omnibase_infra 0.2.1__py3-none-any.whl

omnibase_infra/observability/hooks/hook_observability.py

@@ -0,0 +1,1223 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2025 OmniNode Team
+"""Pipeline hook for cross-cutting observability concerns.
+
+This module provides the HookObservability class, a pipeline hook that enables
+cross-cutting observability instrumentation for infrastructure components.
+The hook tracks operation timing, emits metrics, and maintains execution context
+across async boundaries.
+
+CRITICAL: Concurrency Safety via contextvars
+--------------------------------------------
+This implementation uses contextvars exclusively for all timing and operation
+state. This is a CRITICAL design decision to prevent concurrency bugs in async
+code. Each async task gets its own isolated context, preventing race conditions
+where multiple concurrent operations would corrupt shared timing state.
+
+Why NOT use instance variables:
+    # WRONG - Race condition in async code!
+    class BadHook:
+        def __init__(self):
+            self._start_time = 0.0  # Shared across all concurrent operations!
+
+        def before_operation(self, operation: str):
+            self._start_time = time.perf_counter()  # Overwrites previous!
+
+        def after_operation(self):
+            return time.perf_counter() - self._start_time  # Wrong value!
+
+Why contextvars ARE correct:
+    # CORRECT - Each async task has isolated state
+    _start_time: ContextVar[float | None] = ContextVar("start_time", default=None)
+
+    class GoodHook:
+        def before_operation(self, operation: str):
+            _start_time.set(time.perf_counter())  # Isolated per-task
+
+        def after_operation(self):
+            return time.perf_counter() - _start_time.get()  # Correct per-task
+
+Thread-Safety Guarantees
+------------------------
+This class is thread-safe with the following guarantees:
+
+1. **Timing State**: All timing and operation state is stored in contextvars,
+   which provide per-task isolation in async code and per-thread isolation in
+   threaded code. Concurrent operations NEVER share timing state.
+
+2. **Metrics Sink**: The metrics_sink is accessed via a read-only property
+   from multiple async tasks. The metrics_sink implementation MUST be
+   thread-safe. The built-in SinkMetricsPrometheus satisfies this requirement
+   via internal locking. If using a custom sink, ensure it is thread-safe.
+   The metrics_sink is set once during __init__ and exposed as read-only to
+   prevent accidental modification.
+
+3. **Class Constants**: _HIGH_CARDINALITY_KEYS is a frozenset (immutable),
+   ensuring thread-safe read access.
+
+4. **Instance Variables**: The only instance variable (__metrics_sink) is set
+   once during __init__ and exposed via a read-only property, ensuring
+   thread-safe reads and preventing modification after construction.
+
+5. **Module-Level Singleton**: The global singleton hook instance is protected
+   by a threading.Lock for thread-safe initialization across threads.
+
+Singleton Support
+-----------------
+This module provides optional singleton support via get_global_hook() and
+configure_global_hook(). The singleton pattern is useful when:
+
+- Multiple components need to share the same observability infrastructure
+- You want centralized metrics collection across the application
+- Resource efficiency is important
+
+Important: When a singleton already exists, calling configure_global_hook()
+with different configuration will log a warning and return the existing
+instance. Use clear_global_hook() to reset the singleton if reconfiguration
+is needed.
+
+Usage Example:
+    ```python
+    from omnibase_infra.observability.hooks import HookObservability
+    from omnibase_spi.protocols.observability import ProtocolHotPathMetricsSink
+
+    # Create hook with optional metrics sink
+    sink: ProtocolHotPathMetricsSink = get_metrics_sink()
+    hook = HookObservability(metrics_sink=sink)
+
+    # Use in handler execution
+    hook.before_operation("handler.execute", correlation_id="abc-123")
+    try:
+        result = await handler.execute(payload)
+        hook.record_success()
+    except Exception as e:
+        hook.record_failure(str(type(e).__name__))
+        raise
+    finally:
+        duration_ms = hook.after_operation()
+        logger.info(f"Operation took {duration_ms:.2f}ms")
+
+    # Or use the global singleton
+    from omnibase_infra.observability.hooks import get_global_hook
+    hook = get_global_hook()  # Returns singleton, creates if needed
+    ```
+
+See Also:
+    - ProtocolHotPathMetricsSink: Metrics collection interface
+    - correlation.py: Correlation ID context management pattern
+    - docs/patterns/observability_patterns.md: Observability guidelines
+"""
+
+from __future__ import annotations
+
+import logging
+import threading
+import time
+from contextvars import ContextVar, Token
+from typing import TYPE_CHECKING
+from uuid import UUID
+
+_logger = logging.getLogger(__name__)
+
+# =============================================================================
+# MODULE-LEVEL SINGLETON STATE
+# =============================================================================
+#
+# Thread-safe singleton support for global hook instance. Protected by
+# _global_hook_lock for safe initialization from multiple threads.
+#
+# =============================================================================
+
+# Forward reference resolved by `from __future__ import annotations`
+_global_hook_instance: HookObservability | None = None
+_global_hook_lock = threading.Lock()
+
+if TYPE_CHECKING:
+    from types import TracebackType
+
+    from omnibase_spi.protocols.observability import ProtocolHotPathMetricsSink
+
+# =============================================================================
+# CONTEXT VARIABLES FOR CONCURRENCY-SAFE OPERATION TRACKING
+# =============================================================================
+#
+# These ContextVars provide per-async-task isolation for operation state.
