fix(sampling): ensure agent based sampling is not reset after forking and on tracer.configure (#13560)

Builds on
7fbdc9f

- Fix: Avoids reinitializing the SpanAggregator on
`tracer.configure(...)` and when an application forks. Instead
`SpanAggregator.reset()` is called. This operation ensures global
configurations are re-applied, trace buffer can be reset, and trace
writer is recreated. This ensures agent based sampling rules are not
reset.

- Clean up
- Removes `writer` parameter from `SpanAggregator.__init__(...)` with
this change the intialization of the global writer is an implementation
detail of the SpanAggregator. There is no longer a need to supply the
`SpanAggregator` with a writer on the initialization of the global
tracer.
- Moves all implementation details of resetting the `SpanAggregator`
from `Tracer.configure(...)` and `Tracer._recreate(...)` to
`SpanAggregator.reset(...)`.
- Removes the initialization of the SpanAggregator from
`_default_span_processors_factory`. With this change the global tracer's
SpanAggregator is never re-created. It's only modified when
`tracer.configure(..)` is used.
- Rename `DatadogSampler._service_based_samplers` property to
`DatadogSampler._agent_based_sampler` to improve clarity. These sampling
rules are no longer supplied via environment variables or a programatic
api, they can only be set by the Datadog Agent.
- Splits `SpanAggregator.trace_proccessors` into two properties
`SpanAggregator.dd_proccessors` and `SpanAggregator.user_processors`.
`SpanAggregator.users_proccessors` is set after application startup via
`Tracer.configure(..)` while `SpanAggregator.dd_proccessors` is internal
to the ddtrace library and should only be set by ddtrace components.
This separation allows us to avoid recreating all trace processors when
`tracer.configure()` is called.
 - Adds a more descriptive release note to an unreleased fix.

## Checklist
- [x] PR author has checked that all the criteria below are met
- The PR description includes an overview of the change
- The PR description articulates the motivation for the change
- The change includes tests OR the PR description describes a testing
strategy
- The PR description notes risks associated with the change, if any
- Newly-added code is easy to change
- The change follows the [library release note
guidelines](https://ddtrace.readthedocs.io/en/stable/releasenotes.html)
- The change includes or references documentation updates if necessary
- Backport labels are set (if
[applicable](https://ddtrace.readthedocs.io/en/latest/contributing.html#backporting))

## Reviewer Checklist
- [x] Reviewer has checked that all the criteria below are met
- Title is accurate
- All changes are related to the pull request's stated goal
- Avoids breaking
[API](https://ddtrace.readthedocs.io/en/stable/versioning.html#interfaces)
changes
- Testing strategy adequately addresses listed risks
- Newly-added code is easy to change
- Release note makes sense to a user of the library
- If necessary, author has acknowledged and discussed the performance
implications of this PR as reported in the benchmarks PR comment
- Backport labels are set in a manner that is consistent with the
[release branch maintenance
policy](https://ddtrace.readthedocs.io/en/latest/contributing.html#backporting)

---------

Co-authored-by: Brett Langdon <[email protected]>
(cherry picked from commit ca79351)

Author: 2 people

ddtrace/_trace/processor/__init__.py

@@ -4,7 +4,6 @@
 from os import environ
 from threading import RLock
 from typing import Dict
-from typing import Iterable
 from typing import List
 from typing import Optional
 
@@ -128,7 +127,13 @@ class TraceSamplingProcessor(TraceProcessor):
       Agent even if the dropped trace is not (as is the case when trace stats computation is enabled).
     """
 
