指标 SDK¶

Status: Mixed

Table of Contents

- [MeterProvider](#meterprovider) - [MeterProvider Creation](#meterprovider-creation) - [Meter Creation](#meter-creation) - [Shutdown](#shutdown) - [ForceFlush](#forceflush) - [View](#view) - [Aggregation](#aggregation) - [Drop Aggregation](#drop-aggregation) - [Default Aggregation](#default-aggregation) - [Sum Aggregation](#sum-aggregation) - [Last Value Aggregation](#last-value-aggregation) - [Histogram Aggregations](#histogram-aggregations) - [Explicit Bucket Histogram Aggregation](#explicit-bucket-histogram-aggregation) - [Base2 Exponential Bucket Histogram Aggregation](#base2-exponential-bucket-histogram-aggregation) - [Handle all normal values](#handle-all-normal-values) - [Support a minimum and maximum scale](#support-a-minimum-and-maximum-scale) - [Use the maximum scale for single measurements](#use-the-maximum-scale-for-single-measurements) - [Maintain the ideal scale](#maintain-the-ideal-scale) - [Observations inside asynchronous callbacks](#observations-inside-asynchronous-callbacks) - [Cardinality limits](#cardinality-limits) - [Synchronous instrument cardinality limits](#synchronous-instrument-cardinality-limits) - [Asynchronous instrument cardinality limits](#asynchronous-instrument-cardinality-limits) - [Meter](#meter) - [Duplicate instrument registration](#duplicate-instrument-registration) - [Instrument name](#instrument-name) - [Instrument unit](#instrument-unit) - [Instrument description](#instrument-description) - [Instrument advice](#instrument-advice) - [Attribute limits](#attribute-limits) - [Exemplar](#exemplar) - [ExemplarFilter](#exemplarfilter) - [Built-in ExemplarFilters](#built-in-exemplarfilters) - [AlwaysOn](#alwayson) - [AlwaysOff](#alwaysoff) - [TraceBased](#tracebased) - [ExemplarReservoir](#exemplarreservoir) - [Exemplar defaults](#exemplar-defaults) - [MetricReader](#metricreader) - [MetricReader operations](#metricreader-operations) - [RegisterProducer(metricProducer)](#registerproducermetricproducer) - [Collect](#collect) - [Shutdown](#shutdown-1) - [Periodic exporting MetricReader](#periodic-exporting-metricreader) - [MetricExporter](#metricexporter) - [Push Metric Exporter](#push-metric-exporter) - [Interface Definition](#interface-definition) - [Export(batch)](#exportbatch) - [ForceFlush()](#forceflush) - [Shutdown()](#shutdown) - [Pull Metric Exporter](#pull-metric-exporter) - [MetricProducer](#metricproducer) - [Interface Definition](#interface-definition-1) - [Produce() batch](#produce-batch) - [Defaults and configuration](#defaults-and-configuration) - [Numerical limits handling](#numerical-limits-handling) - [Compatibility requirements](#compatibility-requirements) - [Concurrency requirements](#concurrency-requirements)

Users of OpenTelemetry need a way for instrumentation interactions with the OpenTelemetry API to actually produce telemetry. The OpenTelemetry SDK (henceforth referred to as the SDK) is an implementation of the OpenTelemetry API that provides users with this functionally.

All language implementations of OpenTelemetry MUST provide an SDK.

MeterProvider¶

Status: Stable

A MeterProvider MUST provide a way to allow a Resource to be specified. If a Resource is specified, it SHOULD be associated with all the metrics produced by any Meter from the MeterProvider. The tracing SDK specification has provided some suggestions regarding how to implement this efficiently.

MeterProvider Creation¶

The SDK SHOULD allow the creation of multiple independent MeterProviders.

Meter Creation¶

New Meter instances are always created through a MeterProvider (see API). The name, version (optional), schema_url (optional), and attributes (optional) arguments supplied to the MeterProvider MUST be used to create an InstrumentationScope instance which is stored on the created Meter.

In the case where an invalid name (null or empty string) is specified, a working Meter MUST be returned as a fallback rather than returning null or throwing an exception, its name SHOULD keep the original invalid value, and a message reporting that the specified value is invalid SHOULD be logged.

When a Schema URL is passed as an argument when creating a Meter the emitted telemetry for that Meter MUST be associated with the Schema URL, provided that the emitted data format is capable of representing such association.

Configuration (i.e., MetricExporters, MetricReaders and Views) MUST be managed solely by the MeterProvider and the SDK MUST provide a way to configure all options that are implemented by the SDK. This MAY be done at the time of MeterProvider creation if appropriate.

The MeterProvider MAY provide methods to update the configuration. If configuration is updated (e.g., adding a MetricReader), the updated configuration MUST also apply to all already returned Meters (i.e. it MUST NOT matter whether a Meter was obtained from the MeterProvider before or after the configuration change). Note: Implementation-wise, this could mean that Meter instances have a reference to their MeterProvider and access configuration only via this reference.

Shutdown¶

This method provides a way for provider to do any cleanup required.

Shutdown MUST be called only once for each MeterProvider instance. After the call to Shutdown, subsequent attempts to get a Meter are not allowed. SDKs SHOULD return a valid no-op Meter for these calls, if possible.

Shutdown SHOULD provide a way to let the caller know whether it succeeded, failed or timed out.

Shutdown SHOULD complete or abort within some timeout. Shutdown MAY be implemented as a blocking API or an asynchronous API which notifies the caller via a callback or an event. OpenTelemetry SDK authors MAY decide if they want to make the shutdown timeout configurable.

Shutdown MUST be implemented at least by invoking Shutdown on all registered MetricReader and MetricExporter instances.

