OTLP profiles export foundation (experimental)

logfire/_internal/config.py

@@ -109,6 +109,9 @@
 from .integrations.executors import instrument_executors
 from .logs import ProxyLoggerProvider
 from .metrics import ProxyMeterProvider
+from .profiling.exporter import PROFILES_PATH, ProfilesExporter
+from .profiling.otlp import resource_from_attributes
+from .profiling.supervisor import ProfilingSupervisor
 from .scrubbing import NOOP_SCRUBBER, BaseScrubber, Scrubber, ScrubbingOptions
 from .server_response import ServerResponseCallback, install_logfire_response_hook
 from .stack_info import warn_at_user_stacklevel
@@ -438,6 +441,7 @@ def configure(
     variables: VariablesOptions | LocalVariablesOptions | None = None,
     distributed_tracing: bool | None = None,
     advanced: AdvancedOptions | None = None,
+    profiling: bool = False,
     **deprecated_kwargs: Unpack[DeprecatedKwargs],
 ) -> Logfire:
     """Configure the logfire SDK.
@@ -513,6 +517,9 @@ def configure(
             for more information.
             This setting always applies globally, and the last value set is used, including the default value.
         advanced: Advanced options primarily used for testing by Logfire developers.
+        profiling: **Experimental.** Set to `True` to continuously profile this process with the
+            Python 3.15+ sampling profiler and send CPU profiles to Logfire. A no-op with a
+            warning on older Pythons or unsupported platforms.
     """
     from .. import DEFAULT_LOGFIRE_INSTANCE, Logfire
 
@@ -646,6 +653,7 @@ def configure(
         variables=variables,
         distributed_tracing=distributed_tracing,
         advanced=advanced,
+        profiling=profiling,
     )
 
     if local:
@@ -731,6 +739,9 @@ class _LogfireConfigData:
     advanced: AdvancedOptions
     """Advanced options primarily used for testing by Logfire developers."""
 
+    profiling: bool
+    """Whether to continuously profile this process (experimental)."""
+
     def _load_configuration(
         self,
         # note that there are no defaults here so that the only place
@@ -756,6 +767,7 @@ def _load_configuration(
         variables: VariablesOptions | LocalVariablesOptions | None,
         distributed_tracing: bool | None,
         advanced: AdvancedOptions | None,
+        profiling: bool,
     ) -> None:
         """Merge the given parameters with the environment variables file configurations."""
         self.param_manager = param_manager = ParamManager.create(config_dir)
@@ -861,6 +873,8 @@ def _load_configuration(
             metrics = MetricsOptions()
         self.metrics = metrics
 
+        self.profiling = profiling
+
         if self.service_version is None:
             try:
                 self.service_version = get_git_revision_hash()
@@ -893,6 +907,7 @@ def __init__(
         code_source: CodeSource | None = None,
         distributed_tracing: bool | None = None,
         advanced: AdvancedOptions | None = None,
+        profiling: bool = False,
     ) -> None:
         """Create a new LogfireConfig.
 
@@ -923,6 +938,7 @@ def __init__(
             variables=variables,
             distributed_tracing=distributed_tracing,
             advanced=advanced,
+            profiling=profiling,
         )
         # initialize with no-ops so that we don't impact OTEL's global config just because logfire is installed
         # that is, we defer setting logfire as the otel global config until `configure` is called
@@ -935,6 +951,7 @@ def __init__(
         # This ensures that we only call OTEL's global set_tracer_provider once to avoid warnings.
         self._has_set_providers = False
         self._initialized = False
+        self._profiling_supervisor: ProfilingSupervisor | None = None
         self._lock = RLock()
 
     def configure(
@@ -959,6 +976,7 @@ def configure(
         variables: VariablesOptions | LocalVariablesOptions | None,
         distributed_tracing: bool | None,
         advanced: AdvancedOptions | None,
+        profiling: bool,
     ) -> None:
         with self._lock:
             self._initialized = False
@@ -983,6 +1001,7 @@ def configure(
                 variables,
                 distributed_tracing,
                 advanced,
+                profiling,
             )
             self.initialize()
 
@@ -995,6 +1014,10 @@ def _initialize(self) -> None:
         if self._initialized:  # pragma: no cover
             return
 
+        if self._profiling_supervisor is not None:  # pragma: no cover
+            self._profiling_supervisor.shutdown()
+            self._profiling_supervisor = None
+
         emscripten = platform_is_emscripten()
 
         with suppress_instrumentation():
@@ -1042,6 +1065,7 @@ def _initialize(self) -> None:
             # https://github.com/open-telemetry/semantic-conventions/blob/e44693245eef815071402b88c3a44a8f7f8f24c8/docs/resource/README.md#service-experimental
             # Both recommend generating a UUID.
             resource = Resource({'service.instance.id': uuid4().hex}).merge(resource)
+            profiles_exporter: ProfilesExporter | None = None
 
             head = self.sampling.head
             sampler: Sampler | None = None
@@ -1176,6 +1200,8 @@ def check_tokens():
                         headers = {'User-Agent': f'logfire/{VERSION}', 'Authorization': token}
                         session = OTLPExporterHttpSession()
                         install_logfire_response_hook(session, self.advanced.server_response_hook)
+                        if self.profiling and profiles_exporter is None:  # pragma: no cover
+                            profiles_exporter = ProfilesExporter(session, urljoin(base_url, PROFILES_PATH))
                         span_exporter = BodySizeCheckingOTLPSpanExporter(
                             endpoint=urljoin(base_url, '/v1/traces'),
                             session=session,
@@ -1400,6 +1426,14 @@ def fix_pid():  # pragma: no cover
 
             self._ensure_flush_after_aws_lambda()
 