+# Each concurrent operation gets its own isolated copy of these values,
+# preventing race conditions in high-concurrency environments.
+#
+# DO NOT convert these to instance variables - that would break concurrency!
+# =============================================================================
+
+# Operation start time in perf_counter units (high-resolution monotonic clock)
+_start_time: ContextVar[float | None] = ContextVar("hook_start_time", default=None)
+
+# Current operation name (e.g., "handler.execute", "retry.attempt")
+_operation_name: ContextVar[str | None] = ContextVar(
+    "hook_operation_name", default=None
+)
+
+# Correlation ID for distributed tracing (propagated from request context)
+_correlation_id: ContextVar[str | None] = ContextVar(
+    "hook_correlation_id", default=None
+)
+
+# Additional operation labels for metrics (e.g., handler name, status)
+# Note: ContextVar doesn't support default_factory, so we use None and handle it
+_operation_labels: ContextVar[dict[str, str] | None] = ContextVar(
+    "hook_operation_labels", default=None
+)
+
+
+class HookObservability:
+    """Pipeline hook for cross-cutting observability instrumentation.
+
+    This hook provides timing, metrics, and context management for infrastructure
+    operations. It uses contextvars for all state to ensure concurrency safety
+    in async code paths.
+
+    Key Features:
+        - Concurrency-safe timing via contextvars (NOT instance variables)
+        - Metrics emission via ProtocolHotPathMetricsSink
+        - Operation context propagation across async boundaries
+        - Support for nested operation tracking via context manager
+        - High-cardinality label filtering to prevent metric explosion
+
+    Thread-Safety Guarantees:
+        This class is safe for concurrent use from multiple async tasks.
+        All timing and operation state is stored in contextvars, which provide
+        per-task isolation. The metrics sink (if provided) MUST be thread-safe.
+
+        Specific guarantees:
+        1. Timing state (start_time, operation_name, etc.) is isolated per-task
+        2. The metrics_sink is set once in __init__ and never modified
+        3. _HIGH_CARDINALITY_KEYS is immutable (frozenset)
+        4. No mutable shared state exists between concurrent operations
+
+    High-Cardinality Label Filtering:
+        Labels containing high-cardinality keys (correlation_id, request_id,
+        trace_id, span_id, session_id, user_id) are automatically filtered
+        from metrics to prevent cardinality explosion. These values remain
+        available via get_current_context() for logging and tracing.
+
+        When labels are filtered, a debug log is emitted to aid troubleshooting.
+        Metrics are NEVER dropped entirely - only high-cardinality labels are
+        removed from the label set.
+
+    Metrics Emitted:
+        - `operation_started_total`: Counter incremented when operation starts
+        - `operation_completed_total`: Counter incremented when operation completes
+        - `operation_failed_total`: Counter incremented on failure
+        - `operation_duration_seconds`: Histogram of operation durations
+        - `retry_attempt_total`: Counter for retry attempts
+        - `circuit_breaker_state_change_total`: Counter for circuit state changes
+
+    Attributes:
+        metrics_sink: Read-only property for the metrics sink. Returns None if
+            no sink was provided. The sink MUST be thread-safe for concurrent
+            access from multiple async tasks.
+
+    Example:
+        ```python
+        hook = HookObservability(metrics_sink=sink)
+
+        # Manual instrumentation
+        hook.before_operation("db.query", correlation_id="req-123")
+        try:
+            result = await db.execute(query)
+            hook.record_success()
+        except Exception:
+            hook.record_failure("DatabaseError")
+            raise
+        finally:
+            duration = hook.after_operation()
+
+        # Context manager for automatic timing
+        with hook.operation_context("http.request", correlation_id="req-456"):
+            response = await http_client.get(url)
+        ```
+    """
+
+    # Use __slots__ to prevent accidental attribute addition and improve memory
+    # _metrics_lock provides defense-in-depth thread safety for metrics operations
+    __slots__ = ("__metrics_sink", "_metrics_lock")
+
+    def __init__(
+        self,
+        metrics_sink: ProtocolHotPathMetricsSink | None = None,
+    ) -> None:
+        """Initialize the observability hook.
+
+        Args:
+            metrics_sink: Optional metrics sink for emitting observability data.
+                If None, the hook operates in no-op mode for metrics (timing
+                is still tracked). This allows the hook to be used even when
+                metrics infrastructure is not available.
+
+        Thread-Safety Requirements:
+            The metrics_sink (if provided) MUST be thread-safe for concurrent
+            access from multiple async tasks. The built-in SinkMetricsPrometheus
+            satisfies this requirement. If using a custom sink implementation,
+            ensure all methods (increment_counter, set_gauge, observe_histogram)
+            are thread-safe.
+
+        Note:
+            The metrics_sink is stored as a private instance variable and
+            exposed via a read-only property. This prevents accidental
+            modification after construction, ensuring thread-safe reads.
+            This is intentionally different from timing state which MUST be
+            in contextvars for per-task isolation.
