Sampling is a broad topic; here it refers to the independent decisions made at points in a distributed tracing system of whether to collect a span or not. Multiple sampling decisions can be made before a span is finally consumed. When sampling is to be performed at multiple points in the process, the only way to reason about it effectively is to make sure that the sampling decisions are consistent. In this context, consistency means that a positive sampling decision made for a particular span with probability p1 implies a positive sampling decision for any span belonging to the same trace, if it is made with probability p2 >= p1.
The existing, experimental specification for probability sampling using TraceState is limited to powers-of-two probabilities, and is designed to work without making assumptions about TraceID randomness.
This system can only achieve non-power-of-two sampling using interpolation between powers of two, which is unnecessarily restrictive.
In existing sampling systems, sampling probabilities like 1%, 10%, and 75% are common, and it should be possible to express these without interpolation.
There is also a need for consistent sampling in the collection path (outside of the head-sampling paths) and using inherent randomness in the traceID is a less-expensive solution than referencing a custom r-value
from the tracestate in every span.
This proposal introduces a new value with the key th
as an alternative to the p
value in the previous specification.
The p
value is limited to powers of two, while the th
value in this proposal supports a large range of values.
This proposal allows for the continued expression of randomness using r-value
as specified there using the key r
.
To distinguish the cases, this proposal uses the key rv
.
In the general case, in order to make consistent sampling decisions across the entire path of the trace, two values MUST be present in the SpanContext
:
- A random (or pseudo-random) 56-bit value, called
R
below. - A 56-bit rejection threshold (or just "threshold") as expressed in the TraceState, called
T
below.T
represents the maximum threshold that was applied in all previous consistent sampling stages. If the current sampling stage applies a greater-valued threshold than any stage before, it MUST update (increase) the threshold correspondingly.
One way to think about rejection threshold is that is the number of spans that would be discarded out of 2^56 considered spans. This means that spans where R >= T
will be sampled.
Here is an example involving three participants A
, B
, and C
:
A
-> B
-> C
where -> indicates a parent -> child relationship.
A
uses consistent probability sampling with a sampling probability of 0.25 (this corresponds to a rejection probability of .75).
B
uses consistent probability sampling with a sampling probability of 0.5.
C
uses a parent-based sampler.
When A
samples a span, its outgoing traceparent will have the 'sampled' flag SET and the 'th' in its outgoing tracestate will be set to 0xc0_0000_0000_0000
.
When A
does not sample a span, its outgoing traceparent will have the 'sampled' flag UNSET but the 'th' in its outgoing tracestate will still be set to 0xc0_0000_0000_0000
.
When B samples a span, its outgoing traceparent will have the 'sampled' flag SET and the 'th' in its outgoing tracestate will be set to 0x80_0000_0000_0000
.
C (being a parent based sampler) samples a span purely based on its parent (B in this case), it will use the sampled flag to make the decision. Its outgoing 'th' value will continue to reflect what it got from B (0x80_0000_0000_0000
), and this is useful to understand its adjusted count.
This design requires that as a given span progresses along its collection path, th
is non-decreasing (and, in particular, must be increased at stages that apply lower sampling probabilities).
It does not, however, restrict a span's initial th
in any way (e.g., relating it to that of its parent, if it has one).
It is acceptable for B to have a lesser initial th
than A has. It would not be ok if some later-stage sampler decreased A's th
.
The system has the following invariant:
(R >= T) = sampled flag
The sampling decision is propagated with the following algorithm:
- If the
th
key is not specified, this implies that non-probabilistic sampling may be taking place. - Else derive
T
by parsing theth
key as a hex value as described below. - If
T
is 0, Always Sample. - Compare the 56 bits of
T
with the 56 bits ofR
. IfT > R
, then do not sample.
The R
value MUST be derived as follows:
- If the key
rv
is present in the Tracestate header, thenR = rv
. - Else if the Random Trace ID Flag is
true
in the traceparent header, thenR
is the lowest-order 56 bits of the trace-id. - Else
R
MUST be generated as a random value in the range[0, (2**56)-1]
and added to the Tracestate header with keyrv
.
The preferred way to propagate the R
value is as the lowest 56 bits of the trace-id.
If these bits are in fact random, the random
trace-flag SHOULD be set as specified in the W3C trace context specification.
There are circumstances where trace-id randomness is inadequate (for example, sampling a group of traces together); in these cases, an rv
value is required.
