Skip to content

Latest commit

 

History

History
1579 lines (1222 loc) · 72.2 KB

sdk.md

File metadata and controls

1579 lines (1222 loc) · 72.2 KB
Raw

Metrics SDK

Status: Mixed

Table of Contents

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

It SHOULD only be possible to create Meter instances through a MeterProvider (see API).

The MeterProvider MUST implement the Get a Meter API.

The input provided by the user 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

Configuration (i.e. MetricExporters, MetricReaders and Views) MUST be owned by the MeterProvider. The configuration MAY be applied 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 instances that have an associated Push Metric Exporter, so they can do as much as they could to collect and send the metrics. Note: Pull Metric Exporter can only send the data when it is being asked by the scraper, so ForceFlush would not make much sense.

ForceFlush MUST invoke ForceFlush on all registered MetricReader instances that implement ForceFlush.

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.

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 functionality for a user to create Views for a MeterProvider. This functionality MUST accept as inputs the Instrument selection criteria and the resulting stream configuration.

The SDK MUST provide the means to register Views with a MeterProvider.

Instrument selection criteria

Instrument selection criteria are the predicates that determine if a View will be applied to an Instrument or not.

Criteria SHOULD be treated as additive. This means an Instrument has to match all the provided criteria for the View to be applied. 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).

The SDK MUST accept the following criteria:

  • name: The name of the Instrument(s) to match. This name is evaluated to match an Instrument in the following manner.

    1. If the value of name is *, the criterion matches all Instruments.
    2. If the value of name is exactly the same as an Instrument, then the criterion matches that instrument.

    Additionally, the SDK MAY support wildcard pattern matching for the name criterion using the following characters.

    • A question mark (?): matches any single character
    • An asterisk (*): matches any number of any characters including none

    If wildcard pattern matching is supported, the name criterion will match if the wildcard pattern is evaluated to match the Instrument name.

    If the SDK does not support wildcards in general, it MUST still recognize the special single asterisk (*) character as matching all Instruments.

    Users can provide a name, but it is up to their discretion. Therefore, the instrument selection criteria parameter needs to be structured to accept a name, but MUST NOT obligate a user to provide one.

  • type: The type of Instruments to match. If the value of type is the same as an Instrument's type, then the criterion matches that Instrument.

    Users can provide a type, but it is up to their discretion. Therefore, the instrument selection criteria parameter needs to be structured to accept a type, but MUST NOT obligate a user to provide one.

  • unit: If the value of unit is the same as an Instrument's unit, then the criterion matches that Instrument.

    Users can provide a unit, but it is up to their discretion. Therefore, the instrument selection criteria parameter needs to be structured to accept a unit, but MUST NOT obligate a user to provide one.

  • meter_name: If the value of meter_name is the same as the Meter that created an Instrument, then the criterion matches that Instrument.

    Users can provide a meter_name, but it is up to their discretion. Therefore, the instrument selection criteria parameter needs to be structured to accept a meter_name, but MUST NOT obligate a user to provide one.

  • meter_version: If the value of meter_version is the same version as the Meter that created an Instrument, then the criterion matches that Instrument.

    Users can provide a meter_version, but it is up to their discretion. Therefore, the instrument selection criteria parameter needs to be structured to accept a meter_version, but MUST NOT obligate a user to provide one.

  • meter_schema_url: If the value of meter_schema_url is the same schema URL as the Meter that created an Instrument, then the criterion matches that Instrument.

    Users can provide a meter_schema_url, but it is up to their discretion. Therefore, the instrument selection criteria parameter needs to be structured to accept a meter_schema_url, but MUST NOT obligate a user to provide one.

The SDK MAY accept additional criteria. For example, a strongly typed language may support point type criterion (e.g. allow the users to select Instruments based on whether the underlying number is integral or rational). Users can provide these additional criteria the SDK accepts, but it is up to their discretion. Therefore, the instrument selection criteria can be structured to accept the criteria, but MUST NOT obligate a user to provide them.

Stream configuration

Stream configuration are the parameters that define the metric stream a MeterProvider will use to define telemetry pipelines.

The SDK MUST accept the following stream configuration parameters:

  • name: The metric stream name that SHOULD be used.

    In order to avoid conflicts, if a name is provided the View SHOULD have an instrument selector that selects at most one instrument. If the Instrument selection criteria for a View with a stream configuration name parameter can select more than one instrument (i.e. wildcards) the SDK MAY fail fast in accordance with initialization error handling principles.