+
+        Thread-Safety Implementation:
+            A threading.Lock (_metrics_lock) is used for defense-in-depth protection
+            of all metrics sink operations. While the metrics_sink is expected to be
+            thread-safe, this lock ensures atomic operation sequences and protects
+            against potential subtle thread-safety issues in sink implementations.
+        """
+        # Use name mangling (__) to prevent external modification
+        # Access via the metrics_sink property
+        self.__metrics_sink = metrics_sink
+        # Defense-in-depth lock for metrics operations
+        # Ensures atomic counter increments and metric emissions even if sink
+        # has subtle thread-safety issues
+        self._metrics_lock = threading.Lock()
+
+    @property
+    def metrics_sink(self) -> ProtocolHotPathMetricsSink | None:
+        """Read-only access to the metrics sink.
+
+        Returns:
+            The metrics sink provided during construction, or None if no
+            sink was provided.
+
+        Thread-Safety:
+            This property is thread-safe. The underlying value is set once
+            during __init__ and never modified.
+        """
+        return self.__metrics_sink
+
+    # =========================================================================
+    # CORE TIMING API
+    # =========================================================================
+
+    def before_operation(
+        self,
+        operation: str,
+        correlation_id: str | UUID | None = None,
+        labels: dict[str, str] | None = None,
+    ) -> None:
+        """Mark the start of an operation for timing.
+
+        Sets up the timing context for the current async task. This method
+        MUST be called before after_operation() to establish the start time.
+
+        Concurrency Safety:
+            Uses contextvars to store timing state, ensuring each async task
+            has its own isolated start time. Multiple concurrent operations
+            will not interfere with each other.
+
+        Args:
+            operation: Name of the operation being tracked. Should follow
+                a dotted naming convention (e.g., "handler.execute",
+                "db.query", "http.request"). This is used as the metric name.
+            correlation_id: Optional correlation ID for distributed tracing.
+                Can be a string or UUID. If UUID, it will be converted to string.
+            labels: Optional additional labels to attach to metrics. Keys and
+                values must be strings. Common labels: "handler", "status".
+
+        Side Effects:
+            - Sets _start_time contextvar to current perf_counter value
+            - Sets _operation_name contextvar to operation parameter
+            - Sets _correlation_id contextvar if provided
+            - Sets _operation_labels contextvar if provided
+            - Increments "operation_started_total" counter if metrics sink present
+
+        Example:
+            ```python
+            hook.before_operation(
+                "handler.process",
+                correlation_id="abc-123",
+                labels={"handler": "UserHandler"},
+            )
+            ```
+        """
+        # Store timing state in contextvars for concurrency safety
+        _start_time.set(time.perf_counter())
+        _operation_name.set(operation)
+
+        # Convert UUID to string if needed
+        if correlation_id is not None:
+            _correlation_id.set(str(correlation_id))
+        else:
+            _correlation_id.set(None)
+
+        # Store labels (or empty dict if none provided)
+        _operation_labels.set(labels.copy() if labels else {})
+
+        # Emit start metric if sink is available
+        # Lock ensures atomic counter increment across concurrent calls
+        if self.metrics_sink is not None:
+            metric_labels = self._build_metric_labels(operation)
+            with self._metrics_lock:
+                self.metrics_sink.increment_counter(
+                    name="operation_started_total",
+                    labels=metric_labels,
+                    increment=1,
+                )
+
+    def after_operation(self) -> float:
+        """Mark the end of an operation and calculate duration.
+
+        Calculates the elapsed time since before_operation() was called and
+        optionally emits the duration as a histogram observation.
+
+        Concurrency Safety:
+            Reads timing state from contextvars, which are isolated per async
+            task. The returned duration is specific to the current task's
+            operation timing.
+
+        Returns:
+            Duration in milliseconds since before_operation() was called.
+            Returns 0.0 if before_operation() was not called (start_time is None).
+
+        Side Effects:
+            - Observes "operation_duration_seconds" histogram if metrics sink present
+            - Clears _start_time contextvar (sets to None)
+            - Clears _operation_name contextvar (sets to None)
+            - Does NOT clear correlation_id (may be needed for error handling)
+
+        Example:
+            ```python
+            hook.before_operation("db.query")
+            result = await db.execute(query)
+            duration_ms = hook.after_operation()  # e.g., 42.5
+            logger.info(f"Query took {duration_ms:.2f}ms")
+            ```
+        """
+        start = _start_time.get()
+        operation = _operation_name.get()
+
+        # Handle case where before_operation was not called
+        if start is None:
+            return 0.0
+
+        # Calculate duration
+        end = time.perf_counter()
+        duration_seconds = end - start
+        duration_ms = duration_seconds * 1000.0
+
+        # Emit duration metric if sink is available
+        # Lock ensures atomic histogram observation across concurrent calls
+        if self.metrics_sink is not None and operation is not None:
+            metric_labels = self._build_metric_labels(operation)
+            with self._metrics_lock:
+                self.metrics_sink.observe_histogram(
+                    name="operation_duration_seconds",
+                    labels=metric_labels,
+                    value=duration_seconds,
+                )
+
+        # Clear timing state (but keep correlation_id for potential error handling)
+        _start_time.set(None)
+        _operation_name.set(None)
+        _operation_labels.set(None)
+
+        return duration_ms
+
+    def get_current_context(self) -> dict[str, str | None]:
+        """Get the current operation context from contextvars.
