Table of Contents
Sampling is a mechanism to control the noise and overhead introduced by OpenTelemetry by reducing the number of samples of traces collected and sent to the backend.
Sampling may be implemented on different stages of a trace collection.
OpenTelemetry API defines a Sampler
interface that can be used at
instrumentation points by libraries to check the SamplingResult
early and
optimize the amount of telemetry that needs to be collected.
All other sampling algorithms may be implemented on SDK layer in exporters, or even out of process in Agent or Collector.
The OpenTelemetry API has two properties responsible for the data collection:
IsRecording
field of aSpan
. Iftrue
the currentSpan
records tracing events (attributes, events, status, etc.), otherwise all tracing events are dropped. Users can use this property to determine if expensive trace events can be avoided. Span Processors will receive all spans with this flag set. However, Span Exporter will not receive them unless theSampled
flag was set.Sampled
flag inTraceFlags
onSpanContext
. This flag is propagated via theSpanContext
to child Spans. For more details see the W3C Trace Context specification. This flag indicates that theSpan
has beensampled
and will be exported. Span Processor and Span Exporter will receive spans with theSampled
flag set for processing.
The flag combination SampledFlag == false
and IsRecording == true
means that the current Span
does record information, but most likely the child
Span
will not.
The flag combination SampledFlag == true
and IsRecording == false
could cause gaps in the distributed trace, and because of this OpenTelemetry API
MUST NOT allow this combination.
The SDK defines the two interfaces Sampler
and
Decision
as well as a set of built-in
samplers.
Sampler
interface allows to create custom samplers which will return a
sampling SamplingResult
based on information that is typically available just
before the Span
was created.
Returns the sampling Decision for a Span
to be created.
Required arguments:
SpanContext
of a parentSpan
. Typically extracted from the wire. Can benull
.TraceId
of theSpan
to be created. It can be different from theTraceId
in theSpanContext
. Typically in situations when theSpan
to be created starts a new Trace.SpanId
of theSpan
to be created.- Name of the
Span
to be created. SpanKind
- Initial set of
Attributes
for theSpan
being constructed - Collection of links that will be associated with the
Span
to be created. Typically useful for batch operations, see Links Between Spans.
Return value:
It produces an output called SamplingResult
which contains:
- A sampling
Decision
. One of the following enum values:NOT_RECORD
-IsRecording() == false
, span will not be recorded and all events and attributes will be dropped.RECORD
-IsRecording() == true
, butSampled
flag MUST NOT be set.RECORD_AND_SAMPLED
-IsRecording() == true
ANDSampled
flag` MUST be set.
- A set of span Attributes that will also be added to the
Span
.- The list of attributes returned by
SamplingResult
MUST be immutable. Caller may call this method any number of times and can safely cache the returned value.
- The list of attributes returned by
Returns the sampler name or short description with the configuration. This may
be displayed on debug pages or in the logs. Example:
"ProbabilitySampler{0.000100}"
.
Description MUST NOT change over time and caller can cache the returned value.
These are the default samplers implemented in the OpenTelemetry SDK:
- ALWAYS_ON
- This will be used as a default.
- Description MUST be
AlwaysOnSampler
.
- ALWAYS_OFF
- Description MUST be
AlwaysOffSampler
.
- Description MUST be
- ALWAYS_PARENT
Returns RECORD_AND_SAMPLED
ifSampledFlag
is set to true on parent SpanContext andNOT_RECORD
otherwise.- Description MUST be
AlwaysParentSampler
.
- Probability
- The default behavior should be to trust the parent
SampledFlag
. However there should be configuration to change this. - The default behavior is to apply the sampling probability only for Spans that are root spans (no parent) and Spans with remote parent. However there should be configuration to change this to "root spans only", or "all spans".
- Description MUST be
ProbabilitySampler{0.000100}
.
- The default behavior should be to trust the parent
TODO: Add details about how the probability sampler is implemented as a function
of the TraceID
.
New Tracer
instances are always created through a TracerProvider
(see
API). The name
and version
arguments
supplied to the TracerProvider
must be used to create a
Resource
instance which is stored on the created Tracer
.
All configuration objects (SDK specific) and extension points (span processors,
propagators) must be provided to the TracerProvider
. Tracer
instances must
not duplicate this data (unless for read-only access) to avoid that different
Tracer
instances of a TracerProvider
have different versions of these data.
The readable representations of all Span
instances created by a Tracer
must
provide a getLibraryResource
method that returns this Resource
information
held by the Tracer
.
Span processor is an interface which allows hooks for span start and end method
invocations. The span processors are invoked only when
IsRecording
is true.
Built-in span processors are responsible for batching and conversion of spans to exportable representation and passing batches to exporters.
Span processors can be registered directly on SDK TracerProvider
and they are
invoked in the same order as they were registered.
All Tracer
instances created by a TracerProvider
share the same span processors.
Changes to this collection reflect in all Tracer
instances.
Implementation-wise, this could mean that Tracer
instances have a reference to
their TracerProvider
and can access span processor objects only via this
reference.