The value of the rv
and th
keys MUST be expressed as up to 14 hexadecimal digits from the set [0-9a-f]
. For th
keys only, trailing zeros (but not leading zeros) may be omitted. rv
keys MUST always be exactly 14 hex digits.
Examples:
th
value is missing: non-probabalistic sampling may be taking place.th=4
-- equivalent toth=40000000000000
, which is a 25% rejection threshold, corresponding to a 75% sampling probability.th=c
-- equivalent toth=c0000000000000
, which is a rejection threshold of 75%, corresponding to a sampling probability of 25%.th=08
-- equivalent toth=08000000000000
, which is a rejection threshold of 3.125%, corresponding to a sampling probability of 96.875%.th=0
-- equivalent toth=00000000000000
, which is a 0% rejection threshold, which means Always Sample.
The T
value MUST be derived as follows:
- If the
th
key is not present in the Tracestate header, then non-probabalistic sampling may be in use. - Else the value corresponding to the
th
key should be interpreted as above.
Sampling Decisions MUST be propagated by setting the value of the th
key in the Tracestate header according to the above.
There are two categories of sampler:
- Head samplers: Implementations of
Sampler
, called by aTracer
during span creation. - Downstream samplers: Any component that, given an ended Span, decides whether to drop or forward ("sample") it on to the next component in the system. Also known as "collection-path samplers" or "sampling processors". Tail samplers are a special class of downstream samplers that buffer the spans in a trace and select a sampling probability for the trace as a whole using data from any span in the buffered trace.
This section defines behavior for each kind of sampler.
A head sampler is responsible for computing the rv
and th
values in a new span's initial TraceState
. Notable inputs to that computation include the parent span's trace state (if a parent span exists) and the new span's trace ID.
First, a consistent Sampler
decides which sampling probability to use. The sampler MAY select any value of T. If a valid SpanContext
is provided in the call to ShouldSample
(indicating that the span being created will be a child span),
- Choosing a T greater than the parent span's is expected to result in partial traces (the parent may be sampled but its child, the current span, dropped).
- Choosing a T less than or equal to the parent span is expected to result in complete traces (this is definition of consistent probability sampling).
For the output TraceState,
- The
th
key MUST be defined with a value corresponding to the sampling probability the sampler actually used. - The
rv
value, if present on the input TraceState, MUST be defined and equal to the parent span'srv
. Otherwise,rv
MUST be defined if and only if the effective R was generated during the decision, per the "derive R" algorithm given earlier.
TODO: For new spans, ShouldSample
doesn't currently have a way to know the new Span's TraceFlags
, so it can't determine whether the Random Trace ID Flag is set, and in turn can't execute the "derive R" algorithm. Maybe it should take TraceFlags
as an additional parameter, just like it takes TraceId
?
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 anyth
. - Otherwise, the chosen sampling probability is in
(0, 1)
. In this case the sampler MUST output the span with ath
equal tomax(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.
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 Python3. They are intended as examples only for clarity, and not as a suggested implementation.
To convert a t-value string to a 56-bit integer threshold, pad it on the right with 0s so that it is 14 digits in length, and then parse it as a hexadecimal value.
padded = (tvalue + "00000000000000")[:14]
threshold = int('0x' + padded, 16)
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:
h = hex(tvalue).rstrip('0')
# remove leading 0x
tv = 'tv='+h[2:]
Given rv and threshold as 64-bit integers, a sample should be taken if rv is greater than or equal to the threshold.
shouldSample = (rv >= threshold)
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
The adjusted count indicates the approximate quantity of items from the population that this sample represents. It is equal to 1/probability
. It is not defined for spans that were obtained via non-probabilistic sampling (a sampled span with no th
value).
This proposal is the result of long negotiations on the Sampling SIG over what is required and various alternative forms of expressing it. This issue exhaustively covers the various formats that were discussed and their pros and cons. This proposal is the result of that decision.
The existing specification for r-value
and p-value
attempted to solve this problem, but were limited to powers of 2, which is inadequate.
This specification leaves room for different implementation options. For example, comparing hex strings or converting them to numeric format are both viable alternatives for handling the threshold.
We also know that some implementations prefer to use a sampling probability (in the range from 0-1.0) or a sampling rate (1/probability); this design permits conversion to and from these formats without loss up to at least 6 decimal digits of precision.
This permits sampling systems to propagate consistent sampling information downstream where it can be compensated for. For example, this will enable the tail-sampling processor in the OTel Collector to propagate its sampling decisions to backends in a standard way. This permits backend systems to use the effective sampling probability in data presentations.