+
+        Returns the current operation context including operation name,
+        correlation ID, and any additional labels. Useful for logging
+        and debugging.
+
+        Concurrency Safety:
+            Reads from contextvars, returning context specific to the
+            current async task.
+
+        Returns:
+            Dictionary containing:
+                - "operation": Current operation name or None
+                - "correlation_id": Current correlation ID or None
+                - Plus any additional labels from _operation_labels
+
+        Example:
+            ```python
+            hook.before_operation("handler.process", correlation_id="abc-123")
+            ctx = hook.get_current_context()
+            # ctx = {"operation": "handler.process", "correlation_id": "abc-123"}
+            logger.info("Processing", extra=ctx)
+            ```
+        """
+        result: dict[str, str | None] = {
+            "operation": _operation_name.get(),
+            "correlation_id": _correlation_id.get(),
+        }
+
+        # Add any additional labels (they're already string -> string)
+        labels = _operation_labels.get()
+        if labels is not None:
+            for key, value in labels.items():
+                result[key] = value
+
+        return result
+
+    # =========================================================================
+    # SUCCESS/FAILURE TRACKING
+    # =========================================================================
+
+    def record_success(self, labels: dict[str, str] | None = None) -> None:
+        """Record a successful operation completion.
+
+        Increments the operation_completed_total counter with success status.
+        Should be called after the operation completes successfully, before
+        after_operation().
+
+        Args:
+            labels: Optional additional labels to merge with operation labels.
+
+        Side Effects:
+            - Increments "operation_completed_total" counter with status="success"
+
+        Example:
+            ```python
+            hook.before_operation("handler.process")
+            result = await handler.execute()
+            hook.record_success()
+            duration = hook.after_operation()
+            ```
+        """
+        if self.metrics_sink is None:
+            return
+
+        operation = _operation_name.get()
+        if operation is None:
+            return
+
+        metric_labels = self._build_metric_labels(operation, labels)
+        metric_labels["status"] = "success"
+
+        # Lock ensures atomic counter increment across concurrent calls
+        with self._metrics_lock:
+            self.metrics_sink.increment_counter(
+                name="operation_completed_total",
+                labels=metric_labels,
+                increment=1,
+            )
+
+    def record_failure(
+        self,
+        error_type: str,
+        labels: dict[str, str] | None = None,
+    ) -> None:
+        """Record a failed operation.
+
+        Increments the operation_failed_total counter with the error type.
+        Should be called when an operation fails, before after_operation().
+
+        Args:
+            error_type: Type/class name of the error that occurred.
+                Should be a stable identifier (e.g., "TimeoutError",
+                "DatabaseConnectionError"), not the error message.
+            labels: Optional additional labels to merge with operation labels.
+
+        Side Effects:
+            - Increments "operation_failed_total" counter with error_type label
+
+        Example:
+            ```python
+            hook.before_operation("db.query")
+            try:
+                result = await db.execute(query)
+                hook.record_success()
+            except DatabaseError as e:
+                hook.record_failure("DatabaseError")
+                raise
+            finally:
+                hook.after_operation()
+            ```
+        """
+        if self.metrics_sink is None:
+            return
+
+        operation = _operation_name.get()
+        if operation is None:
+            return
+
+        metric_labels = self._build_metric_labels(operation, labels)
+        metric_labels["status"] = "failure"
+        metric_labels["error_type"] = error_type
+
+        # Lock ensures atomic counter increment across concurrent calls
+        with self._metrics_lock:
+            self.metrics_sink.increment_counter(
+                name="operation_failed_total",
+                labels=metric_labels,
+                increment=1,
+            )
+
+    # =========================================================================
+    # SPECIALIZED TRACKING METHODS
+    # =========================================================================
+
+    def record_retry_attempt(
+        self,
+        attempt_number: int,
+        reason: str,
+        labels: dict[str, str] | None = None,
+    ) -> None:
+        """Record a retry attempt for an operation.
+
+        Tracks retry attempts with attempt number and reason. Useful for
+        monitoring retry behavior and identifying flaky operations.
+
+        Args:
+            attempt_number: The current attempt number (1-based). First attempt
+                is 1, first retry is 2, etc.
+            reason: Reason for the retry (e.g., "timeout", "connection_reset",
+                "rate_limited"). Should be a stable identifier.
+            labels: Optional additional labels to merge with operation labels.
+
+        Side Effects:
+            - Increments "retry_attempt_total" counter
+
+        Example:
+            ```python
+            for attempt in range(1, max_retries + 1):
+                try:
+                    result = await operation()
+                    break
+                except RetryableError as e:
+                    hook.record_retry_attempt(attempt, "transient_error")
+                    if attempt == max_retries:
+                        raise
+            ```
+        """
+        if self.metrics_sink is None:
+            return
+
+        operation = _operation_name.get() or "unknown"
+        metric_labels = self._build_metric_labels(operation, labels)
+        metric_labels["attempt"] = str(attempt_number)
+        metric_labels["reason"] = reason
+
+        # Lock ensures atomic counter increment across concurrent calls
+        with self._metrics_lock:
+            self.metrics_sink.increment_counter(
+                name="retry_attempt_total",
+                labels=metric_labels,
+                increment=1,
+            )
+
+    def record_circuit_breaker_state_change(
+        self,
+        service_name: str,
+        from_state: str,
+        to_state: str,
+        labels: dict[str, str] | None = None,
+    ) -> None:
+        """Record a circuit breaker state transition.