ForceFlush¶

This method provides a way for provider to notify the registered MetricReader and MetricExporter instances, so they can do as much as they could to consume or send the metrics. Note: unlike Push Metric Exporter which can send data on its own schedule, Pull Metric Exporter can only send the data when it is being asked by the scraper, so ForceFlush would not make much sense.

ForceFlush SHOULD provide a way to let the caller know whether it succeeded, failed or timed out. ForceFlush SHOULD return some ERROR status if there is an error condition; and if there is no error condition, it should return some NO ERROR status, language implementations MAY decide how to model ERROR and NO ERROR.

ForceFlush SHOULD complete or abort within some timeout. ForceFlush MAY be implemented as a blocking API or an asynchronous API which notifies the caller via a callback or an event. OpenTelemetry SDK authors MAY decide if they want to make the flush timeout configurable.

ForceFlush MUST invoke ForceFlush on all registered MetricReader and Push Metric Exporter instances.

View¶

A View provides SDK users with the flexibility to customize the metrics that are output by the SDK. Here are some examples when a View might be needed:

Customize which Instruments are to be processed/ignored. For example, an instrumented library can provide both temperature and humidity, but the application developer might only want temperature.
Customize the aggregation - if the default aggregation associated with the Instrument does not meet the needs of the user. For example, an HTTP client library might expose HTTP client request duration as Histogram by default, but the application developer might only want the total count of outgoing requests.
Customize which attribute(s) are to be reported on metrics. For example, an HTTP server library might expose HTTP verb (e.g. GET, POST) and HTTP status code (e.g. 200, 301, 404). The application developer might only care about HTTP status code (e.g. reporting the total count of HTTP requests for each HTTP status code). There could also be extreme scenarios in which the application developer does not need any attributes (e.g. just get the total count of all incoming requests).

The SDK MUST provide the means to register Views with a MeterProvider. Here are the inputs:

The Instrument selection criteria (required), which covers:
The type of the Instrument(s) (optional).
The name of the Instrument(s). OpenTelemetry SDK authors MAY choose to support wildcard characters, with the question mark (?) matching exactly one character and the asterisk character (*) matching zero or more characters. If wildcards are not supported in general, OpenTelemetry SDKs MUST specifically recognize the single * wildcard as matching all instruments.
The unit of the Instrument(s) (optional).
The name of the Meter (optional).
The version of the Meter (optional).
The schema_url of the Meter (optional).
OpenTelemetry SDK authors MAY choose to support more criteria. For example, a strong typed language MAY support point type (e.g. allow the users to select Instruments based on whether the underlying type is integer or double).
The criteria SHOULD be treated as additive, which means the Instrument has to meet all the provided criteria. For example, if the criteria are instrument name == "Foobar" and instrument type is Histogram, it will be treated as (instrument name == "Foobar") AND (instrument type is Histogram).
If no criteria is provided, the SDK SHOULD treat it as an error. It is recommended that the SDK implementations fail fast. Please refer to Error handling in OpenTelemetry for the general guidance.
The name of the View (optional). If not provided, the Instrument name MUST be used by default. This will be used as the name of the metrics stream.
The configuration for the resulting metrics stream:
The description. If not provided, the Instrument description MUST be used by default.
A list of attribute keys (optional). If provided, the attributes that are not in the list will be ignored. If not provided, all the attribute keys will be used by default (TODO: once the Hint API is available, the default behavior should respect the Hint if it is available).
The aggregation (optional) to be used. If not provided, the SDK MUST apply a default aggregation configurable on the basis of instrument kind according to the MetricReader instance.
Status: Feature-freeze - the exemplar_reservoir (optional) to use for storing exemplars. This should be a factory or callback similar to aggregation which allows different reservoirs to be chosen by the aggregation.
Status: Experimental - the aggregation_cardinality_limit (optional) associated with the view. This should be a positive integer to be taken as a hard limit on the number of data points that will be emitted during a single collection by a single instrument. See cardinality limits, below.

In order to avoid conflicts, views which specify a name SHOULD have an instrument selector that selects at most one instrument. For the registration mechanism described above, where selection is provided via configuration, the SDK SHOULD NOT allow Views with a specified name to be declared with instrument selectors that may select more than one instrument (e.g. wild card instrument name) in the same Meter. For this and other cases where registering a view will cause a conflict, SDKs MAY fail fast in accordance with initialization error handling principles.

The SDK SHOULD use the following logic to determine how to process Measurements made with an Instrument:

Determine the MeterProvider which "owns" the Instrument.
If the MeterProvider has no View registered, take the Instrument and apply the default Aggregation on the basis of instrument kind according to the MetricReader instance's aggregation property.
If the MeterProvider has one or more View(s) registered:
For each View, if the Instrument could match the instrument selection criteria:
- Try to apply the View configuration. If applying the View results in conflicting metric identities the implementation SHOULD apply the View and emit a warning. If it is not possible to apply the View without producing semantic errors (e.g. the View sets an asynchronous instrument to use the Explicit bucket histogram aggregation) the implementation SHOULD emit a warning and proceed as if the View did not exist.
If the Instrument could not match with any of the registered View(s), the SDK SHOULD enable the instrument using the default aggregation and temporality. Users can configure match-all Views using Drop aggregation to disable instruments by default.