    Users can provide a name, but it is up to their discretion. Therefore, the stream configuration parameter needs to be structured to accept a name, but MUST NOT obligate a user to provide one. If the user does not provide a name value, name from the Instrument the View matches MUST be used by default.

  • description: The metric stream description that SHOULD be used.

    Users can provide a description, but it is up to their discretion. Therefore, the stream configuration parameter needs to be structured to accept a description, but MUST NOT obligate a user to provide one. If the user does not provide a description value, the description from the Instrument a View matches MUST be used by default.

  • attribute_keys: This is, at a minimum, an allow-list of attribute keys for measurements captured in the metric stream. The allow-list contains attribute keys that identify the attributes that MUST be kept, and all other attributes MUST be ignored.

    Implementations MAY accept additional attribute filtering functionality for this parameter.

    Users can provide attribute_keys, but it is up to their discretion. Therefore, the stream configuration parameter needs to be structured to accept attribute_keys, but MUST NOT obligate a user to provide them. If the user does not provide any value, the SDK SHOULD use the Attributes advisory parameter configured on the instrument instead. If the Attributes advisory parameter is absent, all attributes MUST be kept.

  • aggregation: The name of an aggregation function to use in aggregating the metric stream data.

    Users can provide an aggregation, but it is up to their discretion. Therefore, the stream configuration parameter needs to be structured to accept an aggregation, but MUST NOT obligate a user to provide one. If the user does not provide an aggregation value, the MeterProvider MUST apply a default aggregation configurable on the basis of instrument type according to the MetricReader instance.

  • Status: Feature-freeze - exemplar_reservoir: A functional type that generates an exemplar reservoir a MeterProvider will use when storing exemplars. This functional type needs to be a factory or callback similar to aggregation selection functionality which allows different reservoirs to be chosen by the aggregation.

    Users can provide an exemplar_reservoir, but it is up to their discretion. Therefore, the stream configuration parameter needs to be structured to accept an exemplar_reservoir, but MUST NOT obligate a user to provide one. If the user does not provide an exemplar_reservoir value, the MeterProvider MUST apply a default exemplar reservoir.

  • Status: Experimental - aggregation_cardinality_limit: A positive integer value defining the maximum number of data points allowed to be emitted in a collection cycle by a single instrument. See cardinality limits, below.

    Users can provide an aggregation_cardinality_limit, but it is up to their discretion. Therefore, the stream configuration parameter needs to be structured to accept an aggregation_cardinality_limit, but MUST NOT obligate a user to provide one. If the user does not provide an aggregation_cardinality_limit value, the MeterProvider MUST apply the default aggregation cardinality limit the MetricReader is configured with.

Measurement processing

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's stream 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.

View examples

The following are examples of an SDK's functionality to create Views for a MeterProvider.

# 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:

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 advisory parameters 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
Gauge Last Value Aggregation
Asynchronous Gauge Last Value Aggregation
Histogram Explicit Bucket Histogram Aggregation, with the ExplicitBucketBoundaries advisory parameter 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:

  1. The MaxScale (see configuration parameters), generally used for single-value histogram Aggregations where scale is not otherwise constrained.
  2. 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 use of asynchronous instrument APIs outside of registered callbacks.

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.

The implementation SHOULD NOT produce aggregated metric data for a previously-observed attribute set which is not observed during a successful callback. See MetricReader for more details on the persistence of metrics across successive collections.

Cardinality limits

Status: Experimental

SDKs SHOULD support being configured with a cardinality limit. A cardinality limit is the hard limit on the number of metric streams that can be collected.

Configuration

The cardinality limit for an aggregation is defined in one of three ways:

  1. A view with criteria matching the instrument an aggregation is created for has an aggregation_cardinality_limit value defined for the stream, that value SHOULD be used.
  2. If there is no matching view, but the MetricReader defines a default cardinality limit value based on the instrument an aggregation is created for, that value SHOULD be used.
  3. If none of the previous values are defined, the default value of 2000 SHOULD be used.

Overflow attribute

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

Aggregators for synchronous instruments with cumulative temporality MUST continue to export 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.

Aggregators 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

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

Meter

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

Duplicate instrument registration

A duplicate instrument registration occurs when more than one Instrument of the same name is created for identical Meters from the same MeterProvider but they have different identifying fields.

Whenever this occurs, users still need to be able to make measurements with the duplicate instrument. This means that the Meter MUST return a functional instrument that can be expected to export data even if this will cause semantic error in the data model.

Additionally, users need to be informed about this error. Therefore, 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.

