prometheus_and_openmetrics.md

Prometheus and OpenMetrics Compatibility

Status: Experimental

Table of Contents

• Prometheus Metric points to OTLP
  • Metric Metadata
  • Counters
  • Gauges
  • Info
  • StateSet
  • Unknown-typed
  • Histograms
  • Summaries
  • Dropped Types
  • Start Time
  • Exemplars
  • Instrumentation Scope
  • Resource Attributes
• OTLP Metric points to Prometheus
  • Metric Metadata
  • Instrumentation Scope
  • Gauges
  • Sums
  • Histograms
  • Summaries
  • Dropped Data Points
  • Metric Attributes
  • Exemplars
  • Resource Attributes

This section denotes how to convert metrics scraped in the Prometheus exposition or OpenMetrics formats to the OpenTelemetry metric data model and how to create Prometheus metrics from OpenTelemetry metric data. Since OpenMetrics has a superset of Prometheus' types, "Prometheus" is taken to mean "Prometheus or OpenMetrics". "OpenMetrics" refers to OpenMetrics-only concepts.

Prometheus Metric points to OTLP

Metric Metadata

The OpenMetrics MetricFamily Name MUST be added as the Name of the OTLP metric after the removal of unit and type suffixes described below.

The OpenMetrics UNIT metadata, if present, MUST be converted to the unit of the OTLP metric. After trimming type-specific suffixes, such as _total for counters, the unit MUST be trimmed from the suffix as well, if the metric suffix matches the unit.

The OpenMetrics HELP metadata, if present, MUST be added as the description of the OTLP metric.

The OpenMetrics TYPE metadata, if present, MUST be used to determine the OTLP data type, and dictates type-specific conversion rules listed below. Metric families without type metadata follow rules for unknown-typed metrics below.

Counters

A Prometheus Counter MUST be converted to an OTLP Sum with is_monotonic equal to true. If the counter has a _total suffix, it MUST be removed.

Gauges

A Prometheus Gauge MUST be converted to an OTLP Gauge.

Info

An OpenMetrics Info metric MUST be converted to an OTLP Non-Monotonic Sum unless it is the target_info metric, which is used to populate resource attributes. An OpenMetrics Info can be thought of as a special-case of the OpenMetrics Gauge which has a value of 1, and whose labels generally stays constant over the life of the process. It is converted to a Non-Monotonic Sum, rather than a Gauge, because the value of 1 is intended to be viewed as a count, which should be summed together when aggregating away labels. If it has an _info suffix, the suffix MUST be removed from the metric name.

StateSet

An OpenMetrics StateSet metric MUST be converted to an OTLP Non-Monotonic Sum. An OpenMetrics StateSet can be thought of as a special-case of the OpenMetrics Gauge which has a 0 or 1 value, and has one metric point for every possible state. It is converted to a Non-Monotonic Sum, rather than a Gauge, because the value of 1 is intended to be viewed as a count, which should be summed together when aggregating away labels.

Unknown-typed

A Prometheus Unknown MUST be converted to an OTLP Gauge.

Histograms

A Prometheus Histogram MUST be converted to an OTLP Histogram.

Multiple Prometheus histogram metrics MUST be merged together into a single OTLP Histogram:

• The le label on non-suffixed metrics is used to identify and order histogram bucket boundaries. Each Prometheus line produces one bucket count on the resulting histogram. Each value for the le label except +Inf produces one bucket boundary.
• Lines with _count and _sum suffixes are used to determine the histogram's count and sum.
• If _count is not present, the metric MUST be dropped.
• If _sum is not present, the histogram's sum MUST be unset.

Summaries

Prometheus Summary MUST be converted to an OTLP Summary.

Multiple Prometheus metrics are merged together into a single OTLP Summary:

• The quantile label on non-suffixed metrics is used to identify quantile points in summary metrics. Each Prometheus line produces one quantile on the resulting summary.
• Lines with _count and _sum suffixes are used to determine the summary's count and sum.