Here are some examples:

# Python
'''
+------------------+
| MeterProvider    |
|   Meter A        |
|     Counter X    |
|     Histogram Y  |
|   Meter B        |
|     Gauge Z      |
+------------------+
'''

# metrics from X and Y (reported as Foo and Bar) will be exported
meter_provider
    .add_view("X")
    .add_view("Foo", instrument_name="Y")
    .add_view(
        "Bar",
        instrument_name="Y",
        aggregation=HistogramAggregation(buckets=[5.0, 10.0, 25.0, 50.0, 100.0]))
    .add_metric_reader(PeriodicExportingMetricReader(ConsoleExporter()))

# all the metrics will be exported using the default configuration
meter_provider.add_metric_reader(PeriodicExportingMetricReader(ConsoleExporter()))

# all the metrics will be exported using the default configuration
meter_provider
    .add_view("*") # a wildcard view that matches everything
    .add_metric_reader(PeriodicExportingMetricReader(ConsoleExporter()))

# Counter X will be exported as cumulative sum
meter_provider
    .add_view("X", aggregation=SumAggregation())
    .add_metric_reader(PeriodicExportingMetricReader(ConsoleExporter()))

# Counter X will be exported as delta sum
# Histogram Y and Gauge Z will be exported with 2 attributes (a and b)
meter_provider
    .add_view("X", aggregation=SumAggregation())
    .add_view("*", attribute_keys=["a", "b"])
    .add_metric_reader(PeriodicExportingMetricReader(ConsoleExporter()),
              temporality=lambda kind: Delta if kind in [Counter, AsyncCounter, Histogram] else Cumulative)

Aggregation¶

An Aggregation, as configured via the View, informs the SDK on the ways and means to compute Aggregated Metrics from incoming Instrument Measurements.

Note: the term aggregation is used instead of aggregator. It is recommended that implementors reserve the "aggregator" term for the future when the SDK allows custom aggregation implementations.

An Aggregation specifies an operation (i.e. decomposable aggregate function like Sum, Histogram, Min, Max, Count) and optional configuration parameter overrides. The operation's default configuration parameter values will be used unless overridden by optional configuration parameter overrides.

Note: Implementors MAY choose the best idiomatic practice for their language to represent the semantic of an Aggregation and optional configuration parameters.

e.g. The View specifies an Aggregation by string name (i.e. "ExplicitBucketHistogram").

# Use Histogram with custom boundaries
meter_provider
  .add_view(
    "X",
    aggregation="ExplicitBucketHistogram",
    aggregation_params={"Boundaries": [0, 10, 100]}
    )

e.g. The View specifies an Aggregation by class/type instance.

// Use Histogram with custom boundaries
meterProviderBuilder
  .AddView(
    instrumentName: "X",
    aggregation: new ExplicitBucketHistogramAggregation(
      boundaries: new double[] { 0.0, 10.0, 100.0 }
    )
  );

TODO: after we release the initial Stable version of Metrics SDK specification, we will explore how to allow configuring custom ExemplarReservoirs with the View API.

The SDK MUST provide the following Aggregation to support the Metric Points in the Metrics Data Model.

The SDK SHOULD provide the following Aggregation:

Base2 Exponential Bucket Histogram

Drop Aggregation¶

The Drop Aggregation informs the SDK to ignore/drop all Instrument Measurements for this Aggregation.

This Aggregation does not have any configuration parameters.

Default Aggregation¶

The Default Aggregation informs the SDK to use the Instrument kind to select an aggregation and advice to influence aggregation configuration parameters (as noted in the "Selected Aggregation" column).

Instrument Kind	Selected Aggregation
Counter	Sum Aggregation
Asynchronous Counter	Sum Aggregation
UpDownCounter	Sum Aggregation
Asynchronous UpDownCounter	Sum Aggregation
Asynchronous Gauge	Last Value Aggregation
Histogram	Explicit Bucket Histogram Aggregation, with `ExplicitBucketBoundaries` from advice if provided

This Aggregation does not have any configuration parameters.

Sum Aggregation¶

The Sum Aggregation informs the SDK to collect data for the Sum Metric Point.

The monotonicity of the aggregation is determined by the instrument type:

Instrument Kind	`SumType`
Counter	Monotonic
UpDownCounter	Non-Monotonic
Histogram	Monotonic
Asynchronous Gauge	Non-Monotonic
Asynchronous Counter	Monotonic
Asynchronous UpDownCounter	Non-Monotonic

This Aggregation does not have any configuration parameters.

This Aggregation informs the SDK to collect:

The arithmetic sum of Measurement values.

Last Value Aggregation¶

The Last Value Aggregation informs the SDK to collect data for the Gauge Metric Point.

This Aggregation does not have any configuration parameters.

This Aggregation informs the SDK to collect:

The last Measurement.
The timestamp of the last Measurement.

Histogram Aggregations¶

All histogram Aggregations inform the SDK to collect:

Count of Measurement values in population.
Arithmetic sum of Measurement values in population. This SHOULD NOT be collected when used with instruments that record negative measurements (e.g. UpDownCounter or ObservableGauge).
Min (optional) Measurement value in population.
Max (optional) Measurement value in population.

Explicit Bucket Histogram Aggregation¶

The Explicit Bucket Histogram Aggregation informs the SDK to collect data for the Histogram Metric Point using a set of explicit boundary values for histogram bucketing.

This Aggregation honors the following configuration parameters:

Key	Value	Default Value	Description
Boundaries	double[]	[ 0, 5, 10, 25, 50, 75, 100, 250, 500, 750, 1000, 2500, 5000, 7500, 10000 ]	Array of increasing values representing explicit bucket boundary values. The Default Value represents the following buckets (heavily influenced by the default buckets of Prometheus clients, e.g. Java and Go): (-∞, 0], (0, 5.0], (5.0, 10.0], (10.0, 25.0], (25.0, 50.0], (50.0, 75.0], (75.0, 100.0], (100.0, 250.0], (250.0, 500.0], (500.0, 750.0], (750.0, 1000.0], (1000.0, 2500.0], (2500.0, 5000.0], (5000.0, 7500.0], (7500.0, 10000.0], (10000.0, +∞). SDKs SHOULD use the default value when boundaries are not explicitly provided, unless they have good reasons to use something different (e.g. for backward compatibility reasons in a stable SDK release).
RecordMinMax	true, false	true	Whether to record min and max.

