| Status | |
|---|---|
| Stability | beta: traces |
| Distributions | contrib, k8s |
| Issues |   |
| Code Owners | @jpkrohling |
The tail sampling processor samples traces based on a set of defined policies. All spans for a given trace MUST be received by the same collector instance for effective sampling decisions.
Before performing sampling, spans will be grouped by trace_id. Therefore, the tail sampling processor can be used directly without the need for the groupbytraceprocessor.
This processor must be placed in pipelines after any processors that rely on context, e.g. k8sattributes. It reassembles spans into new batches, causing them to lose their original context.
Please refer to config.go for the config spec.
The following configuration options are required:
policies(no default): Policies used to make a sampling decision
Multiple policies exist today and it is straight forward to add more. These include:
always_sample: Sample all traceslatency: Sample based on the duration of the trace. The duration is determined by looking at the earliest start time and latest end time, without taking into consideration what happened in between. Supplying no upper bound will result in a policy sampling anything greater thanthreshold_ms.numeric_attribute: Sample based on number attributes (resource and record)probabilistic: Sample a percentage of traces. Read a comparison with the Probabilistic Sampling Processor.status_code: Sample based upon the status code (OK,ERRORorUNSET)string_attribute: Sample based on string attributes (resource and record) value matches, both exact and regex value matches are supportedtrace_state: Sample based on TraceState value matchesrate_limiting: Sample based on ratespan_count: Sample based on the minimum and/or maximum number of spans, inclusive. If the sum of all spans in the trace is outside the range threshold, the trace will not be sampled.boolean_attribute: Sample based on boolean attribute (resource and record).ottl_condition: Sample based on given boolean OTTL condition (span and span event).and: Sample based on multiple policies, creates an AND policycomposite: Sample based on a combination of above samplers, with ordering and rate allocation per sampler. Rate allocation allocates certain percentages of spans per policy order. For example if we have set max_total_spans_per_second as 100 then we can set rate_allocation as follows- test-composite-policy-1 = 50 % of max_total_spans_per_second = 50 spans_per_second
- test-composite-policy-2 = 25 % of max_total_spans_per_second = 25 spans_per_second
- To ensure remaining capacity is filled use always_sample as one of the policies
The following configuration options can also be modified:
decision_wait(default = 30s): Wait time since the first span of a trace before making a sampling decisionnum_traces(default = 50000): Number of traces kept in memory.expected_new_traces_per_sec(default = 0): Expected number of new traces (helps in allocating data structures)decision_cache(default =sampled_cache_size: 0): Configures amount of trace IDs to be kept in an LRU cache, persisting the "keep" decisions for traces that may have already been released from memory. By default, the size is 0 and the cache is inactive. If using, configure this as much higher thannum_tracesso decisions for trace IDs are kept longer than the span data for the trace.
Each policy will result in a decision, and the processor will evaluate them to make a final decision:
- When there's an "inverted not sample" decision, the trace is not sampled;
- When there's a "sample" decision, the trace is sampled;
- When there's a "inverted sample" decision and no "not sample" decisions, the trace is sampled;
- In all other cases, the trace is NOT sampled
An "inverted" decision is the one made based on the "invert_match" attribute, such as the one from the string, numeric or boolean tag policy.