Dropped Types

The following Prometheus types MUST be dropped:

• OpenMetrics GaugeHistogram

Start Time

Prometheus Cumulative metrics can include the start time using the _created metric as specified in OpenMetrics. When converting Prometheus Counters to OTLP, conversion SHOULD use _created where available. When no _created metric is available, conversion MUST follow Cumulative streams: handling unknown start time by default. Conversion MAY offer configuration, disabled by default, which allows using the process_start_time_seconds metric to provide the start time. Using process_start_time_seconds is only correct when all counters on the target start after the process and are not reset while the process is running.

Exemplars

OpenMetrics Exemplars can be attached to Prometheus Histogram bucket metric points and counter metric points. Exemplars on histogram buckets SHOULD be converted to exemplars on OpenTelemetry histograms. Exemplars on counter metric points SHOULD be converted to exemplars on OpenTelemetry sums. If present, the timestamp MUST be added to the OpenTelemetry exemplar. The Trace ID and Span ID SHOULD be retrieved from the trace_id and span_id label keys, respectively. All labels not used for the trace and span ids MUST be added to the OpenTelemetry exemplar as attributes.

Instrumentation Scope

Each otel_scope_info metric point present in a batch of metrics SHOULD be dropped from the incoming scrape, and converted to an instrumentation scope. The otel_scope_name and otel_scope_version labels, if present, MUST be converted to the Name and Version of the Instrumentation Scope. Additional labels MUST be added as scope attributes, with keys and values unaltered. Other metrics in the batch which have otel_scope_name and otel_scope_version labels that match an instrumentation scope MUST be placed within the matching instrumentation scope, and MUST remove those labels. For example, the OpenMetrics metrics:

# TYPE otel_scope_info info
otel_scope_info{otel_scope_name="go.opentelemetry.io.contrib.instrumentation.net.http.otelhttp",otel_scope_version="v0.24.0",library_mascot="bear"} 1
# TYPE http_server_duration counter
http_server_duration{otel_scope_name="go.opentelemetry.io.contrib.instrumentation.net.http.otelhttp",otel_scope_version="v0.24.0"...} 1

becomes:

# within a resource_metrics
scope_metrics:
  scope:
    name: go.opentelemetry.io.contrib.instrumentation.net.http.otelhttp
    version: v0.24.0
    attributes:
      library_mascot: bear
  metrics:
  - name: http_server_duration
    data:
      sum:
        data_points:
        - value: 1

Metrics which are not found to be associated with an instrumentation scope MUST all be placed within an empty instrumentation scope, and MUST not have any labels removed.

Resource Attributes

When scraping a Prometheus endpoint, resource attributes MUST be added to the scraped metrics to distinguish them from metrics from other Prometheus endpoints. In particular, service.name and service.instance.id, are needed to ensure Prometheus exporters can disambiguate metrics using job and instance labels as described below.

The following attributes MUST be associated with scraped metrics as resource attributes, and MUST NOT be added as metric attributes:

OTLP Resource Attribute	Description
service.name	The configured name of the service that the target belongs to
service.instance.id	A unique identifier of the target. By default, it should be the <host>:<port> of the scraped URL

The following attributes SHOULD be associated with scraped metrics as resource attributes, and MUST NOT be added as metric attributes:

OTLP Resource Attribute	Description
net.host.name	The <host> portion of the target's URL that was scraped
net.host.port	The <port> portion of the target's URL that was scraped
http.scheme	http or https

In addition to the attributes above, the target_info metric is used to supply additional resource attributes. If present, target_info MUST be dropped from the batch of metrics, and all labels from the target_info metric MUST be converted to resource attributes attached to all other metrics which are part of the scrape. By default, label keys and values MUST NOT be altered (such as replacing _ with . characters in keys).

