Skip to content

Latest commit

 

History

History
238 lines (158 loc) · 18 KB

File metadata and controls

238 lines (158 loc) · 18 KB

TraceState: Probability Sampling

Status: Development

Table of Contents

Overview

Sampling is an important lever to reduce the costs associated with collecting and processing telemetry data. It enables you to choose a representative set of items from an overall population.

There are two key aspects for sampling of tracing data. The first is that sampling decisions can be made independently for each span in a trace. The second is that sampling decisions can be made at multiple points in the telemetry pipeline. For example, the sampling decision for a span at span creation time could have been to keep that span, while the downstream sampling decision for the same span at a later stage (say in an external process in the data collection pipeline) could be to drop it.

For each of the above aspects, if we don't make consistent sampling decisions, we will end up with traces that are unusable and do not contain a coherent set of spans, because of the independent sampling decisions. Instead, we want sampling decisions to be made in a consistent manner so that we can effectively reason about a trace.

This specification describes a mechanism to achieve such consistent sampling decisions using a mechanism called Consistent Probability Sampling. To achieve this, it uses two key building blocks. The first is a common source of randomness (R) that is available to all participants, which includes a set of tracers and collectors. This can be either a custom value (called rv) or taken from the trailing 7 bytes of the TraceID. The second is a concept of a rejection threshold (T). This is derived directly from a participant's sampling rate. This proposal describes how these two values should be propagated and how participants should use them to make sampling decisions.

For more details about this specification, see OTEP 235.

Definitions

Sampling Probability

Sampling probability is the likelihood that a span will be kept. Each participant can choose a different sampling probability for each span. For example, if the sampling probability is 0.25, around 25% of the spans will be kept.

Sampling probability is valid in the range 2^-56 through 1. The value 56 appearing in this expression corresponds with 7 bytes of randomness (i.e., 56 bits) which are specified for W3C Trace Context Level 2 TraceIDs. Note that the zero value is not defined and that "never" sampling is not a form of probability sampling.

Consistent Sampling Decision

A consistent sampling decision means that a positive sampling decision made for a particular span with probability p1 necessarily implies a positive sampling decision for any span belonging to the same trace if it is made with probability p2 >= p1.

Rejection Threshold (T)

This is a 56-bit value directly derived from the sampling probability. One way to think about this is that this is the number of spans that would be dropped out of 2^56 considered spans. This is an alternative to the p value in the previous specification. The p value is limited to powers of two, while this supports a large range of values.

You can derive the rejection threshold from the sampling probability as follows:

Rejection_Threshold = (1 - Sampling_Probability) * 2^56.

For example, if the sampling probability is 100% (keep all spans), the rejection threshold is 0.

Similarly, if the sampling probability is 1% (drop 99% of spans), the rejection threshold with 5 digits of precision would be (1-0.01) * 2^56 = 4458562600304640 = 0xfd70a00000000.

We refer to this rejection threshold conceptually as T. We represent it using the OpenTelemetry TraceState key th, where the value is propagated and also stored with each span. In the example above, the th key has fd70a00000000 as the value.

See tracestate handling for details about encoding threshold values.

Random Value (R)

A common random value (that is known or propagated to all participants) is the main ingredient that enables consistent probability sampling. Each participant can compare this value (R) with their rejection threshold (T) to make a consistent sampling decision across an entire trace (or even across a group of traces).

This proposal supports two sources of randomness:

  • A custom source of randomness: This proposal allows for a random (or pseudo-random) 56-bit value. We refer to this as rv. This can be generated and propagated through the tracestate header and the tracestate attribute in each span.
  • Using TraceID as a source of randomness: This proposal introduces using the last 56 bits of the traceid as the source of randomness. This can be done if the root participant knows that the traceid has been generated in a random or pseudo-random manner.

See tracestate handling for details about encoding randomness values.

Consistent Sampling Decision Approach

Given the above building blocks, let's look at how a participant can make consistent sampling decisions. For this, two values MUST be present in the SpanContext:

  1. The common source of randomness: the 56-bit R value.
  2. The rejection threshold: the 56-bit T value.

If R >= T, keep the span, else drop the span.