+
+        Tracks circuit breaker state changes for monitoring circuit health
+        and identifying unstable services.
+
+        Args:
+            service_name: Name of the service protected by the circuit breaker.
+            from_state: Previous circuit state (e.g., "CLOSED", "OPEN", "HALF_OPEN").
+            to_state: New circuit state after transition.
+            labels: Optional additional labels.
+
+        Side Effects:
+            - Increments "circuit_breaker_state_change_total" counter
+
+        Example:
+            ```python
+            # In circuit breaker implementation
+            hook.record_circuit_breaker_state_change(
+                service_name="database",
+                from_state="CLOSED",
+                to_state="OPEN",
+            )
+            ```
+        """
+        if self.metrics_sink is None:
+            return
+
+        metric_labels: dict[str, str] = {
+            "service": service_name,
+            "from_state": from_state,
+            "to_state": to_state,
+        }
+
+        # Merge additional labels
+        if labels:
+            for key, value in labels.items():
+                if key not in metric_labels:  # Don't overwrite required labels
+                    metric_labels[key] = value
+
+        # Lock ensures atomic counter increment across concurrent calls
+        with self._metrics_lock:
+            self.metrics_sink.increment_counter(
+                name="circuit_breaker_state_change_total",
+                labels=metric_labels,
+                increment=1,
+            )
+
+    def set_gauge(
+        self,
+        name: str,
+        value: float,
+        labels: dict[str, str] | None = None,
+    ) -> None:
+        """Set a gauge metric value.
+
+        Convenience method for setting gauge values through the hook.
+        Useful for tracking current state like queue depths or active connections.
+
+        Note:
+            Gauges can legitimately be negative (e.g., temperature, delta values).
+            For buffer/queue metrics that should never be negative, use
+            set_buffer_gauge() instead, which enforces non-negative values.
+
+        Args:
+            name: Metric name following Prometheus conventions.
+            value: Current gauge value. Can be negative for appropriate metrics.
+            labels: Optional labels for the metric.
+
+        Side Effects:
+            - Sets gauge metric via metrics sink
+
+        Example:
+            ```python
+            hook.set_gauge(
+                "active_handlers",
+                value=len(active_handlers),
+                labels={"handler_type": "http"},
+            )
+            ```
+        """
+        if self.metrics_sink is None:
+            return
+
+        # Lock ensures atomic gauge update across concurrent calls
+        with self._metrics_lock:
+            self.metrics_sink.set_gauge(
+                name=name,
+                labels=labels or {},
+                value=value,
+            )
+
+    def set_buffer_gauge(
+        self,
+        name: str,
+        value: float,
+        labels: dict[str, str] | None = None,
+    ) -> None:
+        """Set a gauge metric value for buffer/queue metrics (non-negative).
+
+        Similar to set_gauge(), but enforces non-negative values. Use this for
+        metrics that represent counts, sizes, or capacities that logically
+        cannot be negative (e.g., queue depth, buffer size, connection pool size).
+
+        The value is clamped to 0.0 if negative, preventing invalid metric values
+        that could occur from race conditions or calculation errors. A warning
+        is logged when clamping occurs to aid debugging.
+
+        Args:
+            name: Metric name following Prometheus conventions.
+            value: Current buffer/queue value. Clamped to 0.0 if negative.
+            labels: Optional labels for the metric.
+
+        Side Effects:
+            - Sets gauge metric via metrics sink with max(0.0, value)
+            - Logs warning if negative value is clamped
+
+        Example:
+            ```python
+            # Queue depth can't go negative even with concurrent updates
+            hook.set_buffer_gauge(
+                "message_queue_depth",
+                value=queue.qsize(),
+                labels={"queue_name": "events"},
+            )
+
+            # Safe even if calculation error produces negative
+            hook.set_buffer_gauge(
+                "buffer_available_slots",
+                value=total_slots - used_slots,  # Clamped to 0 if oversubscribed
+                labels={"buffer": "write"},
+            )
+            ```
+        """
+        if self.metrics_sink is None:
+            return
+
+        # Enforce non-negative values for buffer/count metrics
+        if value < 0.0:
+            _logger.warning(
+                "Buffer gauge received negative value; clamping to 0.0",
+                extra={
+                    "metric_name": name,
+                    "original_value": value,
+                    "labels": labels,
+                },
+            )
+            safe_value = 0.0
+        else:
+            safe_value = value
+
+        # Lock ensures atomic gauge update across concurrent calls
+        with self._metrics_lock:
+            self.metrics_sink.set_gauge(
+                name=name,
+                labels=labels or {},
+                value=safe_value,
+            )
+
+    # =========================================================================
+    # CONTEXT MANAGER SUPPORT
+    # =========================================================================
+
+    def operation_context(
+        self,
+        operation: str,
+        correlation_id: str | UUID | None = None,
+        labels: dict[str, str] | None = None,
+    ) -> OperationScope:
+        """Create a context manager for operation timing.