OTLP Metric points to Prometheus

Metric Metadata

Prometheus SDK exporters MUST NOT allow duplicate UNIT, HELP, or TYPE comments for the same metric name to be returned in a single scrape of the Prometheus endpoint. Exporters MUST drop entire metrics to prevent conflicting TYPE comments, but SHOULD NOT drop metric points as a result of conflicting UNIT or HELP comments. Instead, all but one of the conflicting UNIT and HELP comments (but not metric points) SHOULD be dropped. If dropping a comment or metric points, the exporter SHOULD warn the user through error logging.

The Name of an OTLP metric MUST be added as the OpenMetrics MetricFamily Name, with unit and type suffixes added as described below. The metric name is required to match the regex: [a-zA-Z_:]([a-zA-Z0-9_:])*. Invalid characters in the metric name MUST be replaced with the _ character. Multiple consecutive _ characters MUST be replaced with a single _ character.

The Unit of an OTLP metric point MUST be added as OpenMetrics UNIT metadata. Additionally, the unit MUST be added as a suffix to the metric name, and SHOULD be converted to base units recommended by OpenMetrics when possible. The unit suffix comes before any type-specific suffixes.

The description of an OTLP metrics point MUST be added as OpenMetrics HELP metadata.

The data point type of an OTLP metric MUST be added as OpenMetrics TYPE metadata. It also dictates type-specific conversion rules listed below.

Instrumentation Scope

Prometheus exporters SHOULD generate an Info-typed metric named otel_scope_info. If present, Instrumentation Scope name and version MUST be added as otel_scope_name and otel_scope_version labels. Scope attributes MUST also be added as labels following the rules described in the Metric Attributes section below.

Prometheus exporters MUST add the scope name as the otel_scope_name label and the scope version as the otel_scope_version label on all metric points by default.

Prometheus exporters SHOULD provide a configuration option to disable the otel_scope_info metric and otel_scope_ labels.

Gauges

An OpenTelemetry Gauge MUST be converted to a Prometheus Gauge.

Sums

OpenTelemetry Sums follows this logic:

• If the aggregation temporality is cumulative and the sum is monotonic, it MUST be converted to a Prometheus Counter.
• If the aggregation temporality is cumulative and the sum is non-monotonic, it MUST be converted to a Prometheus Gauge.
• If the aggregation temporality is delta and the sum is monotonic, it SHOULD be converted to a cumulative temporality and become a Prometheus Sum. The following behaviors are expected:
  • The new data point type must be the same as the accumulated data point type.
  • The new data point's start time must match the time of the accumulated data point. If not, see detecting alignment issues.
• Otherwise, it MUST be dropped.

Monotonic Sum metric points MUST have _total added as a suffix to the metric name. Monotonic Sum metric points with StartTimeUnixNano should export the {name}_created metric as well.

Histograms

An OpenTelemetry Histogram with a cumulative aggregation temporality MUST be converted to a Prometheus metric family with the following metrics:

• A single {name}_count metric denoting the count field of the histogram. All attributes of the histogram point are converted to Prometheus labels.
• {name}_sum metric denoting the sum field of the histogram, reported only if the sum is positive and monotonic. The sum is positive and monotonic when all buckets are positive. All attributes of the histogram point are converted to Prometheus labels.
• A series of {name} metric points that contain all attributes of the histogram point recorded as labels. Additionally, a label, denoted as le is added denoting the bucket boundary. The label's value is the stringified floating point value of bucket boundaries, ordered from lowest to highest. The value of each point is the sum of the count of all histogram buckets up the the boundary reported in the le label. These points will include a single exemplar that falls within le label and no other le labelled point. The final bucket metric MUST have an +Inf threshold.
• Histograms with StartTimeUnixNano set should export the {name}_created metric as well.

OpenTelemetry Histograms with Delta aggregation temporality SHOULD be aggregated into a Cumulative aggregation temporality and follow the logic above, or MUST be dropped.

