Documentation ¶
Overview ¶
Package sampler contains all the logic of the agent-side trace sampling
Currently implementation is based on the scoring of the "signature" of each trace Based on the score, we get a sample rate to apply to the given trace
Current score implementation is super-simple, it is a counter with polynomial decay per signature. We increment it for each incoming trace then we periodically divide the score by two every X seconds. Right after the division, the score is an approximation of the number of received signatures over X seconds. It is different from the scoring in the Agent.
Since the sampling can happen at different levels (client, agent, server) or depending on different rules, we have to track the sample rate applied at previous steps. This way, sampling twice at 50% can result in an effective 25% sampling. The rate is stored as a metric in the trace root.
Package sampler contains all the logic of the agent-side trace sampling ¶
Currently implementation is based on the scoring of the "signature" of each trace Based on the score, we get a sample rate to apply to the given trace
Current score implementation is super-simple, it is a counter with polynomial decay per signature. We increment it for each incoming trace then we periodically divide the score by two every X seconds. Right after the division, the score is an approximation of the number of received signatures over X seconds. It is different from the scoring in the Agent.
Since the sampling can happen at different levels (client, agent, server) or depending on different rules, we have to track the sample rate applied at previous steps. This way, sampling twice at 50% can result in an effective 25% sampling. The rate is stored as a metric in the trace root.
Package sampler contains all the logic of the agent-side trace sampling ¶
Currently implementation is based on the scoring of the "signature" of each trace Based on the score, we get a sample rate to apply to the given trace
Current score implementation is super-simple, it is a counter with polynomial decay per signature. We increment it for each incoming trace then we periodically divide the score by two every X seconds. Right after the division, the score is an approximation of the number of received signatures over X seconds. It is different from the scoring in the Agent.
Since the sampling can happen at different levels (client, agent, server) or depending on different rules, we have to track the sample rate applied at previous steps. This way, sampling twice at 50% can result in an effective 25% sampling. The rate is stored as a metric in the trace root.
Index ¶
- Constants
- func CalcPreSampleRate(maxUserAvg, currentUserAvg, currentRate float64) (float64, error)
- func GetTraceAppliedSampleRate(root *model.Span) float64
- func SampleByRate(traceID uint64, sampleRate float64) bool
- func SetTraceAppliedSampleRate(root *model.Span, sampleRate float64)
- type Backend
- type Engine
- type EngineType
- type InternalState
- type MemoryBackend
- func (b *MemoryBackend) CountSample()
- func (b *MemoryBackend) CountSignature(signature Signature)
- func (b *MemoryBackend) GetCardinality() int64
- func (b *MemoryBackend) GetSampledScore() float64
- func (b *MemoryBackend) GetSignatureScore(signature Signature) float64
- func (b *MemoryBackend) GetSignatureScores() map[Signature]float64
- func (b *MemoryBackend) GetTotalScore() float64
- func (b *MemoryBackend) GetUpperSampledScore() float64
- func (b *MemoryBackend) Run()
- func (b *MemoryBackend) Stop()
- type PreSampler
- func (ps *PreSampler) Rate() float64
- func (ps *PreSampler) RealRate() float64
- func (ps *PreSampler) Run()
- func (ps *PreSampler) Sample(req *http.Request) bool
- func (ps *PreSampler) SetError(err error)
- func (ps *PreSampler) SetRate(rate float64)
- func (ps *PreSampler) Stats() *PreSamplerStats
- func (ps *PreSampler) Stop()
- type PreSamplerStats
- type PriorityEngine
- type Sampler
- func (s *Sampler) AdjustScoring()
- func (s *Sampler) GetAllCountScores() map[Signature]float64
- func (s *Sampler) GetAllSignatureSampleRates() map[Signature]float64
- func (s *Sampler) GetCountScore(signature Signature) float64
- func (s *Sampler) GetDefaultCountScore() float64
- func (s *Sampler) GetDefaultSampleRate() float64
- func (s *Sampler) GetMaxTPSSampleRate() float64
- func (s *Sampler) GetSampleRate(trace model.Trace, root *model.Span, signature Signature) float64
- func (s *Sampler) GetSignatureSampleRate(signature Signature) float64
- func (s *Sampler) GetState() InternalState
- func (s *Sampler) Run()
- func (s *Sampler) RunAdjustScoring()
- func (s *Sampler) SetSignatureCoefficients(offset float64, slope float64)
- func (s *Sampler) Stop()
- func (s *Sampler) UpdateExtraRate(extraRate float64)
- func (s *Sampler) UpdateMaxTPS(maxTPS float64)
- type ScoreEngine
- type Signature
Constants ¶
const (
// SamplingPriorityKey is the key of the sampling priority value in the metrics map of the root span
SamplingPriorityKey = "_sampling_priority_v1"
)
const ( // TraceCountHeader is the header client implementation should fill // with the number of traces contained in the payload. TraceCountHeader = "X-Datadog-Trace-Count" )
Variables ¶
This section is empty.
