Strengthen requirements for aggregators

specification/sdk-metric.md

@@ -31,12 +31,12 @@ system-imposed requirements. A _dimensionality reduction_ maps input
 LabelSets with (potentially) a large number of labels into a smaller
 LabelSet containing only labels for an explicit set of label keys.
 Performing dimensionality reduction in an metrics export pipeline
-generally means merging Aggregators computed for original LabelSets
-into a single combined Aggregator for the reduced-dimension LabelSet.
+generally means merging aggregators computed for original LabelSets
+into a single combined aggregator for the reduced-dimension LabelSet.
 
 __Export record__: The _Export record_ is an exporter-independent
 in-memory representation combining the metric instrument, the LabelSet
-for export, and the associated (checkpointed) Aggregator containing
+for export, and the associated (checkpointed) aggregator containing
 its state. Metric instruments are described by a metric descriptor.
 
 __Metric descriptor__: A _metric descriptor_ is an in-memory
@@ -67,7 +67,7 @@ collection. Batcher and Exporter implementations are written with the
 assumption that collection is single-threaded, therefore the Meter
 implementation MUST prevent concurrent `Collect()` calls. During the
 collection pass, the Meter implementation checkpoints each active
-Aggregator and passes it to the Batcher for processing.
+aggregator and passes it to the Batcher for processing.
 
 This document does not specify how to coordinate synchronization
 between user-facing metric updates and metric collection activity,
@@ -80,7 +80,7 @@ updates.
 
 The Meter acts as a short-term store for aggregating metric updates
 within a collection period. The Meter implementation maintains
-Aggregators for active metric instruments according to the complete,
+aggregators for active metric instruments according to the complete,
 original LabelSet. This ensures a relatively simple code path for
 entering metric updates into the Meter implementation.
 
@@ -116,7 +116,7 @@ every record either: (1) has a current, un-released handle pinning it
 in memory, (2) has pending updates that have not been collected, (3)
 is a candidate for removing from memory. The Meter maintains a
 mapping from the pair (Instrument, LabelSet) to an active record.
-Each active record contains an Aggregator implementation, which is
+Each active record contains an aggregator implementation, which is
 responsible for incorporating a series of metric updates into the
 current state.
 
@@ -128,51 +128,124 @@ active records.
 ## Aggregator implementations
 
 The Aggregator interface supports combining multiple metric events
-into a single aggregated state. Different concrete Aggregator types
+into a single aggregated state. Different concrete aggregator types
 provide different functionality and levels of concurrent performance.
 
 Aggregators support `Update()`, `Checkpoint()`, and `Merge()`.
 `Update()` is called directly from the Meter in response to a metric
 event, and may be called concurrently. `Update()` is also passed the
 user's telemetry context, which allows is to access the current trace
-context and distributed correlations, honwever none of the built-in
+context and distributed correlations, however none of the built-in
 aggregators use this information.
 
 The `Checkpoint()` operation is called to atomically save a snapshot
 of the Aggregator, since `Checkpoint()` may be called concurrently
 with `Update()`. The `Merge()` operation supports dimensionality
-reduction by combining state from multiple Aggregators into a single
+reduction by combining state from multiple aggregators into a single
 Aggregator state. 
 
