# Tail Sampling Processor | Status | | | ------------- |-----------| | Stability | [beta]: traces | | Distributions | [contrib], [k8s] | | Issues | [![Open issues](https://img.shields.io/github/issues-search/open-telemetry/opentelemetry-collector-contrib?query=is%3Aissue%20is%3Aopen%20label%3Aprocessor%2Ftailsampling%20&label=open&color=orange&logo=opentelemetry)](https://github.com/open-telemetry/opentelemetry-collector-contrib/issues?q=is%3Aopen+is%3Aissue+label%3Aprocessor%2Ftailsampling) [![Closed issues](https://img.shields.io/github/issues-search/open-telemetry/opentelemetry-collector-contrib?query=is%3Aissue%20is%3Aclosed%20label%3Aprocessor%2Ftailsampling%20&label=closed&color=blue&logo=opentelemetry)](https://github.com/open-telemetry/opentelemetry-collector-contrib/issues?q=is%3Aclosed+is%3Aissue+label%3Aprocessor%2Ftailsampling) | | [Code Owners](https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/main/CONTRIBUTING.md#becoming-a-code-owner) | [@jpkrohling](https://www.github.com/jpkrohling), [@portertech](https://www.github.com/portertech) | [beta]: https://github.com/open-telemetry/opentelemetry-collector/blob/main/docs/component-stability.md#beta [contrib]: https://github.com/open-telemetry/opentelemetry-collector-releases/tree/main/distributions/otelcol-contrib [k8s]: https://github.com/open-telemetry/opentelemetry-collector-releases/tree/main/distributions/otelcol-k8s 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`](https://github.com/open-telemetry/opentelemetry-collector-contrib/tree/main/processor/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](./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 traces - `latency`: 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 than `threshold_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](#probabilistic-sampling-processor-compared-to-the-tail-sampling-processor-with-the-probabilistic-policy). - `status_code`: Sample based upon the status code (`OK`, `ERROR` or `UNSET`) - `string_attribute`: Sample based on string attributes (resource and record) value matches, both exact and regex value matches are supported - `trace_state`: Sample based on [TraceState](https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/trace/api.md#tracestate) value matches - `rate_limiting`: Sample based on the rate of spans per second. - `span_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 policy - `composite`: 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 1. test-composite-policy-1 = 50 % of max_total_spans_per_second = 50 spans_per_second 2. test-composite-policy-2 = 25 % of max_total_spans_per_second = 25 spans_per_second 3. 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 decision - `num_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`: Options for configuring caches for sampling decisions. You may want to vary the size of these caches depending on how many "keep" vs "drop" decisions you expect from your policies. For example, you may allocate a larger `non_sampled_cache_size` if you expect most traces to be dropped. Additionally, if using, configure this as much greater than `num_traces` so decisions for trace IDs are kept longer than the span data for the trace. - `sampled_cache_size` (default = 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. - `non_sampled_cache_size` (default = 0) Configures amount of trace IDs to be kept in an LRU cache, persisting the "drop" decisions for traces that may have already been released from memory. By default, the size is 0 and the cache is inactive. 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: ```yaml processors: tail_sampling: decision_wait: 10s num_traces: 100 expected_new_traces_per_sec: 10 decision_cache: sampled_cache_size: 100_000 non_sampled_cache_size: 100_000 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} }, { 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](./testdata/tail_sampling_config.yaml) for detailed examples on using the processor. ## A Practical Example Imagine that you wish to configure the processor to implement the following rules: 1. **Rule 1:** Not all teams are ready to move to tail sampling. Therefore, sample all traces that are not from the team `team_a`. 1. **Rule 2:** Sample only 0.1 percent of Readiness/liveness probes 1. **Rule 3:** `service-1` has a noisy endpoint `/v1/name/{id}`. Sample only 1 percent of such traces. 1. **Rule 4:** Other traces from `service-1` should be sampled at 100 percent. 1. **Rule 5:** Sample all traces if there is an error in any span in the trace. 1. **Rule 6:** Add an escape hatch. If there is an attribute called `app.force_sample` in the span, then sample the trace at 100 percent. 1. **Rule 7:** Force spans with `app.do_not_sample` set to `true` to not be sampled, even if the result of the other rules yield a sampling decision. Here is what the configuration would look like: ```yaml 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 ] ``` ### Scaling collectors with the tail sampling processor 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][loadbalancing_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][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][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. [probabilistic_sampling_processor]: ../probabilisticsamplerprocessor [loadbalancing_exporter]: ../../exporter/loadbalancingexporter ## FAQ **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. ## Monitoring and Tuning See [documentation.md][documentation_md] for the full list metrics available for this component and their descriptions. ### Dropped Traces 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 ``` **Preemptively 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. ### Late-Arriving Spans 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_traces` length, 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_wait` seconds 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 that `count_spans_sampled` requires enabling the `processor.tailsamplingprocessor.metricstatcountspanssampled` feature gate. - Visualize lateness as a histogram to see how much it can be reduced by increasing `decision_wait`. ### Sampling Decision Frequency **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. ### Tracking sampling policy To better understand _which_ sampling policy made the decision to include a trace, you can enable tracking the policy responsible for sampling a trace via the `processor.tailsamplingprocessor.recordpolicy` feature gate. When this feature gate is set, this will add additional attributes on each sampled span: | Attribute | Description | Present? | |---------------------------------|---------------------------------------------------------------------------|----------------------------| | `tailsampling.policy` | Records the configured name of the policy that sampled a trace | Always | | `tailsampling.composite_policy` | Records the configured name of a composite subpolicy that sampled a trace | When composite policy used | ### Policy Evaluation Errors ``` sampling_policy_evaluation_error ``` [documentation_md]: ./documentation.md