Functions ¶
func CalcPreSampleRate ¶
CalcPreSampleRate gives the new sample rate to apply for a given max user CPU average. It takes the current sample rate and user CPU average as those parameters both have an influence on the result.
func GetTraceAppliedSampleRate ¶
GetTraceAppliedSampleRate gets the sample rate the sample rate applied earlier in the pipeline.
func SampleByRate ¶
SampleByRate tells if a trace (from its ID) with a given rate should be sampled Use Knuth multiplicative hashing to leverage imbalanced traceID generators
func SetTraceAppliedSampleRate ¶
SetTraceAppliedSampleRate sets the currently applied sample rate in the trace data to allow chained up sampling.
Types ¶
type Backend ¶
type Backend interface { // Run runs the blocking execution of the backend main loop. Run() // Stop stops the backend main loop. Stop() // CountSample counts that 1 trace is going through the sampler. CountSample() // CountSignature counts that 1 signature is going through the sampler. CountSignature(signature Signature) // GetTotalScore returns the TPS (Traces Per Second) of all traces ingested. GetTotalScore() float64 // GetSampledScore returns the TPS of all traces sampled. GetSampledScore() float64 // GetUpperSampledScore is similar to GetSampledScore, but with the upper approximation. GetUpperSampledScore() float64 // GetSignatureScore returns the TPS of traces ingested of a given signature. GetSignatureScore(signature Signature) float64 // GetSignatureScores returns the TPS of traces ingested for all signatures. GetSignatureScores() map[Signature]float64 // GetCardinality returns the number of different signatures seen. GetCardinality() int64 }
Backend stores and counts traces and signatures ingested by a sampler.
type Engine ¶
type Engine interface { // Run the sampler. Run() // Stop the sampler. Stop() // Sample a trace. Sample(trace model.Trace, root *model.Span, env string) bool // GetState returns information about the sampler. GetState() interface{} // GetType returns the type of the sampler. GetType() EngineType }
Engine is a common basic interface for sampler engines.
type EngineType ¶
type EngineType int
EngineType represents the type of a sampler engine.
const ( // NormalScoreEngineType is the type of the ScoreEngine sampling non-error traces. NormalScoreEngineType EngineType = iota // ErrorsScoreEngineType is the type of the ScoreEngine sampling error traces. ErrorsScoreEngineType // PriorityEngineType is type of the priority sampler engine type. PriorityEngineType )
type InternalState ¶
type InternalState struct { Offset float64 Slope float64 Cardinality int64 InTPS float64 OutTPS float64 MaxTPS float64 }
InternalState exposes all the main internal settings of the score sampler
type MemoryBackend ¶
type MemoryBackend struct {
// contains filtered or unexported fields
}
MemoryBackend storing any state required to run the sampling algorithms.
