obsreport

Documentation

Overview

Package obsreport provides unified and consistent observability signals ( metrics, tracing, etc) for components of the OpenTelemetry collector.

The function Configure is used to control which signals are going to be generated. It provides functions for the typical operations of receivers, processors, and exporters.

Receivers should use the respective start and end according to the data type being received, ie.:

• Traces receive operations should use the pair: StartTracesOp/EndTracesOp

• Metrics receive operations should use the pair: StartMetricsOp/EndMetricsOp

• Logs receive operations should use the pair: StartLogsOp/EndLogsOp

Similar for exporters:

• Traces export operations should use the pair: StartTracesOp/EndTracesOp

• Metrics export operations should use the pair: StartMetricsOp/EndMetricsOp

• Metrics export operations should use the pair: StartLogsOp/EndLogsOp

The package is capable of generating legacy metrics by using the observability package allowing a controlled transition from legacy to the new metrics. The goal is to eventually remove the legacy metrics and use only the new metrics.

The main differences regarding the legacy metrics are:

1. "Amount of metric data" is measured as metric points (ie.: a single value in time), contrast it with number of time series used legacy. Number of metric data points is a more general concept regarding various metric formats.

2. Exporters measure the number of items, ie.: number of spans or metric points, that were sent and the ones for which the attempt to send failed. For more information about this see Notes below about reporting data loss.

3. All measurements of "amount of data" used in the new metrics for receivers and exporters should reflect their native formats, not the internal format used in the Collector. This is to facilitate reconciliation between Collector, client and backend. For instance: certain metric formats do not provide direct support for histograms and have predefined conventions to represent those, this conversion may end with a different number of time series and data points than the internal Collector format.

Notes:

* Data loss should be recorded only when the component itself remove the data from the pipeline. Legacy metrics for receivers used "dropped" in their names but these could be non-zero under normal operations and reflected no actual data loss when exporters with "sending_queue" are used. New metrics were renamed to avoid this misunderstanding. Here are the general recommendations to report data loss:

• Receivers reporting errors to clients typically result in the client re-sending the same data so it is more correct to report "receive errors", not actual data loss.

• Exporters need to report individual failures to send data, but on typical production pipelines processors usually take care of retries, so these should be reported as "send errors".

• Data "filtered out" should have its own metrics and not be confused with dropped data.

Naming Convention for New Metrics

Common Metrics: Metrics shared by different components should follow the convention below:

`<component>/<metric_name>`

As a label the metric should have at least `{<component>="<name>"}` where `<name>` is the name used in the configuration for the instance of the component, eg.:

`receiver/accepted_spans{receiver="otlp",...}` `exporter/sent_spans{exporter="otlp/prod",...}`

Component Specific Metrics: These metrics are implemented by specific components, eg.: batch processor. The general pattern is the same as the common metrics but with the addition of the component type (as it appears in the configuration) before the actual metric:

`<component>/<type>/<metric_name>`

Even metrics exclusive to a single type should follow the conventions above and also include the type (as written in the configuration) as part of the metric name since there could be multiple instances of the same type in different pipelines, eg.:

`processor/batch/batch_size_trigger_send{processor="batch/dev",...}`