T represents the maximum threshold that was applied in all previous consistent sampling stages. If the current sampling stage applies a greater threshold value than any stage before, it MUST update (increase) the threshold correspondingly.

Explanation

Sampler behavior for initializing and updating T and R values

There are two categories of samplers:

  • Head samplers: Implementations of Sampler, called by a Tracer during span creation.
  • Downstream samplers: Any component that, given an ended Span, decides whether to drop it or keep it (by forwarding it to the next component in the pipeline). This category is also known as "collection path samplers" or "sampling processors". Note that Tail samplers are a special class of downstream samplers that buffer spans of a trace and make a sampling decision for the trace as a whole using data from any span in the buffered trace.

This section defines the behavior for these two categories of samplers.

Head samplers

A head sampler is responsible for computing the rv and th values in a new span's initial TraceState. The main inputs to that computation include the parent span's trace state (if a parent span exists), the new span's trace ID, and possibly the trace flags (to know if the trace ID has been generated in a random manner).

When a span is sampled by in accordance with this specification, the output TraceState SHOULD be set to convey probability sampling:

  • The th key MUST be defined with a threshold value corresponding to the sampling probability the sampler used.
  • If trace randomness was derived from a TraceState rv value, the same rv value MUST be defined and equal to the incoming Context's rv value.

Downstream samplers

A downstream sampler, in contrast, may output a given ended Span with a modified trace state, complying with following rules:

  • If the chosen sampling probability is 1, the sampler MUST NOT modify any existing th, nor set any th.
  • Otherwise, the chosen sampling probability is in (0, 1). In this case the sampler MUST output the span with a th equal to max(input th, chosen th). In other words, th MUST NOT be decreased (as it is not possible to retroactively adjust an earlier stage's sampling probability), and it MUST be increased if a lower sampling probability was used. This case represents the common case where a downstream sampler is reducing span throughput in the system.

Migration to consistent probability samplers

The OpenTelemetry specification for TraceIdRatioBased samplers was not completed until after the SDK specification was declared stable, and the exact behavior of that sampler was left unspecified. The th and rv sub-keys defined in the OpenTelemetry TraceState now address this behavior specifically.

As the OpenTelemetry TraceIdRatioBased sampler changes definition, users must consider how to avoid incomplete traces due to inconsistent sampling during the transition between old and new logic.

The original TraceIdRatioBased sampler specification gave a workaround for the underspecified behavior, that it was safe to use for root spans: "It is recommended to use this sampler algorithm only for root spans (in combination with ParentBased) because different language SDKs or even different versions of the same language SDKs may produce inconsistent results for the same input."

To avoid inconsistency during this transition, users SHOULD follow this guidance until all TraceIdRatioBased samplers used in a system have been upgraded to the modern TraceIdRatioBased specification based on W3C Trace Context Level 2 randomness. After all TraceIdRatioBased samplers have been upgraded, it is safe to use TraceIdRatioBased sampler without also using the ParentBased sampler.

Algorithms

The th and rv values may be represented and manipulated in a variety of forms depending on the capabilities of the processor and needs of the implementation. As 56-bit values, they are compatible with byte arrays and 64-bit integers, and can also be manipulated with 64-bit floating point with a truly negligible loss of precision.

The following examples are in Golang and Python3. They are intended as examples only for clarity, and not as a suggested implementation.

Converting floating-point probability to threshold value

Threshold values are encoded with trailing zeros removed, which allows for variable precision. This can be accomplished by rounding, and there are several practical ways to do this with built-in string formatting libraries.

With up to 56 bits of precision available, implementations that use built-in floating point number support will be limited by the precision of the underlying number support. One way to encode thresholds uses the IEEE 754-2008-standard hexadecimal floating point representation as a simple solution. For example, in Golang,

// ProbabilityToThresholdWithPrecision assumes the probability value is in the range
// [0x1p-56, 1] and precision is in the range [1, 14].
func ProbabilityToThresholdWithPrecision(probability float64, precision int) string {
    if probability == 1 {
        // Special case
        return "0"
    }
    // Raise precision by the number of leading 0s or Fs
    _, expF := math.Frexp(probability)
    _, expR := math.Frexp(1 - probability)
    precision = min(14, max(precision+expF/-4, precision+expR/-4))