Current implementation is only based on counters with polynomial decay. Its bias with steady counts is 1 * decayFactor. The stored scores represent approximation of the real count values (with a countScaleFactor factor).
func NewMemoryBackend ¶
func NewMemoryBackend(decayPeriod time.Duration, decayFactor float64) *MemoryBackend
NewMemoryBackend returns an initialized Backend.
func (*MemoryBackend) CountSample ¶
func (b *MemoryBackend) CountSample()
CountSample counts a trace sampled by the sampler.
func (*MemoryBackend) CountSignature ¶
func (b *MemoryBackend) CountSignature(signature Signature)
CountSignature counts an incoming signature.
func (*MemoryBackend) GetCardinality ¶
func (b *MemoryBackend) GetCardinality() int64
GetCardinality returns the number of different signatures seen recently.
func (*MemoryBackend) GetSampledScore ¶
func (b *MemoryBackend) GetSampledScore() float64
GetSampledScore returns the global score of all sampled traces.
func (*MemoryBackend) GetSignatureScore ¶
func (b *MemoryBackend) GetSignatureScore(signature Signature) float64
GetSignatureScore returns the score of a signature. It is normalized to represent a number of signatures per second.
func (*MemoryBackend) GetSignatureScores ¶
func (b *MemoryBackend) GetSignatureScores() map[Signature]float64
GetSignatureScores returns the scores for all signatures. It is normalized to represent a number of signatures per second.
func (*MemoryBackend) GetTotalScore ¶
func (b *MemoryBackend) GetTotalScore() float64
GetTotalScore returns the global score of all sampled traces.
func (*MemoryBackend) GetUpperSampledScore ¶
func (b *MemoryBackend) GetUpperSampledScore() float64
GetUpperSampledScore returns a certain upper bound of the global count of all sampled traces.
func (*MemoryBackend) Run ¶
func (b *MemoryBackend) Run()
Run runs and block on the Sampler main loop.
type PreSampler ¶
type PreSampler struct {
// contains filtered or unexported fields
}
PreSampler tries to tell wether we should keep a payload, even before fully processing it. Its only clues are the unparsed payload and the HTTP headers. It should remain very light and fast.
func NewPreSampler ¶
func NewPreSampler(rate float64) *PreSampler
NewPreSampler returns an initialized presampler
func (*PreSampler) Rate ¶
func (ps *PreSampler) Rate() float64
Rate returns the current target pre-sample rate, thread-safe. The target pre-sample rate is the value set with SetRate, ideally this is the sample rate, but depending on what is received, the real rate might defer.
func (*PreSampler) RealRate ¶
func (ps *PreSampler) RealRate() float64
RealRate returns the current real pre-sample rate, thread-safe. This is the value obtained by counting what was kept and dropped.
func (*PreSampler) Sample ¶
func (ps *PreSampler) Sample(req *http.Request) bool
Sample tells wether a given request should be kept (true means: "yes, keep it"). Calling this alters the statistics, it affects the result of RealRate() so only call it once per payload.
func (*PreSampler) SetError ¶
func (ps *PreSampler) SetError(err error)
SetError set the pre-sample error, thread-safe.
func (*PreSampler) SetRate ¶
func (ps *PreSampler) SetRate(rate float64)
SetRate set the pre-sample rate, thread-safe.
func (*PreSampler) Stats ¶
func (ps *PreSampler) Stats() *PreSamplerStats
Stats returns a copy of the currrent pre-sampler stats.
type PreSamplerStats ¶
type PreSamplerStats struct { // Rate is the target pre-sampling rate. Rate float64 // Error is the last error got when trying to calc the pre-sampling rate. // Stored as a string as this is easier to marshal & publish in JSON. Error string // RecentPayloadsSeen is the number of payloads that passed by. RecentPayloadsSeen float64 // RecentTracesSeen is the number of traces that passed by. RecentTracesSeen float64 // RecentTracesDropped is the number of traces that were dropped. RecentTracesDropped float64 }
PreSamplerStats contains pre-sampler data. The public content might be interesting for statistics, logging.
func (*PreSamplerStats) RealRate ¶
func (stats *PreSamplerStats) RealRate() float64
RealRate calcuates the current real pre-sample rate from the stats data. If no data is available, returns the target rate.