+        if profiles_exporter is not None:  # pragma: no cover
+            self._profiling_supervisor = ProfilingSupervisor(
+                profiles_exporter,
+                resource=resource_from_attributes(resource.attributes),
+                scope_version=VERSION,
+            )
+            self._profiling_supervisor.start()
+
     def force_flush(self, timeout_millis: int = 30_000) -> bool:
         """Force flush all spans and metrics.
 

logfire/_internal/main.py

@@ -2491,6 +2491,10 @@ def remaining_ms() -> int:
         if not remaining:  # pragma: no cover
             return False
 
+        profiling_supervisor = self.config._profiling_supervisor  # pyright: ignore[reportPrivateUsage]
+        if profiling_supervisor is not None:
+            profiling_supervisor.shutdown(timeout=remaining_ms() / 1000.0)
+
         if flush:  # pragma: no branch
             self._tracer_provider.force_flush(remaining)
             remaining = remaining_ms()

logfire/_internal/profiling/__init__.py

@@ -0,0 +1,10 @@
+"""Experimental: ship Python sampling-profiler data to the OpenTelemetry profiles signal.
+
+This package converts the output of the Python 3.15 `profiling.sampling`
+profiler (Tachyon) into OTLP profiles and exports them over HTTP.
+
+Status: work in progress. Implemented here are the conversion (`collapsed` +
+`otlp`) and the HTTP `exporter`. Still to come: the supervisor that actually
+runs the profiler subprocess, wiring into `logfire.configure()`, and per-span
+correlation.
+"""

logfire/_internal/profiling/_proto/README.md

@@ -0,0 +1,48 @@
+# Vendored OTLP profiles protobuf bindings
+
+`profiles_pb2.py` / `profiles_service_pb2.py` (and their `.pyi` stubs) are
+**generated** code, vendored here rather than imported from the
+`opentelemetry-proto` PyPI package.
+
+## Why vendored
+
+The OpenTelemetry profiles signal is alpha (`v1development`). The
+`opentelemetry-proto` PyPI package (≤ 1.41.x) is generated from the
+`opentelemetry-proto` repo pinned at `v1.9.0`, which carries an **outdated**
+`Sample` message: its field numbers (`values`, `attribute_indices`,
+`link_index`) were renumbered in proto `v1.10.0`. A `Sample` serialized with
+the stale numbering is silently misread by current consumers (e.g. Grafana
+Pyroscope), so we cannot use the PyPI bindings.
+
+These files are generated from `opentelemetry-proto` **v1.10.0** instead.
+
+## When this directory can be deleted
+
+Once the `opentelemetry-proto` PyPI package ships bindings generated from
+proto ≥ `v1.10.0`, delete this directory and import from
+`opentelemetry.proto.profiles...` directly. The upstream pin bump is tracked
+in <https://github.com/open-telemetry/opentelemetry-python/pull/5223>.
+
+## Regeneration
+
+The proto `package` / file paths are renamed to `lf_otlp.*` so the descriptors
+do not collide with the (stale) copy registered by the installed
+`opentelemetry-proto` package — protobuf's descriptor pool rejects two files
+with the same name. This rename is purely about the descriptor registry; it
+does not affect the wire format (only field numbers do).
+
+To regenerate, fetch `profiles.proto` and `profiles_service.proto` from
+`opentelemetry-proto` at the desired tag, place them under an `lf_otlp/`
+directory with `opentelemetry.proto.{profiles.v1development,collector.profiles.v1development}`
+rewritten to `lf_otlp.{profiles,collector}`, then:
+
+```sh
+uvx --from grpcio-tools --with mypy-protobuf python -m grpc_tools.protoc \
+  --proto_path=<root> --python_out=<out> --mypy_out=<out> \
+  lf_otlp/profiles.proto lf_otlp/profiles_service.proto
+```
+
+Finally rewrite `from lf_otlp import profiles_pb2` to
+`from logfire._internal.profiling._proto import profiles_pb2` in the generated
+`profiles_service_pb2.{py,pyi}`. `common`/`resource` come from the installed
+`opentelemetry-proto` package (those signals are stable).

logfire/_internal/profiling/_proto/__init__.py

@@ -0,0 +1,11 @@
+"""Vendored OTLP profiles protobuf bindings.
+
+Generated from the canonical `open-telemetry/opentelemetry-proto` repo (tag
+`v1.10.0`, the `v1development` profiles signal). Vendored because the
+`opentelemetry-proto` PyPI package ships an outdated snapshot of this
+still-alpha proto: its `Sample` field numbers predate a renumbering and are
+wire-incompatible with current consumers (e.g. Grafana Pyroscope).
+
+See `README.md` in this directory for the regeneration recipe and the
+condition under which this directory can be deleted.
+"""