Explicit buckets are stated in terms of their upper boundary. Buckets are exclusive of their lower boundary and inclusive of their upper bound (except at positive infinity). A measurement is defined to fall into the greatest-numbered bucket with boundary that is greater than or equal to the measurement.

Base2 Exponential Bucket Histogram Aggregation¶

The Base2 Exponential Histogram Aggregation informs the SDK to collect data for the Exponential Histogram Metric Point, which uses a base-2 exponential formula to determine bucket boundaries and an integer scale parameter to control resolution. Implementations adjust scale as necessary given the data.

This Aggregation honors the following configuration parameters:

Key	Value	Default Value	Description
MaxSize	integer	160	Maximum number of buckets in each of the positive and negative ranges, not counting the special zero bucket.
MaxScale	integer	20	Maximum `scale` factor.
RecordMinMax	true, false	true	Whether to record min and max.

The default of 160 buckets is selected to establish default support for a high-resolution histogram able to cover a long-tail latency distribution from 1ms to 100s with less than 5% relative error. Because 160 can be factored into 10 * 2**K, maximum contrast is relatively simple to derive for scale K:

Scale	Maximum data contrast at 10 * 2**K buckets
K+2	5.657 (2**(10/4))
K+1	32 (2**(10/2))
K	1024 (2**10)
K-1	1048576 (2**20)

The following table shows how the ideal scale for 160 buckets is calculated as a function of the input range:

Input range	Contrast	Ideal Scale	Base	Relative error
1ms - 4ms	4	6	1.010889	0.542%
1ms - 20ms	20	5	1.021897	1.083%
1ms - 1s	10**3	4	1.044274	2.166%
1ms - 100s	10**5	3	1.090508	4.329%
1μs - 10s	10**7	2	1.189207	8.643%

Note that relative error is calculated as half of the bucket width divided by the bucket midpoint, which is the same in every bucket. Using the bucket from [1, base), we have (bucketWidth / 2) / bucketMidpoint = ((base - 1) / 2) / ((base + 1) / 2) = (base - 1) / (base + 1).

This Aggregation uses the notion of "ideal" scale. The ideal scale is either:

The MaxScale (see configuration parameters), generally used for single-value histogram Aggregations where scale is not otherwise constrained.
The largest value of scale such that no more than the maximum number of buckets are needed to represent the full range of input data in either of the positive or negative ranges.

Handle all normal values¶

Implementations are REQUIRED to accept the entire normal range of IEEE floating point values (i.e., all values except for +Inf, -Inf and NaN values).

Implementations SHOULD NOT incorporate non-normal values (i.e., +Inf, -Inf, and NaNs) into the sum, min, and max fields, because these values do not map into a valid bucket.

Implementations MAY round subnormal values away from zero to the nearest normal value.

Support a minimum and maximum scale¶

The implementation MUST maintain reasonable minimum and maximum scale parameters that the automatic scale parameter will not exceed. The maximum scale is defined by the MaxScale configuration parameter.

Use the maximum scale for single measurements¶

When the histogram contains not more than one value in either of the positive or negative ranges, the implementation SHOULD use the maximum scale.

Maintain the ideal scale¶

Implementations SHOULD adjust the histogram scale as necessary to maintain the best resolution possible, within the constraint of maximum size (max number of buckets). Best resolution (highest scale) is achieved when the number of positive or negative range buckets exceeds half the maximum size, such that increasing scale by one would not be possible given the size constraint.

Observations inside asynchronous callbacks¶

Callback functions MUST be invoked for the specific MetricReader performing collection, such that observations made or produced by executing callbacks only apply to the intended MetricReader during collection.

The implementation SHOULD disregard the accidental use of APIs appurtenant to asynchronous instruments outside of registered callbacks in the context of a single MetricReader collection.

The implementation SHOULD use a timeout to prevent indefinite callback execution.

The implementation MUST complete the execution of all callbacks for a given instrument before starting a subsequent round of collection.

Cardinality limits¶

Status: Experimental

Views SHOULD support being configured with a cardinality limit to be applied to all aggregators not configured by a specific view, specified via MetricReader configuration.

View configuration SHOULD support applying per-aggregation cardinality limits.

The cardinality limit is taken as an exact, hard limit on the number of data points that can be written per collection, per aggregation. Each aggregation configured view MUST NOT output more than the configured aggregation_cardinality_limit number of data points per period.

The RECOMMENDED default aggregation cardinality limit is 2000.

An overflow attribute set is defined, containing a single attribute otel.metric.overflow having (boolean) value true, which is used to report a synthetic aggregation of the metric events that could not be independently aggregated because of the limit.

The SDK MUST create an Aggregator with the overflow attribute set prior to reaching the cardinality limit and use it to aggregate events for which the correct Aggregator could not be created. The maximum number of distinct, non-overflow attributes is one less than the limit, as a result.

Synchronous instrument cardinality limits¶

Views of synchronous instruments with cumulative aggregation temporality MUST continue to export the all attribute sets that were observed prior to the beginning of overflow. Metric events corresponding with attribute sets that were not observed prior to the overflow will be reflected in a single data point described by (only) the overflow attribute.

Views of synchronous instruments with delta aggregation temporality MAY choose an arbitrary subset of attribute sets to output to maintain the stated cardinality limit.

Regardless of aggregation temporality, the SDK MUST ensure that every metric event is reflected in exactly one Aggregator, which is either an Aggregator associated with the correct attribute set or an aggregator associated with the overflow attribute set.

Events MUST NOT be double-counted or dropped during an overflow.

Asynchronous instrument cardinality limits¶

Views of asynchronous instruments SHOULD prefer the first-observed attributes in the callback when limiting cardinality, regardless of aggregation temporality.