+
+        Returns a context manager that automatically calls before_operation()
+        on entry and after_operation() on exit. This is the recommended way
+        to instrument operations as it ensures proper cleanup even on exceptions.
+
+        Args:
+            operation: Name of the operation to track.
+            correlation_id: Optional correlation ID for tracing.
+            labels: Optional additional labels for metrics.
+
+        Returns:
+            A context manager that yields the duration in milliseconds on exit.
+
+        Example:
+            ```python
+            # Basic usage
+            with hook.operation_context("handler.process") as ctx:
+                result = await handler.execute()
+            print(f"Operation took {ctx.duration_ms:.2f}ms")
+
+            # With correlation ID
+            with hook.operation_context(
+                "db.query",
+                correlation_id=request_id,
+                labels={"table": "users"},
+            ):
+                rows = await db.fetch_all(query)
+            ```
+        """
+        return OperationScope(
+            hook=self,
+            operation=operation,
+            correlation_id=correlation_id,
+            labels=labels,
+        )
+
+    # =========================================================================
+    # INTERNAL HELPERS
+    # =========================================================================
+
+    # High-cardinality label keys that must be excluded from metrics.
+    # These would cause metrics cardinality explosion and be dropped by
+    # ModelMetricsPolicy when on_violation is WARN_AND_DROP or DROP_SILENT.
+    _HIGH_CARDINALITY_KEYS: frozenset[str] = frozenset(
+        {
+            "correlation_id",
+            "request_id",
+            "trace_id",
+            "span_id",
+            "session_id",
+            "user_id",
+        }
+    )
+
+    def _build_metric_labels(
+        self,
+        operation: str,
+        extra_labels: dict[str, str] | None = None,
+    ) -> dict[str, str]:
+        """Build the complete label set for a metric.
+
+        Combines operation name, stored operation labels, and any extra labels
+        into a single label dictionary. High-cardinality keys are automatically
+        filtered out to prevent metrics from being dropped by the policy.
+
+        CRITICAL - Metric Recording Guarantee:
+            This method NEVER drops or suppresses metric recording. When called,
+            the metric WILL be recorded with the returned labels. The only
+            filtering that occurs is removal of high-cardinality label KEYS
+            from the label dictionary - the metric itself is ALWAYS recorded.
+
+            Example with correlation_id:
+                - Input: operation="db.query", labels={"correlation_id": "abc-123"}
+                - Output: {"operation": "db.query"}  # correlation_id filtered
+                - Metric: RECORDED with {"operation": "db.query"}
+
+            The correlation_id and other high-cardinality values remain available
+            via the _correlation_id contextvar for structured logging and
+            distributed tracing - they are just excluded from Prometheus labels
+            to prevent cardinality explosion.
+
+        Why Filter High-Cardinality Labels:
+            High-cardinality values (correlation_id, request_id, trace_id, etc.)
+            are unique per request. Including them in Prometheus labels would:
+            1. Create millions of unique time series (cardinality explosion)
+            2. Cause metrics to be dropped by ModelMetricsPolicy (WARN_AND_DROP)
+            3. Overwhelm Prometheus storage and query performance
+
+            By filtering these keys early, we ensure metrics are ALWAYS recorded
+            with stable, low-cardinality labels.
+
+        Args:
+            operation: Operation name to include. Always present in output.
+            extra_labels: Optional additional labels to merge.
+
+        Returns:
+            Complete label dictionary for metric emission, with high-cardinality
+            keys filtered out. GUARANTEED to contain at least {"operation": operation}.
+            Never returns empty dict. Never returns None.
+        """
+        # INVARIANT: labels always contains at least {"operation": operation}
+        # This ensures metrics are NEVER dropped due to empty labels
+        labels: dict[str, str] = {"operation": operation}
+        filtered_keys: list[str] = []
+
+        # Merge stored operation labels, filtering out high-cardinality keys
+        # Note: stored_labels dict is a copy made in before_operation(), safe to iterate
+        stored_labels = _operation_labels.get()
+        if stored_labels is not None:
+            for key, value in stored_labels.items():
+                if key in self._HIGH_CARDINALITY_KEYS:
+                    filtered_keys.append(key)
+                else:
+                    labels[key] = value
+
+        # Merge extra labels (overrides stored if same key), filtering high-cardinality
+        if extra_labels:
+            for key, value in extra_labels.items():
+                if key in self._HIGH_CARDINALITY_KEYS:
+                    # Only add to filtered_keys if not already tracked
+                    if key not in filtered_keys:
+                        filtered_keys.append(key)
+                else:
+                    labels[key] = value
+
+        # Log when high-cardinality keys are filtered (debug level)
+        # IMPORTANT: This is informational logging only - the metric IS being recorded
+        # We only remove specific label KEYS, NOT the entire metric
+        # The log includes correlation_id for tracing even though it's filtered from labels
+        if filtered_keys:
+            # Include correlation_id in log for tracing purposes
+            correlation_id = _correlation_id.get()
+            _logger.debug(
+                "Removed high-cardinality keys from metric labels - "
+                "metric WILL be recorded with %d remaining labels (keys removed: %s)",
+                len(labels),
+                filtered_keys,
+                extra={
+                    "operation": operation,
+                    "filtered_keys": filtered_keys,
+                    "remaining_labels": list(labels.keys()),
+                    "correlation_id": correlation_id,  # Available for log correlation
+                },
+            )
+
+        # CRITICAL: Defense-in-depth guarantee that metrics are NEVER dropped.