-    def __init__(self, compute_stats_enabled: bool, single_span_rules: List[SpanSamplingRule], apm_opt_out: bool):
+    def __init__(
+        self,
+        compute_stats_enabled: bool,
+        single_span_rules: List[SpanSamplingRule],
+        apm_opt_out: bool,
+        agent_based_samplers: Optional[dict] = None,
+    ):
         super(TraceSamplingProcessor, self).__init__()
         self._compute_stats_enabled = compute_stats_enabled
         self.single_span_rules = single_span_rules
@@ -137,9 +142,14 @@ def __init__(self, compute_stats_enabled: bool, single_span_rules: List[SpanSamp
             # If ASM is enabled but tracing is disabled,
             # we need to set the rate limiting to 1 trace per minute
             # for the backend to consider the service as alive.
-            self.sampler = DatadogSampler(rate_limit=1, rate_limit_window=60e9, rate_limit_always_on=True)
+            self.sampler = DatadogSampler(
+                rate_limit=1,
+                rate_limit_window=60e9,
+                rate_limit_always_on=True,
+                agent_based_samplers=agent_based_samplers,
+            )
         else:
-            self.sampler = DatadogSampler()
+            self.sampler = DatadogSampler(agent_based_samplers=agent_based_samplers)
 
     def process_trace(self, trace: List[Span]) -> Optional[List[Span]]:
         if trace:
@@ -260,8 +270,8 @@ def __init__(
         self,
         partial_flush_enabled: bool,
         partial_flush_min_spans: int,
-        trace_processors: Iterable[TraceProcessor],
-        writer: Optional[TraceWriter] = None,
+        dd_processors: Optional[List[TraceProcessor]] = None,
+        user_processors: Optional[List[TraceProcessor]] = None,
     ):
         # Set partial flushing
         self.partial_flush_enabled = partial_flush_enabled
@@ -271,12 +281,10 @@ def __init__(
             config._trace_compute_stats, get_span_sampling_rules(), asm_config._apm_opt_out
         )
         self.tags_processor = TraceTagsProcessor()
-        self.trace_processors = trace_processors
-        # Initialize writer
-        if writer is not None:
-            self.writer: TraceWriter = writer
-        elif SpanAggregator._use_log_writer():
-            self.writer = LogWriter()
+        self.dd_processors = dd_processors or []
+        self.user_processors = user_processors or []
+        if SpanAggregator._use_log_writer():
+            self.writer: TraceWriter = LogWriter()
         else:
             verify_url(agent_config.trace_agent_url)
             self.writer = AgentWriter(
@@ -307,7 +315,9 @@ def __repr__(self) -> str:
             f"{self.partial_flush_min_spans}, "
             f"{self.sampling_processor},"
             f"{self.tags_processor},"
-            f"{self.trace_processors}, "
+            f"{self.dd_processors}, "
+            f"{self.user_processors}, "
+            f"{self._span_metrics}, "
             f"{self.writer})"
         )
 
@@ -369,7 +379,9 @@ def on_span_finish(self, span: Span) -> None:
                     finished[0].set_metric("_dd.py.partial_flush", num_finished)
 
                 spans: Optional[List[Span]] = finished
-                for tp in chain(self.trace_processors, [self.sampling_processor, self.tags_processor]):
+                for tp in chain(
+                    self.dd_processors, self.user_processors, [self.sampling_processor, self.tags_processor]
+                ):
                     try:
                         if spans is None:
                             return
@@ -480,3 +492,58 @@ def _queue_span_count_metrics(self, metric_name: str, tag_name: str, min_count:
                     TELEMETRY_NAMESPACE.TRACERS, metric_name, count, tags=((tag_name, tag_value),)
                 )
             self._span_metrics[metric_name] = defaultdict(int)
+
+    def reset(
+        self,
+        user_processors: Optional[List[TraceProcessor]] = None,
+        compute_stats: Optional[bool] = None,
+        apm_opt_out: Optional[bool] = None,
+        appsec_enabled: Optional[bool] = None,
+        reset_buffer: bool = True,
+    ) -> None:
+        """
+        Resets the internal state of the SpanAggregator, including the writer, sampling processor,
+        user-defined processors, and optionally the trace buffer and span metrics.
+
+        This method is typically used after a process fork or during runtime reconfiguration.
+        Arguments that are None will not override existing values.
+        """
+        try:
+            # Stop the writer to ensure it is not running while we reconfigure it.
+            self.writer.stop()
+        except ServiceStatusError:
+            # Writers like AgentWriter may not start until the first trace is encoded.
+            # Stopping them before that will raise a ServiceStatusError.
+            pass
+
+        if isinstance(self.writer, AgentWriter) and appsec_enabled:
+            # Ensure AppSec metadata is encoded by setting the API version to v0.4.
+            self.writer._api_version = "v0.4"
+        # Re-create the writer to ensure it is consistent with updated configurations (ex: api_version)
+        self.writer = self.writer.recreate()
+
+        # Recreate the sampling processor using new or existing config values.
+        # If an argument is None, the current value is preserved.
+        if compute_stats is None:
+            compute_stats = self.sampling_processor._compute_stats_enabled
+        if apm_opt_out is None:
+            apm_opt_out = self.sampling_processor.apm_opt_out
+        self.sampling_processor = TraceSamplingProcessor(
+            compute_stats,
+            get_span_sampling_rules(),
+            apm_opt_out,
+            self.sampling_processor.sampler._agent_based_samplers,
+        )
+
+        # Update user processors if provided.
+        if user_processors is not None:
+            self.user_processors = user_processors
+
+        # Reset the trace buffer and span metrics.
+        # Useful when forking to prevent sending duplicate spans from parent and child processes.
+        if reset_buffer:
+            self._traces = defaultdict(lambda: _Trace())
+            self._span_metrics = {
+                "spans_created": defaultdict(int),
+                "spans_finished": defaultdict(int),
+            }