type PriorityEngine ¶
type PriorityEngine struct { // Sampler is the underlying sampler used by this engine, sharing logic among various engines. Sampler *Sampler // contains filtered or unexported fields }
PriorityEngine is the main component of the sampling logic
func NewPriorityEngine ¶
func NewPriorityEngine(extraRate float64, maxTPS float64, rateByService *config.RateByService) *PriorityEngine
NewPriorityEngine returns an initialized Sampler
func (*PriorityEngine) GetState ¶
func (s *PriorityEngine) GetState() interface{}
GetState collects and return internal statistics and coefficients for indication purposes It returns an interface{}, as other samplers might return other informations.
func (*PriorityEngine) GetType ¶
func (s *PriorityEngine) GetType() EngineType
GetType return the type of the sampler engine
func (*PriorityEngine) Run ¶
func (s *PriorityEngine) Run()
Run runs and block on the Sampler main loop
type Sampler ¶
type Sampler struct { // Storage of the state of the sampler Backend Backend // contains filtered or unexported fields }
Sampler is the main component of the sampling logic
func (*Sampler) AdjustScoring ¶
func (s *Sampler) AdjustScoring()
AdjustScoring modifies sampler coefficients to fit better the `maxTPS` condition
func (*Sampler) GetAllCountScores ¶
GetAllCountScores scores all signatures based on their recent throughput The score value can be seeing as the sample rate if the count were the only factor Since other factors can intervene (such as extra global sampling), its value can be larger than 1
func (*Sampler) GetAllSignatureSampleRates ¶
GetAllSignatureSampleRates gives the sample rate to apply to all signatures. For now, only based on count score.
func (*Sampler) GetCountScore ¶
GetCountScore scores any signature based on its recent throughput The score value can be seeing as the sample rate if the count were the only factor Since other factors can intervene (such as extra global sampling), its value can be larger than 1
func (*Sampler) GetDefaultCountScore ¶
GetDefaultCountScore returns a default score when not knowing the signature for real. Since other factors can intervene (such as extra global sampling), its value can be larger than 1
func (*Sampler) GetDefaultSampleRate ¶
GetDefaultSampleRate gives the sample rate to apply to an unknown signature. For now, only based on count score.
func (*Sampler) GetMaxTPSSampleRate ¶
GetMaxTPSSampleRate returns an extra sample rate to apply if we are above maxTPS.
func (*Sampler) GetSampleRate ¶
GetSampleRate returns the sample rate to apply to a trace.
func (*Sampler) GetSignatureSampleRate ¶
GetSignatureSampleRate gives the sample rate to apply to any signature. For now, only based on count score.
func (*Sampler) GetState ¶
func (s *Sampler) GetState() InternalState
GetState collects and return internal statistics and coefficients for indication purposes
func (*Sampler) RunAdjustScoring ¶
func (s *Sampler) RunAdjustScoring()
RunAdjustScoring is the sampler feedback loop to adjust the scoring coefficients
func (*Sampler) SetSignatureCoefficients ¶
SetSignatureCoefficients updates the internal scoring coefficients used by the signature scoring
func (*Sampler) UpdateExtraRate ¶
UpdateExtraRate updates the extra sample rate
func (*Sampler) UpdateMaxTPS ¶
UpdateMaxTPS updates the max TPS limit
type ScoreEngine ¶
type ScoreEngine struct { // Sampler is the underlying sampler used by this engine, sharing logic among various engines. Sampler *Sampler // contains filtered or unexported fields }
ScoreEngine is the main component of the sampling logic
func NewErrorsEngine ¶
func NewErrorsEngine(extraRate float64, maxTPS float64) *ScoreEngine
NewErrorsEngine returns an initialized Sampler dedicate to errors. It behaves just like the the normal ScoreEngine except for its GetType method (useful for reporting).
func NewScoreEngine ¶
func NewScoreEngine(extraRate float64, maxTPS float64) *ScoreEngine
NewScoreEngine returns an initialized Sampler
func (*ScoreEngine) GetState ¶
func (s *ScoreEngine) GetState() interface{}
GetState collects and return internal statistics and coefficients for indication purposes It returns an interface{}, as other samplers might return other informations.
func (*ScoreEngine) GetType ¶
func (s *ScoreEngine) GetType() EngineType
GetType returns the type of the sampler