Co-authored-by: Dave Henderson <dhenderson@gmail.com> Co-authored-by: Jack Baldry <jack.baldry@grafana.com>
12 KiB
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
-
Ensure you have Docker installed and running on your machine.
-
Start Prometheus.
make devenv sources=prometheus
-
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. -
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 span’s 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
-
Start Jaeger
make devenv sources=jaeger
-
To enable tracing in Grafana, you must set the address in your
config.ini
file:[tracing.opentelemetry.jaeger] address = http://localhost:14268/api/traces
-
Search/browse collected logs and traces in Grafana Explore
You need provisioned
gdev-jaeger
andgdev-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 clickgdev-jaeger
to split the view and inspect the trace in question. -
Search or browse collected traces in Jaeger UI
You can open
http://localhost:16686
to use the Jaeger UI for browsing and searching traces.