+        # The invariant at line 889 ensures "operation" is always present, but this
+        # explicit check provides runtime safety if the invariant is ever violated.
+        # This protects against data loss from unexpected edge cases.
+        if "operation" not in labels:
+            _logger.error(
+                "BUG: operation key missing from labels after filtering; "
+                "restoring to prevent metric data loss",
+                extra={"operation": operation, "labels_keys": list(labels.keys())},
+            )
+            labels["operation"] = operation
+
+        # GUARANTEE: This method ALWAYS returns a non-empty dict containing at least
+        # {"operation": operation}. Metrics are NEVER dropped - only high-cardinality
+        # label keys (correlation_id, request_id, etc.) are removed from the label set.
+        # The metric itself is ALWAYS recorded with the remaining labels.
+        return labels
+
+
+class OperationScope:
+    """Context manager for scoped operation timing.
+
+    This internal class provides context manager support for HookObservability.
+    It automatically calls before_operation() on entry and after_operation()
+    on exit, ensuring proper cleanup even when exceptions occur.
+
+    Attributes:
+        duration_ms: Duration of the operation in milliseconds, available
+            after the context exits. Will be 0.0 if accessed before exit.
+
+    Note:
+        This class stores tokens for contextvar restoration, enabling proper
+        nesting of operation contexts. Each context saves and restores the
+        previous contextvar values.
+    """
+
+    def __init__(
+        self,
+        hook: HookObservability,
+        operation: str,
+        correlation_id: str | UUID | None = None,
+        labels: dict[str, str] | None = None,
+    ) -> None:
+        """Initialize the context manager.
+
+        Args:
+            hook: The HookObservability instance to use.
+            operation: Operation name to track.
+            correlation_id: Optional correlation ID.
+            labels: Optional additional labels.
+        """
+        self._hook = hook
+        self._operation = operation
+        self._correlation_id = correlation_id
+        self._labels = labels
+        self.duration_ms: float = 0.0
+
+        # Tokens for restoring previous contextvar values (for nesting support)
+        self._start_time_token: Token[float | None] | None = None
+        self._operation_name_token: Token[str | None] | None = None
+        self._correlation_id_token: Token[str | None] | None = None
+        self._labels_token: Token[dict[str, str] | None] | None = None
+
+    def __enter__(self) -> OperationScope:
+        """Enter the operation context.
+
+        Saves current contextvar values and calls before_operation().
+
+        Returns:
+            Self, for accessing duration_ms after exit.
+        """
+        # Save current values for restoration on exit (nesting support)
+        self._start_time_token = _start_time.set(_start_time.get())
+        self._operation_name_token = _operation_name.set(_operation_name.get())
+        self._correlation_id_token = _correlation_id.set(_correlation_id.get())
+        current_labels = _operation_labels.get()
+        # NOTE: Shallow copy is sufficient here because:
+        # 1. Labels dict is typed as dict[str, str] (string keys and values)
+        # 2. Strings are immutable in Python, so no aliasing issues can occur
+        # 3. We only need isolation of the dict structure, not deep cloning of values
+        self._labels_token = _operation_labels.set(
+            current_labels.copy() if current_labels is not None else None
+        )
+
+        # Now start the new operation
+        self._hook.before_operation(
+            operation=self._operation,
+            correlation_id=self._correlation_id,
+            labels=self._labels,
+        )
+
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
+        """Exit the operation context.
+
+        Calls after_operation() and restores previous contextvar values.
+        Records success or failure based on whether an exception occurred.
+
+        Concurrency Safety:
+            Uses try/finally to ensure contextvar tokens are ALWAYS restored,
+            even if record_failure/record_success/after_operation raise exceptions.
+            This prevents contextvar state leakage in error scenarios.
+
+        Args:
+            exc_type: Exception type if an exception was raised.
+            exc_val: Exception value if an exception was raised.
+            exc_tb: Traceback if an exception was raised.
+        """
+        try:
+            # Record success or failure before timing
+            if exc_type is not None:
+                self._hook.record_failure(exc_type.__name__)
+            else:
+                self._hook.record_success()
+
+            # Get duration
+            self.duration_ms = self._hook.after_operation()
+        finally:
+            # CRITICAL: Always restore previous contextvar values (for nesting support)
+            # This must happen even if the above code raises an exception to prevent
+            # contextvar state leakage.