Meter¶

Distinct meters MUST be treated as separate namespaces for the purposes of detecting duplicate instrument registrations.

Duplicate instrument registration¶

When more than one Instrument of the same name is created for identical Meters, denoted duplicate instrument registration, the Meter MUST create a valid Instrument in every case. Here, "valid" means an instrument that is functional and can be expected to export data, despite potentially creating a semantic error in the data model.

It is unspecified whether or under which conditions the same or different Instrument instance will be returned as a result of duplicate instrument registration. The term identical applied to Instruments describes instances where all identifying fields are equal. The term distinct applied to Instruments describes instances where at least one field value is different.

Based on the recommendations from the data model, the SDK MUST aggregate data from identical Instruments together in its export pipeline.

When a duplicate instrument registration occurs, and it is not corrected with a View, a warning SHOULD be emitted. The emitted warning SHOULD include information for the user on how to resolve the conflict, if possible.

If the potential conflict involves multiple description properties, setting the description through a configured View SHOULD avoid the warning.
If the potential conflict involves instruments that can be distinguished by a supported View selector (e.g., instrument type) a renaming View recipe SHOULD be included in the warning.
Otherwise (e.g., use of multiple units), the SDK SHOULD pass through the data by reporting both Metric objects and emit a generic warning describing the duplicate instrument registration.

Instrument name¶

When a Meter creates an instrument, it SHOULD validate the instrument name conforms to the instrument name syntax

If the instrument name does not conform to this syntax, the Meter SHOULD emit an error notifying the user about the invalid name. It is left unspecified if a valid instrument is also returned.

Instrument unit¶

When a Meter creates an instrument, it SHOULD NOT validate the instrument unit. If a unit is not provided or the unit is null, the Meter MUST treat it the same as an empty unit string.

Instrument description¶

When a Meter creates an instrument, it SHOULD NOT validate the instrument description. If a description is not provided or the description is null, the Meter MUST treat it the same as an empty description string.

Instrument advice¶

Status: Experimental

When a Meter creates an instrument, it SHOULD validate the instrument advice parameters. If an advice parameter is not valid, the Meter SHOULD emit an error notifying the user and proceed as if the parameter was not provided.

Attribute limits¶

Status: Stable

Attributes which belong to Metrics are exempt from the common rules of attribute limits at this time. Attribute truncation or deletion could affect identity of metric time series and the topic requires further analysis.

Exemplar¶

Status: Feature-freeze

Exemplars are example data points for aggregated data. They provide specific context to otherwise general aggregations. Exemplars allow correlation between aggregated metric data and the original API calls where measurements are recorded. Exemplars work for trace-metric correlation across any metric, not just those that can also be derived from Spans.

An Exemplar is a recorded Measurement that exposes the following pieces of information:

The value of the Measurement that was recorded by the API call.
The time the API call was made to record a Measurement.
The set of Attributes associated with the Measurement not already included in a metric data point.
The associated trace id and span id of the active Span within Context of the Measurement at API call time.

For example, if a user has configured a View to preserve the attributes: X and Y, but the user records a measurement as follows:

const span = tracer.startSpan('makeRequest');
api.context.with(api.trace.setSpan(api.context.active(), span), () => {
  // Record a measurement.
  cache_miss_counter.add(1, {"X": "x-value", "Y": "y-value", "Z": "z-value"});
  ...
  span.end();
})

Then an exemplar output in OTLP would consist of:

The value of 1.
The time when the add method was called
The Attributes of {"Z": "z-value"}, as these are not preserved in the resulting metric point.
The trace/span id for the makeRequest span.

While the metric data point for the counter would carry the attributes X and Y.

A Metric SDK MUST provide a mechanism to sample Exemplars from measurements via the ExemplarFilter and ExemplarReservoir hooks.

Exemplar sampling SHOULD be turned off by default. If Exemplar sampling is off, the SDK MUST NOT have overhead related to exemplar sampling.

A Metric SDK MUST allow exemplar sampling to leverage the configuration of metric aggregation. For example, Exemplar sampling of histograms should be able to leverage bucket boundaries.

A Metric SDK SHOULD provide extensible hooks for Exemplar sampling, specifically:

ExemplarFilter: filter which measurements can become exemplars.
ExemplarReservoir: storage and sampling of exemplars.

ExemplarFilter¶

The ExemplarFilter interface MUST provide a method to determine if a measurement should be sampled. Sampled here simply makes the measurement eligible for being included as an exemplar. ExemplarReservoir makes the final decision if a measurement becomes an exemplar.

This interface SHOULD have access to:

The value of the measurement.
The complete set of Attributes of the measurement.
The Context of the measurement, which covers the Baggage and the current active Span.
A timestamp that best represents when the measurement was taken.

Built-in ExemplarFilters¶

OpenTelemetry supports a number of built-in exemplar filters to choose from. The default is TraceBased.

AlwaysOn¶

An ExemplarFilter which makes all measurements eligible for being an Exemplar.

AlwaysOff¶

An ExemplarFilter which makes no measurements eligible for being an Exemplar. Using this ExemplarFilter is as good as disabling Exemplar feature.

TraceBased¶

An ExemplarFilter which makes those measurements eligible for being an Exemplar, which are recorded in the context of a sampled parent span.

ExemplarReservoir¶

The ExemplarReservoir interface MUST provide a method to offer measurements to the reservoir and another to collect accumulated Exemplars.

The "offer" method SHOULD accept measurements, including:

The value of the measurement.
The complete set of Attributes of the measurement.
The Context of the measurement, which covers the Baggage and the current active Span.
A timestamp that best represents when the measurement was taken.

The "offer" method SHOULD have the ability to pull associated trace and span information without needing to record full context. In other words, current span context and baggage can be inspected at this point.

The "offer" method does not need to store all measurements it is given and MAY further sample beyond the ExemplarFilter.