Manipulation of the span processors collection must only happen on TracerProvider
instances. This means methods like addSpanProcessor
must be implemented on
TracerProvider
.
Each processor registered on TracerProvider
is a start of pipeline that consist
of span processor and optional exporter. SDK MUST allow to end each pipeline with
individual exporter.
SDK MUST allow users to implement and configure custom processors and decorate built-in processors for advanced scenarios such as tagging or filtering.
The following diagram shows SpanProcessor
's relationship to other components
in the SDK:
+-----+--------------+ +-------------------------+ +-------------------+
| | | | | | |
| | | | BatchExporterProcessor | | SpanExporter |
| | +---> SimpleExporterProcessor +---> (JaegerExporter) |
| | | | | | |
| SDK | Span.start() | +-------------------------+ +-------------------+
| | Span.end() |
| | | +---------------------+
| | | | |
| | +---> ZPagesProcessor |
| | | | |
+-----+--------------+ +---------------------+
OnStart
is called when a span is started. This method is called synchronously
on the thread that started the span, therefore it should not block or throw
exceptions.
Parameters:
Span
- a readable span object.
Returns: Void
OnEnd
is called when a span is ended. This method is called synchronously on
the execution thread, therefore it should not block or throw an exception.
Parameters:
Span
- a readable span object.
Returns: Void
Shuts down the processor. Called when SDK is shut down. This is an opportunity for processor to do any cleanup required.
Shutdown should be called only once for each Processor
instance. After the
call to shutdown subsequent calls to onStart
, onEnd
, or forceFlush
are not allowed.
Shutdown should not block indefinitely. Language library authors can decide if they want to make the shutdown timeout configurable.
Export all ended spans to the configured Exporter
that have not yet been exported.
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 Processor
exports the completed spans.
ForceFlush
should not block indefinitely. Language library authors can decide if they want to make the flush timeout configurable.
The standard OpenTelemetry SDK MUST implement both simple and batch processors, as described below. Other common processing scenarios should be first considered for implementation out-of-process in OpenTelemetry Collector
This is an implementation of SpanProcessor
which passes finished spans
and passes the export-friendly span data representation to the configured
SpanExporter
, as soon as they are finished.
Configurable parameters:
exporter
- the exporter where the spans are pushed.
This is an implementation of the SpanProcessor
which create batches of finished
spans and passes the export-friendly span data representations to the
configured SpanExporter
.
Configurable parameters:
exporter
- the exporter where the spans are pushed.maxQueueSize
- the maximum queue size. After the size is reached spans are dropped. The default value is2048
.scheduledDelayMillis
- the delay interval in milliseconds between two consecutive exports. The default value is5000
.exporterTimeoutMillis
- how long the export can run before it is cancelled. The default value is30000
.maxExportBatchSize
- the maximum batch size of every export. It must be smaller or equal tomaxQueueSize
. The default value is512
.
Span Exporter
defines the interface that protocol-specific exporters must
implement so that they can be plugged into OpenTelemetry SDK and support sending
of telemetry data.
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.
The exporter must support two functions: Export and Shutdown. In
strongly typed languages typically there will be 2 separate Exporter
interfaces, one that accepts spans (SpanExporter) and one that accepts metrics
(MetricsExporter).
Exports a batch of telemetry data. Protocol exporters that will implement this function are typically expected to serialize and transmit the data to the destination.
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 spans are being sent to.
Parameters:
batch - a batch of telemetry data. The exact data type of the batch is language
specific, typically it is a list of telemetry items, e.g. for spans in Java it
will be typically Collection<ExportableSpan>
.
Note that the data type for a span for illustration purposes here is written as an imaginary type ExportableSpan (similarly for metrics it would be e.g. ExportableMetrics). The actual data type must be specified by language library authors, it should be able to represent the span data that can be read by the exporter.
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.
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 Exporter
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). Language library authors can decide if they
want to make the shutdown timeout configurable.
Based on the generic interface definition laid out above library authors must define the exact interface for the particular language.
Authors are encouraged to use efficient data structures on the interface boundary that are well suited for fast serialization to wire formats by protocol exporters and minimize the pressure on memory managers. The latter typically requires understanding of how to optimize the rapidly-generated, short-lived telemetry data structures to make life easier for the memory manager of the specific language. General recommendation is to minimize the number of allocations and use allocation arenas where possible, thus avoiding explosion of allocation/deallocation/collection operations in the presence of high rate of telemetry data generation.
These are examples on what the Exporter
interface can look like in specific
languages. Examples are for illustration purposes only. Language library authors
are free to deviate from these provided that their design remain true to the
spirit of Exporter
concept.
type SpanExporter interface {
Export(batch []ExportableSpan) ExportResult
Shutdown()
}
type ExportResult struct {
Code ExportResultCode
WrappedError error
}
type ExportResultCode int
const (
Success ExportResultCode = iota
Failure
)
public interface SpanExporter {
public enum ResultCode {
Success, Failure
}
ResultCode export(Collection<ExportableSpan> batch);
void shutdown();
}