    // Change the probability to rejection probability, with range [0, 1),
    // translate rejection probability by +1, into range [1, 2).
    // Format the significand of this expression as hexadecimal floating point
    asHex := strconv.FormatFloat(2-probability, 'x', precision, 64)

    // Strip the leading "0x1.", use the requested number of digits.
    // Strip additional trailing zeros.
    digits := asHex[4 : 4+precision]
    return strings.TrimRight(digits, "0")
}

To translate directly from floating point probability into a 56-bit unsigned integer representation using math.Round() and shift operations, see the OpenTelemetry Collector-Contrib pkg/sampling package package. This package demonstrates how to directly calculate integer thresholds from probabilities.

OpenTelemetry SDKs are recommended to use 4 digits of precision by default. The following table shows values computed by the method above for 1-in-N probability sampling, with precision 3, 4, and 5.

1-in-N Input probability Threshold (precision 3, 4, 5) Actual probability (precision 3, 4, 5) Exact Adjusted Count (precision 3, 4, 5)
1 1 0
0
0
1
1
1
1
1
1
2 0.5 8
8
8
0.5
0.5
0.5
2
2
2
3 0.3333333333333333 aab
aaab
aaaab
0.333251953125
0.3333282470703125
0.33333301544189453
3.0007326007326007
3.00004577706569
3.0000028610256777
4 0.25 c
c
c
0.25
0.25
0.25
4
4
4
5 0.2 ccd
cccd
ccccd
0.199951171875
0.1999969482421875
0.19999980926513672
5.001221001221001
5.0000762951094835
5.0000047683761295
8 0.125 e
e
e
0.125
0.125
0.125
8
8
8
10 0.1 e66
e666
e6666
0.10009765625
0.100006103515625
0.10000038146972656
9.990243902439024
9.99938968568813
9.999961853172863
16 0.0625 f
f
f
0.0625
0.0625
0.0625
16
16
16
100 0.01 fd71
fd70a
fd70a4
0.0099945068359375
0.010000228881835938
0.009999990463256836
100.05496183206107
99.99771123402633
100.00009536752259
1000 0.001 ffbe7
ffbe77
ffbe76d
0.0010004043579101562
0.0009999871253967285
0.000999998301267624
999.5958055290753
1000.012874769029
1000.0016987352618
10000 0.0001 fff972
fff9724
fff97247
0.00010001659393310547
0.00010000169277191162
0.00010000006295740604
9998.340882002383
9999.830725674266
9999.99370426336
100000 0.00001 ffff584
ffff583a
ffff583a5
9.998679161071777e-06
1.00000761449337e-05
1.0000003385357559e-05
100013.21013412817
99999.238556461
99999.96614643588
1000000 0.000001 ffffef4
ffffef39
ffffef391
9.98377799987793e-07
1.00000761449337e-06
9.999930625781417e-07
1.0016248358208955e+06
999992.38556461
1.0000069374699865e+06

Converting integer threshold to a th-value

To convert a 56-bit integer threshold value to the t-value representation, emit it as a hexadecimal value (without a leading '0x'), optionally with trailing zeros omitted:

if tvalue == 0: 
  add_otel_trace_state('tv:0')
else:
  h = hex(tvalue).rstrip('0')
  # remove leading 0x
  add_otel_trace_state('tv:'+h[2:])

Testing randomness vs threshold

Given randomness and threshold as 64-bit integers, a sample should be taken if randomness is greater than or equal to the threshold.

shouldSample = (randomness >= threshold)

Converting threshold to a sampling probability

The sampling probability is a value from 0.0 to 1.0, which can be calculated using floating point by dividing by 2^56:

# embedded _ in numbers for clarity (permitted by Python3)
maxth = 0x100_0000_0000_0000  # 2^56
prob = float(maxth - threshold) / maxth

Converting threshold to an adjusted count (sampling rate)

The adjusted count indicates the approximate quantity of items from the population that this sample represents. It is equal to 1/probability.

maxth = 0x100_0000_0000_0000  # 2^56
adjCount = maxth / float(maxth - threshold)

Adjusted count is not defined for spans that were obtained via non-probabilistic sampling (a sampled span with no th value).