ddtrace/_trace/sampler.py

@@ -75,7 +75,7 @@ class DatadogSampler:
         "limiter",
         "rules",
         "_rate_limit_always_on",
-        "_by_service_samplers",
+        "_agent_based_samplers",
     )
     _default_key = "service:,env:"
 
@@ -85,6 +85,7 @@ def __init__(
         rate_limit: Optional[int] = None,
         rate_limit_window: float = 1e9,
         rate_limit_always_on: bool = False,
+        agent_based_samplers: Optional[Dict[str, RateSampler]] = None,
     ):
         """
         Constructor for DatadogSampler sampler
@@ -101,7 +102,7 @@ def __init__(
         else:
             self.rules: List[SamplingRule] = rules or []
         # Set Agent based samplers
-        self._by_service_samplers: Dict[str, RateSampler] = {}
+        self._agent_based_samplers = agent_based_samplers or {}
         # Set rate limiter
         self._rate_limit_always_on: bool = rate_limit_always_on
         if rate_limit is None:
@@ -119,10 +120,10 @@ def update_rate_by_service_sample_rates(self, rate_by_service: Dict[str, float])
         samplers: Dict[str, RateSampler] = {}
         for key, sample_rate in rate_by_service.items():
             samplers[key] = RateSampler(sample_rate)
-        self._by_service_samplers = samplers
+        self._agent_based_samplers = samplers
 
     def __str__(self):
-        rates = {key: sampler.sample_rate for key, sampler in self._by_service_samplers.items()}
+        rates = {key: sampler.sample_rate for key, sampler in self._agent_based_samplers.items()}
         return "{}(agent_rates={!r}, limiter={!r}, rules={!r}), rate_limit_always_on={!r}".format(
             self.__class__.__name__,
             rates,
@@ -181,11 +182,11 @@ def sample(self, span: Span) -> bool:
             sample_rate = matched_rule.sample_rate
         else:
             key = self._key(span.service, span.get_tag(ENV_KEY))
-            if key in self._by_service_samplers:
+            if key in self._agent_based_samplers:
                 # Agent service based sampling
                 agent_service_based = True
-                sampled = self._by_service_samplers[key].sample(span)
-                sample_rate = self._by_service_samplers[key].sample_rate
+                sampled = self._agent_based_samplers[key].sample(span)
+                sample_rate = self._agent_based_samplers[key].sample_rate
 
         if matched_rule or self._rate_limit_always_on:
             # Avoid rate limiting when trace sample rules and/or sample rates are NOT provided