The "collect" method MUST return accumulated Exemplars. Exemplars are expected to abide by the AggregationTemporality of any metric point they are recorded with. In other words, Exemplars reported against a metric data point SHOULD have occurred within the start/stop timestamps of that point. SDKs are free to decide whether "collect" should also reset internal storage for delta temporal aggregation collection, or use a more optimal implementation.

Exemplars MUST retain any attributes available in the measurement that are not preserved by aggregation or view configuration. Specifically, at a minimum, joining together attributes on an Exemplar with those available on its associated metric data point should result in the full set of attributes from the original sample measurement.

The ExemplarReservoir SHOULD avoid allocations when sampling exemplars.

Exemplar defaults¶

The SDK will come with two types of built-in exemplar reservoirs:

SimpleFixedSizeExemplarReservoir
AlignedHistogramBucketExemplarReservoir

By default, explicit bucket histogram aggregation with more than 1 bucket will use AlignedHistogramBucketExemplarReservoir. All other aggregations will use SimpleFixedSizeExemplarReservoir.

SimpleExemplarReservoir This Exemplar reservoir MAY take a configuration parameter for the size of the reservoir pool. The reservoir will accept measurements using an equivalent of the naive reservoir sampling algorithm

bucket = random_integer(0, num_measurements_seen)
if bucket < num_buckets then
  reservoir[bucket] = measurement
end

Additionally, the num_measurements_seen count SHOULD be reset at every collection cycle.

AlignedHistogramBucketExemplarReservoir This Exemplar reservoir MUST take a configuration parameter that is the configuration of a Histogram. This implementation MUST keep the last seen measurement that falls within a histogram bucket. The reservoir will accept measurements using the equivalent of the following naive algorithm:

bucket = find_histogram_bucket(measurement)
if bucket < num_buckets then
  reservoir[bucket] = measurement
end

def find_histogram_bucket(measurement):
  for boundary, idx in bucket_boundaries do
    if value <= boundary then
      return idx
    end
  end
  return boundaries.length

MetricReader¶

Status: Stable

MetricReader is an SDK implementation object that provides the common configurable aspects of the OpenTelemetry Metrics SDK and determines the following capabilities:

Registering MetricProducer(s)
Collecting metrics from the SDK and any registered MetricProducers on demand.
Handling the ForceFlush and Shutdown signals from the SDK.

To construct a MetricReader when setting up an SDK, the caller SHOULD provide at least the following:

The exporter to use, which is a MetricExporter instance.
The default output aggregation (optional), a function of instrument kind. If not configured, the default aggregation SHOULD be used.
The default output temporality (optional), a function of instrument kind. If not configured, the Cumulative temporality SHOULD be used.
The default aggregation cardinality limit to use, a function of instrument kind. If not configured, a default value of 2000 SHOULD be used.

The MetricReader.Collect method allows general-purpose MetricExporter instances to explicitly initiate collection, commonly used with pull-based metrics collection. A common sub-class of MetricReader, the periodic exporting MetricReader SHOULD be provided to be used typically with push-based metrics collection.

The MetricReader MUST ensure that data points from OpenTelemetry instruments are output in the configured aggregation temporality for each instrument kind. For synchronous instruments being output with Cumulative temporality, this means converting Delta to Cumulative aggregation temporality. For asynchronous instruments being output with Delta temporality, this means converting Cumulative to Delta aggregation temporality.

The MetricReader is not required to ensure data points from a non-SDK MetricProducer are output in the configured aggregation temporality, as these data points are not collected using OpenTelemetry instruments.

The SDK MUST support multiple MetricReader instances to be registered on the same MeterProvider, and the MetricReader.Collect invocation on one MetricReader instance SHOULD NOT introduce side-effects to other MetricReader instances. For example, if a MetricReader instance is receiving metric data points that have delta temporality, it is expected that SDK will update the time range - e.g. from (T_n, T_n+1] to (T_n+1, T_n+2] - ONLY for this particular MetricReader instance.

The SDK MUST NOT allow a MetricReader instance to be registered on more than one MeterProvider instance.

+-----------------+            +--------------+
|                 | Metrics... |              |
| In-memory state +------------> MetricReader |
|                 |            |              |
+-----------------+            +--------------+

+-----------------+            +--------------+
|                 | Metrics... |              |
| In-memory state +------------> MetricReader |
|                 |            |              |
+-----------------+            +--------------+

The SDK SHOULD provide a way to allow MetricReader to respond to MeterProvider.ForceFlush and MeterProvider.Shutdown. OpenTelemetry SDK authors MAY decide the language idiomatic approach, for example, as OnForceFlush and OnShutdown callback functions.

MetricReader operations¶

RegisterProducer(metricProducer)¶

Status: Experimental

RegisterProducer causes the MetricReader to use the provided MetricProducer as a source of aggregated metric data in subsequent invocations of Collect. RegisterProducer is expected to be called during initialization, but MAY be invoked later. Multiple registrations of the same MetricProducer MAY result in duplicate metric data being collected.

If the MeterProvider is an instance of MetricProducer, this MAY be used to register the MeterProvider, but MUST NOT allow multiple MeterProviders to be registered with the same MetricReader.

Collect¶

Collects the metrics from the SDK and any registered MetricProducers. If there are asynchronous SDK Instruments involved, their callback functions will be triggered.

Collect SHOULD provide a way to let the caller know whether it succeeded, failed or timed out. When the Collect operation fails or times out on some of the instruments, the SDK MAY return successfully collected results and a failed reasons list to the caller.

Collect does not have any required parameters, however, OpenTelemetry SDK authors MAY choose to add parameters (e.g. callback, filter, timeout). OpenTelemetry SDK authors MAY choose the return value type, or do not return anything.