  1. If the potential conflict involves multiple description properties, setting the description through a configured View SHOULD avoid the warning.
  2. If the potential conflict involves instruments that can be distinguished by a supported View selector (e.g. name, instrument kind) a renaming View recipe SHOULD be included in the warning.
  3. 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.

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.

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

Name conflict

The name of an Instrument is defined to be case-insensitive. If an SDK uses a case-sensitive encoding to represent this name, a duplicate instrument registration will occur when a user passes multiple casings of the same name. When this happens, the Meter MUST return an instrument using the first-seen instrument name and log an appropriate error as described above.

For example, if a user creates an instrument with the name requestCount and then makes another request to the same Meter to create an instrument with the name RequestCount, in both cases an instrument with the name requestCount needs to be returned to the user and a log message needs to be emitted for the second request.

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 advisory parameters

Status: Experimental

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

If multiple identical Instruments are created with different advisory parameters, the Meter MUST return an instrument using the first-seen advisory parameters and log an appropriate error as described in duplicate instrument registrations.

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 SHOULD include two types of built-in exemplar reservoirs:

  1. SimpleFixedSizeExemplarReservoir
  2. AlignedHistogramBucketExemplarReservoir

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

SimpleFixedSizeExemplarReservoir

This reservoir MUST use an uniformly-weighted sampling algorithm based on the number of samples the reservoir has seen so far to determine if the offered measurements should be sampled. For example, the simple reservoir sampling algorithm can be used:

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

Any stateful portion of sampling computation SHOULD be reset every collection cycle. For the above example, that would mean that the num_measurements_seen count is reset every time the reservoir is collected.

This Exemplar reservoir MAY take a configuration parameter for the size of the reservoir pool. If no size configuration is provided, the default size of 1 SHOULD be used.

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:

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.
  • Status: Experimental - The default aggregation cardinality limit to use, a function of instrument kind. If not configured, a default value of 2000 SHOULD be used.
  • Zero of more MetricProducers (optional) to collect metrics from in addition to metrics from the SDK.

The MetricReader.Collect method allows general-purpose MetricExporter instances to explicitly initiate collection, commonly used with pull-based metrics collection. A common implementation 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 with Cumulative aggregation temporality, this means converting Delta to Cumulative aggregation temporality. For asynchronous instruments 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 MetricReader selection of temporality as a function of instrument kind influences the persistence of metric data points across collections. For synchronous instruments with Cumulative aggregation temporality, MetricReader.Collect MUST receive data points exposed in previous collections regardless of whether new measurements have been recorded. For synchronous instruments with Delta aggregation temporality, MetricReader.Collect MUST only receive data points with measurements recorded since the previous collection. For asynchronous instruments with Delta or Cumulative aggregation temporality, MetricReader.Collect MUST only receive data points with measurements recorded since the previous collection. These rules apply to all metrics, not just those whose point kinds includes an aggregation temporality field.

The MetricReader selection of temporality as a function of instrument kind influences the starting timestamp (i.e. StartTimeUnixNano) of metrics data points received by MetricReader.Collect. For instruments with Cumulative aggregation temporality, successive data points received by successive calls to MetricReader.Collect MUST repeat the same starting timestamps (e.g. (T0, T1], (T0, T2], (T0, T3]). For instruments with Delta aggregation temporality, successive data points received by successive calls to MetricReader.Collect MUST advance the starting timestamp ( e.g. (T0, T1], (T1, T2], (T2, T3]). The ending timestamp (i.e. TimeUnixNano) MUST always be equal to time the metric data point took effect, which is equal to when MetricReader.Collect was invoked. These rules apply to all metrics, not just those whose point kinds includes an aggregation temporality field. See data model temporality for more details.

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 (Tn, Tn+1] to (Tn+1, Tn+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

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.

Collect SHOULD invoke Produce on registered MetricProducers. If the batch of metric points from Produce includes Resource information, Collect MAY replace the Resource from the MetricProducer with the Resource provided when constructing the MeterProvider instead.

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.

ForceFlush

This method provides a way for the periodic exporting MetricReader so it can do as much as it could to collect and send the metrics.

ForceFlush SHOULD collect metrics, call Export(batch) and ForceFlush() on the configured Push Metric Exporter.

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.

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 the 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: Stable

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.

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

When new OpenTelemetry integrations are added, the API is the preferred integration point. The MetricProducer is only meant for integrations that bridge pre-processed data.

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. If the batch of Metric points includes resource information, Produce SHOULD require a resource as a parameter. Produce does not have any other required parameters, however, OpenTelemetry SDK authors MAY choose to add required or optional 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, ForceFlush (for periodic exporting MetricReader) and Shutdown are safe to be called concurrently.

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