Examples:
processors:
tail_sampling:
decision_wait: 10s
num_traces: 100
expected_new_traces_per_sec: 10
decision_cache:
sampled_cache_size: 100000
policies:
[
{
name: test-policy-1,
type: always_sample
},
{
name: test-policy-2,
type: latency,
latency: {threshold_ms: 5000, upper_threshold_ms: 10000}
},
{
name: test-policy-3,
type: numeric_attribute,
numeric_attribute: {key: key1, min_value: 50, max_value: 100}
},
{
name: test-policy-4,
type: probabilistic,
probabilistic: {sampling_percentage: 10}
},
{
name: test-policy-5,
type: status_code,
status_code: {status_codes: [ERROR, UNSET]}
},
{
name: test-policy-6,
type: string_attribute,
string_attribute: {key: key2, values: [value1, value2]}
},
{
name: test-policy-7,
type: string_attribute,
string_attribute: {key: key2, values: [value1, val*], enabled_regex_matching: true, cache_max_size: 10}
},
{
name: test-policy-8,
type: rate_limiting,
rate_limiting: {spans_per_second: 35}
},
{
name: test-policy-9,
type: string_attribute,
string_attribute: {key: url.path, values: [\/health, \/metrics], enabled_regex_matching: true, invert_match: true}
},
{
name: test-policy-10,
type: span_count,
span_count: {min_spans: 2, max_spans: 20}
},
{
name: test-policy-11,
type: trace_state,
trace_state: { key: key3, values: [value1, value2] }
},
{
name: test-policy-12,
type: boolean_attribute,
boolean_attribute: {key: key4, value: true}
},
{
name: test-policy-13,
type: ottl_condition,
ottl_condition: {
error_mode: ignore,
span: [
"attributes[\"test_attr_key_1\"] == \"test_attr_val_1\"",
"attributes[\"test_attr_key_2\"] != \"test_attr_val_1\"",
],
spanevent: [
"name != \"test_span_event_name\"",
"attributes[\"test_event_attr_key_2\"] != \"test_event_attr_val_1\"",
]
}
},
{
name: and-policy-1,
type: and,
and: {
and_sub_policy:
[
{
name: test-and-policy-1,
type: numeric_attribute,
numeric_attribute: { key: key1, min_value: 50, max_value: 100 }
},
{
name: test-and-policy-2,
type: string_attribute,
string_attribute: { key: key2, values: [ value1, value2 ] }
},
]
}
},
{
name: composite-policy-1,
type: composite,
composite:
{
max_total_spans_per_second: 1000,
policy_order: [test-composite-policy-1, test-composite-policy-2, test-composite-policy-3],
composite_sub_policy:
[
{
name: test-composite-policy-1,
type: numeric_attribute,
numeric_attribute: {key: key1, min_value: 50, max_value: 100}
},
{
name: test-composite-policy-2,
type: string_attribute,
string_attribute: {key: key2, values: [value1, value2]}
},
{
name: test-composite-policy-3,
type: always_sample
}
],
rate_allocation:
[
{
policy: test-composite-policy-1,
percent: 50
},
{
policy: test-composite-policy-2,
percent: 25
}
]
}
},
]Refer to tail_sampling_config.yaml for detailed examples on using the processor.
Imagine that you wish to configure the processor to implement the following rules:
-
Rule 1: Not all teams are ready to move to tail sampling. Therefore, sample all traces that are not from the team
team_a. -
Rule 2: Sample only 0.1 percent of Readiness/liveness probes
-
Rule 3:
service-1has a noisy endpoint/v1/name/{id}. Sample only 1 percent of such traces. -
Rule 4: Other traces from
service-1should be sampled at 100 percent. -
Rule 5: Sample all traces if there is an error in any span in the trace.
-
Rule 6: Add an escape hatch. If there is an attribute called
app.force_samplein the span, then sample the trace at 100 percent. -
Rule 7: Force spans with
app.do_not_sampleset totrueto not be sampled, even if the result of the other rules yield a sampling decision.
