of metrics before any observations have happened, you can use the 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 then create a Prometheus data source if you do not have one yet. Set the server URL to http://localhost:9090, enable basic authentication, and enter the same authentication 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 access Tracer you need to get it injected as a dependency of your service. Refer to Services for more details. For more information, you may also refer to The OpenTelemetry documentation.

For 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"),
   ))
   // 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 spans status to Error to make
      // the span tracking a failed operation as an error span and
      // record the 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.

These are a few examples of good attributes:

  • service.version
  • http.status_code

Refer to 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 or accept. Using or allowing too many span names can result in high cardinality problems.

Validate and sanitize input coming from user input

If span names, attribute values, or event values originate 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; for example, secrets, credentials, and so on.

Span attributes

Consider using attributes.<Type>("<key>", <value>) instead of attributes.Key("<key>").<Type>(<value>) since it requires fewer characters and is easier to read.

For 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 data sources. Refer to developer dashboard and data sources for set up instructions.

    Open Grafana explore and select the gdev-loki data source 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 gdev-jaeger to split the view and inspect the trace in question.

  4. Search or browse collected traces in Jaeger UI

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