ddtrace/_trace/tracer.py

@@ -46,12 +46,10 @@
 from ddtrace.internal.processor.endpoint_call_counter import EndpointCallCounterProcessor
 from ddtrace.internal.runtime import get_runtime_id
 from ddtrace.internal.schema.processor import BaseServiceProcessor
-from ddtrace.internal.service import ServiceStatusError
 from ddtrace.internal.utils import _get_metas_to_propagate
 from ddtrace.internal.utils.formats import format_trace_id
 from ddtrace.internal.writer import AgentWriter
 from ddtrace.internal.writer import HTTPWriter
-from ddtrace.internal.writer import TraceWriter
 from ddtrace.settings._config import config
 from ddtrace.settings.asm import config as asm_config
 from ddtrace.settings.peer_service import _ps_config
@@ -89,20 +87,9 @@ def _start_appsec_processor() -> Optional["AppSecSpanProcessor"]:
 
 
 def _default_span_processors_factory(
-    trace_filters: List[TraceProcessor],
-    writer: Optional[TraceWriter],
-    partial_flush_enabled: bool,
-    partial_flush_min_spans: int,
     profiling_span_processor: EndpointCallCounterProcessor,
-) -> Tuple[List[SpanProcessor], Optional["AppSecSpanProcessor"], SpanAggregator]:
+) -> Tuple[List[SpanProcessor], Optional["AppSecSpanProcessor"]]:
     """Construct the default list of span processors to use."""
-    trace_processors: List[TraceProcessor] = []
-    trace_processors += [
-        PeerServiceProcessor(_ps_config),
-        BaseServiceProcessor(),
-    ]
-    trace_processors += trace_filters
-
     span_processors: List[SpanProcessor] = []
     span_processors += [TopLevelSpanProcessor()]
 
@@ -140,14 +127,7 @@ def _default_span_processors_factory(
 
     span_processors.append(profiling_span_processor)
 
-    # These need to run after all the other processors
-    span_aggregagtor = SpanAggregator(
-        partial_flush_enabled=partial_flush_enabled,
-        partial_flush_min_spans=partial_flush_min_spans,
-        trace_processors=trace_processors,
-        writer=writer,
-    )
-    return span_processors, appsec_processor, span_aggregagtor
+    return span_processors, appsec_processor
 
 
 class Tracer(object):
@@ -181,8 +161,6 @@ def __init__(self) -> None:
                     "Initializing multiple Tracer instances is not supported. Use ``ddtrace.trace.tracer`` instead.",
                 )
 
-        self._user_trace_processors: List[TraceProcessor] = []
-
         # globally set tags
         self._tags = config.tags.copy()
 
@@ -199,12 +177,13 @@ def __init__(self) -> None:
             config._trace_compute_stats = False
         # Direct link to the appsec processor
         self._endpoint_call_counter_span_processor = EndpointCallCounterProcessor()
-        self._span_processors, self._appsec_processor, self._span_aggregator = _default_span_processors_factory(
-            self._user_trace_processors,
-            None,
-            config._partial_flush_enabled,
-            config._partial_flush_min_spans,
-            self._endpoint_call_counter_span_processor,
+        self._span_processors, self._appsec_processor = _default_span_processors_factory(
+            self._endpoint_call_counter_span_processor
+        )
+        self._span_aggregator = SpanAggregator(
+            partial_flush_enabled=config._partial_flush_enabled,
+            partial_flush_min_spans=config._partial_flush_min_spans,
+            dd_processors=[PeerServiceProcessor(_ps_config), BaseServiceProcessor()],
         )
         if config._data_streams_enabled:
             # Inline the import to avoid pulling in ddsketch or protobuf
@@ -389,13 +368,6 @@ def configure(
         if compute_stats_enabled is not None:
             config._trace_compute_stats = compute_stats_enabled
 
-        if isinstance(self._span_aggregator.writer, AgentWriter):
-            if appsec_enabled:
-                self._span_aggregator.writer._api_version = "v0.4"
-
-        if trace_processors:
-            self._user_trace_processors = trace_processors
-
         if any(
             x is not None
             for x in [
@@ -405,7 +377,9 @@ def configure(
                 iast_enabled,
             ]
         ):
-            self._recreate()
+            self._recreate(
+                trace_processors, compute_stats_enabled, asm_config._apm_opt_out, appsec_enabled, reset_buffer=False
+            )
 
         if context_provider is not None:
             self.context_provider = context_provider
@@ -433,31 +407,31 @@ def _generate_diagnostic_logs(self):
 
     def _child_after_fork(self):
         self._pid = getpid()
-        self._recreate()
+        self._recreate(reset_buffer=True)
         self._new_process = True
 