-The Metric SDK comes with six built-in Aggregator types, two of which
-are standard for use with counters and gauges.
-
-1. Counter: This aggregator maintains a Sum using only a single word of memory.
-1. Gauge: This aggregator maintains a pair containing the last value and its timestamp.
-
-Four aggregators are intended for use with [Measure metric instruments](api-metrics.md#measure).
-
-1. MinMaxSumCount: This aggregator computes the min, max, sum, and count using only four words of memory.
-1. Sketch: This aggregator computes an approximate data structure that can estimate quantiles. Example algorithms include GK-Sketch, Q-Digest, T-Digest, DDSketch, and HDR-Histogram. The choice of algorithm should be made based on available libraries in each language.
-1. Histogram: This aggregator computes a histogram with pre-determined boundaries. This may be used to estimate quantiles, but is generally intended for cases where a histogram will be exported directly.
-1. Exact: This aggregator computes an array of all values, supporting exact quantile computations in the exporter.
+The Metric SDK SHOULD include six built-in aggregator types. Two
+standard aggregators MUST be included that implement standard counter
+and gauge aggregations.
+
+1. Counter: This aggregator MUST maintain a Sum. In languages with
+support for atomic operations, the Counter aggregator SHOULD be
+implemented using only a single word of memory for the current state
+and a single word of memory for its checkpoint.
+1. Gauge: This aggregator MUST maintain the last value and its
+timestamp. In languages with support for atomic operations, this
+aggregator's update operation SHOULD be implemented by a single memory
+allocation--to store the value and timestamp--followed by an atomic
+pointer swap; if the gauge is defined as monotonic, the should use
+atomic compare-and-swap to ensure monotonicity.
+
+Aggregators for [Measure metric instruments](api-metrics.md#measure)
+are more challenging in nature than the Counter and Gauge aggregators,
+since their task is to aggregate a series of individual measurements.
+To perform this duty exactly requires storing the entire set of
+measurements, which may be cost-prohibitive. The common mechanisms
+for exporting recorded measurements from a Measure metric instrument
+are: as a series of raw measurements, as a summary of pre-determined
+quantiles, and as a histogram with pre-determined boundaries. A
+definition for _Quantile_ is given below.
+
+Four aggregators SHOULD be provided for use with Measure metric
+instruments that support the common mechanisms for exporting recorded
+measurements with a range of performance options.
+
+1. MinMaxSumCount: This aggregator is intended as an inexpensive
+alternative to the Sketch, Histogram, and Exact aggregators for
+Measure instruments. This aggregator MUST compute the min, max, sum,
+and count of recorded measurements. In languages with support for
+atomic operations, this aggregator's update operation SHOULD maintain
+its state using four independent atomic updates. In this case, the
+aggregator's update operation SHOULD NOT be atomic with respect to its
+checkpoint operation, implying that a checkpoint could witness an
+inconsistent state; that is intentional given the inexpensive nature
+of this aggregator.
+1. Sketch: This aggregator computes an approximate data structure that
+MUST estimate quantiles of the distribution of recorded measurements.
+Example algorithms that could be used to implement this aggregator are
+include: GK-Sketch, DDSketch, Q-Digest, T-Digest, and HDR-Histogram.
+The choice of algorithm should be made based on available libraries in
+each language, but implementations with well-defined error bounds
+SHOULD be preferred.
+1. Histogram: This aggregator MUST compute a histogram with
+pre-determined boundaries. This aggregator MAY support quantile
+estimation, but is generally intended for cases where a histogram will
+be exported directly and the exporter wants explicit control over
+histogram boundaries.
+1. Exact: This aggregator MUST store an array of all recorded
+measurements. This aggregator MUST support exact quantile
+computations and it MUST support exporting raw values in the order
+they were recorded, however it is not required to support both of
+these modes simultaneously (since computing quantiles requires sorting
+the measurements).
+
+### Quantile definition
+
+When exporting a summary of recorded measurements for a Measure metric
+instrument, it is common to report _quantiles_ of the distribution.
+When computing quantiles from an exact aggregation (i.e., the complete
+data set), the "nearest rank" definition of quantile SHOULD be used.
+The nearest-rank definition ensures that the resulting value belongs
+to the original data set. Interpolation is not used in this method.
+
+The definition for the nearest-rank quantile given here makes use of
+the _cumulative distribution function_, a standard concept from
+probability theory. Quantiles are parameterized by `q`, where `0 <= q
+<= 1`. The value for quantile `q` is the least element of the
+original data set at or above the point where the cumulative
+distribution function equals `q`.
+
+For example, taking a data set of five values `{10, 20, 30, 40, 50}`,
+the `q=0.5` quantile (i.e., the median) equals 30, which is precisely
+the point where the cumulative distribution function equals 0.5.
+
+With an even-sized data set, for example `{10, 20, 30, 40}`, the
+`q=0.5` quantile equals 30. In this case, the cumulative distribution
+function equals 0.5 halfway between 20 and 30 and the greater value is
+selected as the nearest rank.
+
+When using an approximate aggregator to compute estimated quantile
+values, the nearest-rank quantile definition does not apply.
 
 ## Batcher implementation
 
 The Batcher acts as the primary source of configuration for exporting
 metrics from the SDK. The two kinds of configuration are:
 
-1. Given a metric instrument, choose which concrete Aggregator type to apply for in-process aggregation.
+1. Given a metric instrument, choose which concrete aggregator type to apply for in-process aggregation.
 1. Given a metric instrument, choose which dimensions to export by (i.e., the "grouping" function).
 
-The first choice--which concrete Aggregator type to apply--is made
+The first choice--which concrete aggregator type to apply--is made
 whenever the Meter implementation encounters a new (Instrument,
-LabelSet) pair. Each concrete type of Aggregator will perform a
+LabelSet) pair. Each concrete type of aggregator will perform a
 different function. Aggregators for counter and gauge instruments are
-relatively straightforward, but many concrete Aggregators are possible
+relatively straightforward, but many concrete aggregators are possible
 for measure metric instruments. The Batcher has an opportunity to
 disable instruments at this point simply by returning a `nil`
-Aggregator.
+aggregator.
 
 The second choice--which dimensions to export by--affects how the
 batcher processes records emitted by the Meter implementation during
@@ -181,7 +254,7 @@ export record for each metric instrument with pending updates to the
 Batcher.
 
 During the collection pass, the Batcher receives a full set of
-checkpointed Aggregators corresponding to each (Instrument, LabelSet)
+checkpointed aggregators corresponding to each (Instrument, LabelSet)
 pair with an active record managed by the Meter implementation.
 According to its own configuration, the Batcher at this point
 determines which dimensions to aggregate for export; it computes a
@@ -236,7 +309,7 @@ The metric export pipeline specified here does not include explicit
 support for multiple export pipelines. In principle, any one of the
 interfaces here could be satisfied by a multiplexing implementation,
 but in practice, it will be costly to run multiple Batchers or
-Aggregators in parallel.
+aggregators in parallel.
 
 If multiple exporters are required, therefore, it is best if they can
 share a single Batcher configuration.
@@ -245,7 +318,7 @@ share a single Batcher configuration.
 
 The Meter implementation and some Batcher implementations are required
 to compute a unique key corresponding to a LabelSet, for the purposes
-of locating an Aggregator to use for metric updates. Where possible,
+of locating an aggregator to use for metric updates. Where possible,
 Exporters can avoid a duplicate computation by providing a
 LabelEncoder to the Meter implementation.