Here is what the configuration would look like:
tail_sampling:
decision_wait: 10s
num_traces: 100
expected_new_traces_per_sec: 10
policies: [
{
# Rule 1: use always_sample policy for services that don't belong to team_a and are not ready to use tail sampling
name: backwards-compatibility-policy,
type: and,
and:
{
and_sub_policy:
[
{
name: services-using-tail_sampling-policy,
type: string_attribute,
string_attribute:
{
key: service.name,
values:
[
list,
of,
services,
using,
tail_sampling,
],
invert_match: true,
},
},
{ name: sample-all-policy, type: always_sample },
],
},
},
# BEGIN: policies for team_a
{
# Rule 2: low sampling for readiness/liveness probes
name: team_a-probe,
type: and,
and:
{
and_sub_policy:
[
{
# filter by service name
name: service-name-policy,
type: string_attribute,
string_attribute:
{
key: service.name,
values: [service-1, service-2, service-3],
},
},
{
# filter by route
name: route-live-ready-policy,
type: string_attribute,
string_attribute:
{
key: http.route,
values: [/live, /ready],
enabled_regex_matching: true,
},
},
{
# apply probabilistic sampling
name: probabilistic-policy,
type: probabilistic,
probabilistic: { sampling_percentage: 0.1 },
},
],
},
},
{
# Rule 3: low sampling for a noisy endpoint
name: team_a-noisy-endpoint-1,
type: and,
and:
{
and_sub_policy:
[
{
name: service-name-policy,
type: string_attribute,
string_attribute:
{ key: service.name, values: [service-1] },
},
{
# filter by route
name: route-name-policy,
type: string_attribute,
string_attribute:
{
key: http.route,
values: [/v1/name/.+],
enabled_regex_matching: true,
},
},
{
# apply probabilistic sampling
name: probabilistic-policy,
type: probabilistic,
probabilistic: { sampling_percentage: 1 },
},
],
},
},
{
# Rule 4: high sampling for other endpoints
name: team_a-service-1,
type: and,
and:
{
and_sub_policy:
[
{
name: service-name-policy,
type: string_attribute,
string_attribute:
{ key: service.name, values: [service-1] },
},
{
# invert match - apply to all routes except the ones specified
name: route-name-policy,
type: string_attribute,
string_attribute:
{
key: http.route,
values: [/v1/name/.+],
enabled_regex_matching: true,
invert_match: true,
},
},
{
# apply probabilistic sampling
name: probabilistic-policy,
type: probabilistic,
probabilistic: { sampling_percentage: 100 },
},
],
},
},
{
# Rule 5: always sample if there is an error
name: team_a-status-policy,
type: and,
and:
{
and_sub_policy:
[
{
name: service-name-policy,
type: string_attribute,
string_attribute:
{
key: service.name,
values:
[
list,
of,
services,
using,
tail_sampling,
],
},
},
{
name: trace-status-policy,
type: status_code,
status_code: { status_codes: [ERROR] },
},
],
},
},
{
# Rule 6:
# always sample if the force_sample attribute is set to true
name: team_a-force-sample,
type: boolean_attribute,
boolean_attribute: { key: app.force_sample, value: true },
},
{
# Rule 7:
# never sample if the do_not_sample attribute is set to true
name: team_a-do-not-sample,
type: boolean_attribute,
boolean_attribute: { key: app.do_not_sample, value: true, invert_match: true },
},
# END: policies for team_a
]This processor requires all spans for a given trace to be sent to the same collector instance for the correct sampling decision to be derived. When scaling the collector, you'll then need to ensure that all spans for the same trace are reaching the same collector. You can achieve this by having two layers of collectors in your infrastructure: one with the load balancing exporter, and one with the tail sampling processor.
While it's technically possible to have one layer of collectors with two pipelines on each instance, we recommend separating the layers in order to have better failure isolation.
Probabilistic Sampling Processor compared to the Tail Sampling Processor with the Probabilistic policy
The probabilistic sampling processor and the probabilistic tail sampling processor policy work very similar: based upon a configurable sampling percentage they will sample a fixed ratio of received traces. But depending on the overall processing pipeline you should prefer using one over the other.
As a rule of thumb, if you want to add probabilistic sampling and...
...you are not using the tail sampling processor already: use the probabilistic sampling processor. Running the probabilistic sampling processor is more efficient than the tail sampling processor. The probabilistic sampling policy makes decision based upon the trace ID, so waiting until more spans have arrived will not influence its decision.