-    def _recreate(self):
-        """Re-initialize the tracer's processors and trace writer. This method should only be used in tests."""
+    def _recreate(
+        self,
+        trace_processors: Optional[List[TraceProcessor]] = None,
+        compute_stats_enabled: Optional[bool] = None,
+        apm_opt_out: Optional[bool] = None,
+        appsec_enabled: Optional[bool] = None,
+        reset_buffer: bool = True,
+    ) -> None:
+        """Re-initialize the tracer's processors and trace writer"""
         # Stop the writer.
         # This will stop the periodic thread in HTTPWriters, preventing memory leaks and unnecessary I/O.
-        try:
-            self._span_aggregator.writer.stop()
-        except ServiceStatusError:
-            # Some writers (ex: AgentWriter), start when the first trace chunk is encoded. Stopping
-            # the writer before that point will raise a ServiceStatusError.
-            pass
-        # Re-create the background writer thread
-        rules = self._span_aggregator.sampling_processor.sampler._by_service_samplers
-        self._span_aggregator.writer = self._span_aggregator.writer.recreate()
         self.enabled = config._tracing_enabled
-        self._span_processors, self._appsec_processor, self._span_aggregator = _default_span_processors_factory(
-            self._user_trace_processors,
-            self._span_aggregator.writer,
-            self._span_aggregator.partial_flush_enabled,
-            self._span_aggregator.partial_flush_min_spans,
+        self._span_aggregator.reset(
+            user_processors=trace_processors,
+            compute_stats=compute_stats_enabled,
+            apm_opt_out=apm_opt_out,
+            appsec_enabled=appsec_enabled,
+            reset_buffer=reset_buffer,
+        )
+        self._span_processors, self._appsec_processor = _default_span_processors_factory(
             self._endpoint_call_counter_span_processor,
         )
-        self._span_aggregator.sampling_processor.sampler._by_service_samplers = rules.copy()
 
     def _start_span_after_shutdown(
         self,

ddtrace/internal/ci_visibility/recorder.py

@@ -630,7 +630,7 @@ def disable(cls) -> None:
         log.debug("%s disabled", cls.__name__)
 
     def _start_service(self) -> None:
-        tracer_filters = self.tracer._user_trace_processors
+        tracer_filters = self.tracer._span_aggregator.user_processors
         if not any(isinstance(tracer_filter, TraceCiVisibilityFilter) for tracer_filter in tracer_filters):
             tracer_filters += [TraceCiVisibilityFilter(self._tags, self._service)]  # type: ignore[arg-type]
             self.tracer.configure(trace_processors=tracer_filters)

ddtrace/internal/writer/writer.py

@@ -553,6 +553,7 @@ def recreate(self) -> HTTPWriter:
             api_version=self._api_version,
             headers=self._headers,
             report_metrics=self._report_metrics,
+            response_callback=self._response_cb,
         )
         return new_instance
 

ddtrace/opentracer/tracer.py

@@ -104,7 +104,7 @@ def __init__(
         trace_processors = None
         if isinstance(self._config.get(keys.SETTINGS), dict) and self._config[keys.SETTINGS].get("FILTERS"):  # type: ignore[union-attr]
             trace_processors = self._config[keys.SETTINGS]["FILTERS"]  # type: ignore[index]
-            self._dd_tracer._user_trace_processors = trace_processors
+            self._dd_tracer._span_aggregator.user_processors = trace_processors
 
         if self._config[keys.ENABLED]:
             self._dd_tracer.enabled = self._config[keys.ENABLED]

releasenotes/notes/fix-agent-based-sampling-8877694a37053e51.yaml

@@ -0,0 +1,4 @@
+---
+fixes:
+  - |
+    tracing: Resolves a sampling issue where agent-based sampling rates were not correctly applied after a process forked or the tracer was reconfigured.