Note: it is expected that the MetricReader.Collect implementations will be provided by the SDK, so it is RECOMMENDED to prevent the user from accidentally overriding it, if possible (e.g. final in C++ and Java, sealed in C#).

Shutdown¶

This method provides a way for the MetricReader to do any cleanup required.

Shutdown MUST be called only once for each MetricReader instance. After the call to Shutdown, subsequent invocations to Collect are not allowed. SDKs SHOULD return some failure for these calls, if possible.

Shutdown SHOULD provide a way to let the caller know whether it succeeded, failed or timed out.

Shutdown SHOULD complete or abort within some timeout. Shutdown MAY be implemented as a blocking API or an asynchronous API which notifies the caller via a callback or an event. OpenTelemetry SDK authors MAY decide if they want to make the shutdown timeout configurable.

Periodic exporting MetricReader¶

This is an implementation of the MetricReader which collects metrics based on a user-configurable time interval, and passes the metrics to the configured Push Metric Exporter.

Configurable parameters:

exportIntervalMillis - the time interval in milliseconds between two consecutive exports. The default value is 60000 (milliseconds).
exportTimeoutMillis - how long the export can run before it is cancelled. The default value is 30000 (milliseconds).

One possible implementation of periodic exporting MetricReader is to inherit from MetricReader and start a background task which calls the inherited Collect() method at the requested exportIntervalMillis. The reader's Collect() method may still be invoked by other callers. For example,

A user configures periodic exporting MetricReader with a push exporter and a 30 second interval.
At the first 30 second interval, the background task calls Collect() which passes metrics to the push exporter.
After 15 seconds, the user decides to flush metrics for just this reader. They call Collect() which passes metrics to the push exporter.
After another 15 seconds (at the end of the second 30 second interval), the background task calls Collect() which passes metrics to the push exporter.

MetricExporter¶

Status: Stable

MetricExporter defines the interface that protocol-specific exporters MUST implement so that they can be plugged into OpenTelemetry SDK and support sending of telemetry data.

Metric Exporters always have an associated MetricReader. The aggregation and temporality properties used by the OpenTelemetry Metric SDK are determined when registering Metric Exporters through their associated MetricReader. OpenTelemetry language implementations MAY support automatically configuring the MetricReader to use for an Exporter.

The goal of the interface is to minimize burden of implementation for protocol-dependent telemetry exporters. The protocol exporter is expected to be primarily a simple telemetry data encoder and transmitter.

Metric Exporter has access to the aggregated metrics data. Metric Exporters SHOULD report an error condition for data output by the MetricReader with unsupported Aggregation or Aggregation Temporality, as this condition can be corrected by a change of MetricReader configuration.

There could be multiple Push Metric Exporters or Pull Metric Exporters or even a mixture of both configured at the same time on a given MeterProvider using one MetricReader for each exporter. Different exporters can run at different schedule, for example:

Exporter A is a push exporter which sends data every 1 minute.
Exporter B is a push exporter which sends data every 5 seconds.
Exporter C is a pull exporter which reacts to a scraper over HTTP.
Exporter D is a pull exporter which reacts to another scraper over a named pipe.

Push Metric Exporter¶

Push Metric Exporter sends metric data it receives from a paired MetricReader. Here are some examples:

Sends the data based on a user configured schedule, e.g. every 1 minute. This MAY be accomplished by pairing the exporter with a periodic exporting MetricReader.
Sends the data when there is a severe error.

The following diagram shows Push Metric Exporter's relationship to other components in the SDK:

+-----------------+            +---------------------------------+
|                 | Metrics... |                                 |
| In-memory state +------------> Periodic exporting MetricReader |
|                 |            |                                 |
+-----------------+            |    +-----------------------+    |
                               |    |                       |    |
                               |    | MetricExporter (push) +-------> Another process
                               |    |                       |    |
                               |    +-----------------------+    |
                               |                                 |
                               +---------------------------------+

Interface Definition¶

A Push Metric Exporter MUST support the following functions:

Export(batch)¶

Exports a batch of Metric points. Protocol exporters that will implement this function are typically expected to serialize and transmit the data to the destination.

The SDK MUST provide a way for the exporter to get the Meter information (e.g. name, version, etc.) associated with each Metric point.

Export will never be called concurrently for the same exporter instance. Export can be called again only after the current call returns.

Export MUST NOT block indefinitely, there MUST be a reasonable upper limit after which the call must time out with an error result (Failure).

Any retry logic that is required by the exporter is the responsibility of the exporter. The default SDK SHOULD NOT implement retry logic, as the required logic is likely to depend heavily on the specific protocol and backend the metrics are being sent to.

Parameters:

batch - a batch of Metric points. The exact data type of the batch is language specific, typically it is some kind of list. The exact type of Metric point is language specific, and is typically optimized for high performance. Here are some examples:

       +--------+ +--------+     +--------+
Batch: | Metric | | Metric | ... | Metric |
       +---+----+ +--------+     +--------+
           |
           +--> name, unit, description, meter information, ...
           |
           |                  +-------------+ +-------------+     +-------------+
           +--> MetricPoints: | MetricPoint | | MetricPoint | ... | MetricPoint |
                              +-----+-------+ +-------------+     +-------------+
                                    |
                                    +--> timestamps, attributes, value (or buckets), exemplars, ...

Refer to the Metric points section from the Metrics Data Model specification for more details.

Note: it is highly recommended that implementors design the Metric data type based on the Data Model, rather than directly use the data types generated from the proto files (because the types generated from proto files are not guaranteed to be backward compatible).

Returns: ExportResult

ExportResult is one of:

Success - The batch has been successfully exported. For protocol exporters this typically means that the data is sent over the wire and delivered to the destination server.
Failure - exporting failed. The batch must be dropped. For example, this can happen when the batch contains bad data and cannot be serialized.