...you are already using the tail sampling processor: add the probabilistic sampling policy. You are already incurring the cost of running the tail sampling processor, adding the probabilistic policy will be negligible. Additionally, using the policy within the tail sampling processor will ensure traces that are sampled by other policies will not be dropped.
Q. Why am I seeing high values for the error metric sampling_trace_dropped_too_early?
A. This is likely a load issue. If the collector is processing more traces in-memory than the num_traces configuration
option allows, some will have to be dropped before they can be sampled. Increasing the value of num_traces can
help resolve this error, at the expense of increased memory usage.
See documentation.md for the full list metrics available for this component and their descriptions.
A circular buffer is used to ensure the number of traces in-memory doesn't exceed num_traces. When a new trace arrives, the oldest trace is removed. This can cause a trace to be dropped before it's sampled. To reduce the chance of this happening, either increase num_traces or decrease decision_wait. Both of those options increase memory usage.
Number of Traces Dropped
otelcol_processor_tail_sampling_sampling_trace_dropped_too_early
Pre-emptively Preventing Dropped Traces
A trace is dropped without sampling if it's removed from the circular buffer before decision_wait.
To track how long traces remain in the buffer use:
otelcol_processor_tail_sampling_sampling_trace_removal_age
It may be useful to calculate latency percentiles like p1 and compare that value to decision_wait. Values close to decision_wait are at risk of being dropped if trace volume increases.
Slow Sampling Evaluation
otelcol_processor_tail_sampling_sampling_decision_timer_latency
This measures latency of sampling a batch of traces and passing sampled traces through the remainder of the collector pipeline. A latency exceeding 1 second can delay sampling decisions beyond decision_wait, increasing the chance of traces being dropped before sampling.
It's therefore recommended to consume this component's output with components that are fast or trigger asynchronous processing.
A span's arrival is considered "late" if it arrives after its trace's sampling decision is made. Late spans can cause different sampling decisions for different parts of the trace.
There are two scenarios for late arriving spans:
- Scenario 1: While the sampling decision of the trace remains in the circular buffer of
num_traceslength, the late spans inherit that decision. That means late spans do not influence the trace's sampling decision. - Scenario 2: (Default, no decision cache configured) After the sampling decision is removed from the buffer, it's as if this component has never seen the trace before: The late spans are buffered for
decision_waitseconds and then a new sampling decision is made. - Scenario 3: (Decision cache is configured) When a "keep" decision is made on a trace, the trace ID is cached. The component will remember which trace IDs it sampled even after it releases the span data from memory. Unless it has been evicted from the cache after some time, it will remember the same "keep trace" decision.
Occurrences of Scenario 1 where late spans are not sampled can be tracked with the below histogram metric.
otelcol_processor_tail_sampling_sampling_late_span_age
It may also be useful to:
- Calculate the percentage of spans arriving late with
otelcol_processor_tail_sampling_sampling_late_span_age{le="+Inf"} / otelcol_processor_tail_sampling_count_spans_sampled. Note thatcount_spans_sampledrequires enabling theprocessor.tailsamplingprocessor.metricstatcountspanssampledfeature gate. - Visualize lateness as a histogram to see how much it can be reduced by increasing
decision_wait.
Sampled Frequency
To track the percentage of traces that were actually sampled, use:
otelcol_processor_tail_sampling_global_count_traces_sampled{sampled="true"} /
otelcol_processor_tail_sampling_global_count_traces_sampled
Sampling Policy Decision Frequency
To see how often each policy votes to sample a trace, use:
sum (otelcol_processor_tail_sampling_count_traces_sampled{sampled="true"}) by (policy) /
sum (otelcol_processor_tail_sampling_count_traces_sampled) by (policy)
As a reminder, a policy voting to sample the trace does not guarantee sampling; an "inverted not" decision from another policy would still discard the trace.
sampling_policy_evaluation_error