Summaries

An OpenTelemetry Summary MUST be converted to a Prometheus metric family with the following metrics:

• A single {name}_count metric denoting the count field of the summary. All attributes of the summary point are converted to Prometheus labels.
• {name}_sum metric denoting the sum field of the summary, reported only if the sum is positive and monotonic. All attributes of the summary point are converted to Prometheus labels.
• A series of {name} metric points that contain all attributes of the summary point recorded as labels. Additionally, a label, denoted as quantile is added denoting a reported quantile point, and having its value be the stringified floating point value of quantiles (between 0.0 and 1.0), starting from lowest to highest, and all being non-negative. The value of each point is the computed value of the quantile point.
• Summaries with StartTimeUnixNano set should export the {name}_created metric as well.

Dropped Data Points

The following OTLP data points MUST be dropped:

• ExponentialHistogram

Metric Attributes

OpenTelemetry Metric Attributes MUST be converted to Prometheus labels. String Attribute values are converted directly to Metric Attributes, and non-string Attribute values MUST be converted to string attributes following the attribute specification. Prometheus metric label keys are required to match the following regex: [a-zA-Z_]([a-zA-Z0-9_])*. Metrics from OpenTelemetry with unsupported Attribute names MUST replace invalid characters with the _ character. Multiple consecutive _ characters MUST be replaced with a single _ character. This may cause ambiguity in scenarios where multiple similar-named attributes share invalid characters at the same location. In such unlikely cases, if multiple key-value pairs are converted to have the same Prometheus key, the values MUST be concatenated together, separated by ;, and ordered by the lexicographical order of the original keys.

Exemplars

Exemplars on OpenTelemetry Histograms and Monotonic Sums SHOULD be converted to OpenMetrics exemplars. Exemplars on other OpenTelemetry data points MUST be dropped. For Prometheus push exporters, multiple exemplars are able to be added to each bucket, so all exemplars SHOULD be converted. For Prometheus pull endpoints, only a single exemplar is able to be added to each bucket, so the largest exemplar from each bucket MUST be used, if attaching exemplars. If no exemplars exist on a bucket, the highest exemplar from a lower bucket MUST be used, even though it is a duplicate of another bucket's exemplar. OpenMetrics Exemplars MUST use the trace_id and span_id keys for the trace and span IDs, respectively. Timestamps MUST be added as timestamps on the OpenMetrics exemplar, and filtered_attributes MUST be added as labels on the OpenMetrics exemplar unless they would exceed the OpenMetrics limit on characters.

Resource Attributes

In SDK Prometheus (pull) exporters, resource attributes SHOULD be converted to a single target_info metric; otherwise, they MUST be dropped, and MUST NOT be attached as labels to other metric families. The target_info metric MUST be an info-typed metric whose labels MUST include the resource attributes, and MUST NOT include any other labels. There MUST be at most one target_info metric exposed on an SDK Prometheus endpoint.

In the Collector's Prometheus pull and push (remote-write) exporters, it is possible for metrics from multiple targets to be sent together, so targets must be disambiguated from one another. However, the Prometheus exposition format and remote-write formats do not include a notion of resource, and expect metric labels to distinguish scraped targets. By convention, job and instance labels distinguish targets and are expected to be present on metrics exposed on a Prometheus pull exporter (a "federated" Prometheus endpoint) or pushed via Prometheus remote-write. In OTLP, the service.name, service.namespace, and service.instance.id triplet is required to be unique, which makes them good candidates to use to construct job and instance. In the collector Prometheus exporters, the service.name and service.namespace attributes MUST be combined as <service.namespace>/<service.name>, or <service.name> if namespace is empty, to form the job metric label. The service.instance.id attribute, if present, MUST be converted to the instance label; otherwise, instance should be added with an empty value. Other resource attributes SHOULD be converted to a target_info metric, or MUST be dropped. The target_info metric is an info-typed metric whose labels MUST include the resource attributes, and MUST NOT include any other labels other than job and instance. There MUST be at most one target_info metric exported for each unique combination of job and instance.

If info-typed metric families are not yet supported by the language Prometheus client library, a gauge-typed metric family named target_info with a constant value of 1 MUST be used instead.

To convert OTLP resource attributes to Prometheus labels, string Attribute values are converted directly to labels, and non-string Attribute values MUST be converted to string attributes following the attribute specification.