Note: this result may be returned via an async mechanism or a callback, if that is idiomatic for the language implementation.

ForceFlush()¶

This is a hint to ensure that the export of any Metrics the exporter has received prior to the call to ForceFlush SHOULD be completed as soon as possible, preferably before returning from this method.

ForceFlush SHOULD provide a way to let the caller know whether it succeeded, failed or timed out.

ForceFlush SHOULD only be called in cases where it is absolutely necessary, such as when using some FaaS providers that may suspend the process after an invocation, but before the exporter exports the completed metrics.

ForceFlush SHOULD complete or abort within some timeout. ForceFlush can be implemented as a blocking API or an asynchronous API which notifies the caller via a callback or an event. OpenTelemetry SDK authors MAY decide if they want to make the flush timeout configurable.

Shutdown()¶

Shuts down the exporter. Called when SDK is shut down. This is an opportunity for exporter to do any cleanup required.

Shutdown SHOULD be called only once for each MetricExporter instance. After the call to Shutdown subsequent calls to Export are not allowed and should return a Failure result.

Shutdown SHOULD NOT block indefinitely (e.g. if it attempts to flush the data and the destination is unavailable). OpenTelemetry SDK authors MAY decide if they want to make the shutdown timeout configurable.

Pull Metric Exporter¶

Pull Metric Exporter reacts to the metrics scrapers and reports the data passively. This pattern has been widely adopted by Prometheus.

Unlike Push Metric Exporter which can send data on its own schedule, pull exporter can only send the data when it is being asked by the scraper, and ForceFlush would not make sense.

Implementors MAY choose the best idiomatic design for their language. For example, they could generalize the Push Metric Exporter interface design and use that for consistency, they could model the pull exporter as MetricReader, or they could design a completely different pull exporter interface. If the pull exporter is modeled as MetricReader, implementors MAY name the MetricExporter interface as PushMetricExporter to prevent naming confusion.

The following diagram gives some examples on how Pull Metric Exporter can be modeled to interact with other components in the SDK:

Model the pull exporter as MetricReader

+-----------------+            +-----------------------------+
|                 | Metrics... |                             |
| In-memory state +------------> PrometheusExporter (pull)   +---> Another process (scraper)
|                 |            | (modeled as a MetricReader) |
+-----------------+            |                             |
                               +-----------------------------+

Use the same MetricExporter design for both push and pull exporters

+-----------------+            +-----------------------------+
|                 | Metrics... |                             |
| In-memory state +------------> Exporting MetricReader      |
|                 |            |                             |
+-----------------+            |  +-----------------------+  |
                               |  |                       |  |
                               |  | MetricExporter (pull) +------> Another process (scraper)
                               |  |                       |  |
                               |  +-----------------------+  |
                               |                             |
                               +-----------------------------+

MetricProducer¶

Status: Experimental

MetricProducer defines the interface which bridges to third-party metric sources MUST implement so they can be plugged into an OpenTelemetry MetricReader as a source of aggregated metric data. The SDK's in-memory state MAY implement the MetricProducer interface for convenience.

MetricProducer implementations SHOULD accept configuration for the AggregationTemporality of produced metrics. SDK authors MAY provide utility libraries to facilitate conversion between delta and cumulative temporalities.

If the batch of Metric points returned by Produce() includes a Resource, the MetricProducer MUST accept configuration for the Resource.

+-----------------+            +--------------+
|                 | Metrics... |              |
| In-memory state +------------> MetricReader |
|                 |            |              |
+-----------------+            |              |
                               |              |
+-----------------+            |              |
|                 | Metrics... |              |
| MetricProducer  +------------>              |
|                 |            |              |
+-----------------+            +--------------+

Interface Definition¶

A MetricProducer MUST support the following functions:

Produce() batch¶

Produce provides metrics from the MetricProducer to the caller. Produce MUST return a batch of Metric points. Produce does not have any required parameters, however, OpenTelemetry SDK authors MAY choose to add parameters (e.g. timeout).

Produce SHOULD provide a way to let the caller know whether it succeeded, failed or timed out. When the Produce operation fails, the MetricProducer MAY return successfully collected results and a failed reasons list to the caller.

If a batch of Metric points can include InstrumentationScope information, Produce SHOULD include a single InstrumentationScope which identifies the MetricProducer.

Defaults and configuration¶

The SDK MUST provide configuration according to the SDK environment variables specification.

Numerical limits handling¶

The SDK MUST handle numerical limits in a graceful way according to Error handling in OpenTelemetry.

If the SDK receives float/double values from Instruments, it MUST handle all the possible values. For example, if the language runtime supports IEEE 754, the SDK needs to handle NaNs and Infinites.

It is unspecified how the SDK should handle the input limits. The SDK authors MAY leverage/follow the language runtime behavior for better performance, rather than perform a check on each value coming from the API.

It is unspecified how the SDK should handle the output limits (e.g. integer overflow). The SDK authors MAY rely on the language runtime behavior as long as errors/exceptions are taken care of.

Compatibility requirements¶

Status: Stable

All the metrics components SHOULD allow new methods to be added to existing components without introducing breaking changes.

All the metrics SDK methods SHOULD allow optional parameter(s) to be added to existing methods without introducing breaking changes, if possible.

Concurrency requirements¶

Status: Stable

For languages which support concurrent execution the Metrics SDKs provide specific guarantees and safeties.

MeterProvider - Meter creation, ForceFlush and Shutdown are safe to be called concurrently.

ExemplarFilter - all methods are safe to be called concurrently.

ExemplarReservoir - all methods are safe to be called concurrently.

MetricReader - Collect and Shutdown are safe to be called concurrently.

MetricExporter - ForceFlush and Shutdown are safe to be called concurrently.