+            if self._start_time_token is not None:
+                _start_time.reset(self._start_time_token)
+            if self._operation_name_token is not None:
+                _operation_name.reset(self._operation_name_token)
+            if self._correlation_id_token is not None:
+                _correlation_id.reset(self._correlation_id_token)
+            if self._labels_token is not None:
+                _operation_labels.reset(self._labels_token)
+
+
+# =============================================================================
+# MODULE-LEVEL SINGLETON FUNCTIONS
+# =============================================================================
+#
+# These functions provide optional singleton access to a global HookObservability
+# instance. The singleton is thread-safe and lazily initialized.
+#
+# =============================================================================
+
+
+def get_global_hook(
+    metrics_sink: ProtocolHotPathMetricsSink | None = None,
+) -> HookObservability:
+    """Get or create the global singleton HookObservability instance.
+
+    Returns the cached singleton hook if one exists, otherwise creates a new
+    one with the provided configuration and caches it. This is the recommended
+    way to access a shared observability hook across the application.
+
+    Thread Safety:
+        This function is thread-safe. Multiple concurrent calls will receive
+        the same singleton instance. A threading.Lock ensures only one thread
+        creates the singleton.
+
+    Note:
+        The metrics_sink parameter is only used on first call when the singleton
+        is created. Subsequent calls ignore this parameter and return the
+        existing instance. A warning is logged when configuration is provided
+        but ignored due to an existing singleton.
+
+        To reconfigure the singleton, call clear_global_hook() first.
+
+    Args:
+        metrics_sink: Optional metrics sink for the hook. Only used if no
+            singleton exists yet. Subsequent calls ignore this parameter.
+
+    Returns:
+        The global singleton HookObservability instance.
+
+    Example:
+        ```python
+        # First call creates the singleton with the provided sink
+        hook1 = get_global_hook(metrics_sink=my_sink)
+
+        # Subsequent calls return the same instance
+        hook2 = get_global_hook()  # metrics_sink parameter ignored
+        assert hook1 is hook2
+
+        # Reconfigure if needed
+        clear_global_hook()
+        hook3 = get_global_hook(metrics_sink=new_sink)  # New singleton
+        ```
+    """
+    global _global_hook_instance  # noqa: PLW0603 - Standard singleton pattern
+
+    with _global_hook_lock:
+        if _global_hook_instance is None:
+            _global_hook_instance = HookObservability(metrics_sink=metrics_sink)
+            _logger.debug(
+                "Created global singleton HookObservability",
+                extra={"has_metrics_sink": metrics_sink is not None},
+            )
+        elif metrics_sink is not None:
+            # Log warning when config is provided but ignored for existing singleton
+            _logger.warning(
+                "Global HookObservability singleton already exists; "
+                "provided metrics_sink configuration ignored. "
+                "Call clear_global_hook() first to reconfigure.",
+                extra={
+                    "existing_has_metrics_sink": (
+                        _global_hook_instance.metrics_sink is not None
+                    ),
+                    "provided_metrics_sink_type": type(metrics_sink).__name__,
+                },
+            )
+        return _global_hook_instance
+
+
+def configure_global_hook(
+    metrics_sink: ProtocolHotPathMetricsSink | None = None,
+) -> HookObservability:
+    """Configure the global singleton HookObservability instance.
+
+    This is an alias for get_global_hook() that makes the intent clearer when
+    you want to explicitly configure the singleton on first use.
+
+    Args:
+        metrics_sink: Optional metrics sink for the hook.
+
+    Returns:
+        The global singleton HookObservability instance.
+
+    See Also:
+        get_global_hook(): Primary singleton access function.
+        clear_global_hook(): Reset the singleton for reconfiguration.
+    """
+    return get_global_hook(metrics_sink=metrics_sink)
+
+
+def clear_global_hook() -> None:
+    """Clear the global singleton HookObservability instance.
+
+    Removes the reference to the cached singleton hook, allowing it to be
+    garbage collected if no other references exist. Subsequent calls to
+    get_global_hook() will create a new instance.
+
+    Use Cases:
+        - Testing: Reset singleton state between tests
+        - Reconfiguration: Allow new singleton with different metrics_sink
+        - Shutdown: Release resources before application exit
+
+    Thread Safety:
+        This function is thread-safe.
+
+    Warning:
+        Existing references to the old singleton remain valid. Only future
+        get_global_hook() calls will create a new instance.
+
+    Example:
+        ```python
+        hook1 = get_global_hook(metrics_sink=sink1)
+        clear_global_hook()
+        hook2 = get_global_hook(metrics_sink=sink2)
+        assert hook1 is not hook2  # Different instances
+        # hook1 still works but is no longer the singleton
+        ```
+    """
+    global _global_hook_instance  # noqa: PLW0603 - Standard singleton pattern
+
+    with _global_hook_lock:
+        if _global_hook_instance is not None:
+            _logger.debug("Cleared global HookObservability singleton")
+        _global_hook_instance = None
+
+
+def has_global_hook() -> bool:
+    """Check if a global singleton HookObservability exists.
+
+    Returns:
+        True if a singleton hook has been created, False otherwise.
+
+    Thread Safety:
+        This function is thread-safe.
+    """
+    with _global_hook_lock:
+        return _global_hook_instance is not None
+
+
+__all__ = [
+    "HookObservability",
+    "get_global_hook",
+    "configure_global_hook",
+    "clear_global_hook",
+    "has_global_hook",
+]