bakery-ia/shared/monitoring/otel_config.py

"""
Centralized OpenTelemetry Configuration
Manages OTEL endpoints and settings for traces, metrics, and logs
"""

import os
from typing import Optional, Tuple
from dataclasses import dataclass
import structlog

logger = structlog.get_logger()


@dataclass
class OTelEndpoints:
    """
    Container for OpenTelemetry endpoints.

    SigNoz uses different protocols for different signals:
    - Traces: gRPC (port 4317)
    - Metrics: gRPC (port 4317) or HTTP (port 4318)
    - Logs: HTTP (port 4318)
    """
    traces_grpc: str  # gRPC endpoint for traces (e.g., "host:4317")
    metrics_grpc: str  # gRPC endpoint for metrics (e.g., "host:4317")
    metrics_http: str  # HTTP endpoint for metrics (e.g., "http://host:4318/v1/metrics")
    logs_http: str     # HTTP endpoint for logs (e.g., "http://host:4318/v1/logs")


class OTelConfig:
    """
    Centralized configuration for OpenTelemetry exporters.

    This class manages endpoint URLs and ensures proper protocol usage:
    - gRPC endpoints: host:port (no protocol prefix)
    - HTTP endpoints: http://host:port/path (with protocol and path)
    """

    # Default base endpoint (can be overridden by environment variables)
    DEFAULT_OTEL_COLLECTOR_HOST = "signoz-otel-collector.bakery-ia.svc.cluster.local"
    DEFAULT_GRPC_PORT = 4317
    DEFAULT_HTTP_PORT = 4318

    @classmethod
    def get_endpoints(cls) -> OTelEndpoints:
        """
        Get OpenTelemetry endpoints from environment variables with proper fallbacks.

        Environment variables (in order of precedence):
        1. OTEL_EXPORTER_OTLP_ENDPOINT - Base endpoint (gRPC format: host:port)
        2. OTEL_EXPORTER_OTLP_TRACES_ENDPOINT - Specific traces endpoint
        3. OTEL_EXPORTER_OTLP_METRICS_ENDPOINT - Specific metrics endpoint
        4. OTEL_EXPORTER_OTLP_LOGS_ENDPOINT - Specific logs endpoint
        5. OTEL_COLLECTOR_ENDPOINT - Legacy variable (HTTP format)

        Returns:
            OTelEndpoints with all configured endpoints
        """
        # Get base endpoint from environment
        base_endpoint = os.getenv("OTEL_EXPORTER_OTLP_ENDPOINT")

        if base_endpoint:
            # Clean and parse base endpoint
            base_grpc = cls._clean_grpc_endpoint(base_endpoint)
            base_http_host = cls._extract_host(base_endpoint)
            
            # Validate that the endpoint doesn't contain secret references or malformed data
            if cls._contains_secret_reference(base_grpc):
                logger.error("OTEL endpoint contains secret reference, falling back to default", 
                           malformed_endpoint=base_endpoint)
                base_grpc = f"{cls.DEFAULT_OTEL_COLLECTOR_HOST}:{cls.DEFAULT_GRPC_PORT}"
                base_http_host = f"http://{cls.DEFAULT_OTEL_COLLECTOR_HOST}:{cls.DEFAULT_HTTP_PORT}"
        else:
            # Use default collector
            base_grpc = f"{cls.DEFAULT_OTEL_COLLECTOR_HOST}:{cls.DEFAULT_GRPC_PORT}"
            base_http_host = f"http://{cls.DEFAULT_OTEL_COLLECTOR_HOST}:{cls.DEFAULT_HTTP_PORT}"

        # Get signal-specific endpoints (or use base endpoint)
        traces_endpoint = os.getenv("OTEL_EXPORTER_OTLP_TRACES_ENDPOINT", base_grpc)
        metrics_endpoint = os.getenv("OTEL_EXPORTER_OTLP_METRICS_ENDPOINT", base_grpc)
        logs_endpoint = os.getenv("OTEL_EXPORTER_OTLP_LOGS_ENDPOINT")

        # Validate and clean signal-specific endpoints
        traces_grpc = cls._clean_and_validate_grpc_endpoint(traces_endpoint)
        metrics_grpc = cls._clean_and_validate_grpc_endpoint(metrics_endpoint)

        # For metrics HTTP, convert gRPC endpoint to HTTP if needed
        metrics_http = cls._grpc_to_http_endpoint(metrics_grpc, "/v1/metrics")

        # For logs, use HTTP endpoint
        if logs_endpoint:
            logs_http = cls._ensure_http_endpoint(logs_endpoint, "/v1/logs")
        else:
            logs_http = cls._grpc_to_http_endpoint(base_grpc, "/v1/logs")

        endpoints = OTelEndpoints(
            traces_grpc=traces_grpc,
            metrics_grpc=metrics_grpc,
            metrics_http=metrics_http,
            logs_http=logs_http
        )

        logger.info(
            "OpenTelemetry endpoints configured",
            traces_grpc=endpoints.traces_grpc,
            metrics_grpc=endpoints.metrics_grpc,
            metrics_http=endpoints.metrics_http,
            logs_http=endpoints.logs_http
        )

        return endpoints

    @staticmethod
    def _clean_grpc_endpoint(endpoint: str) -> str:
        """
        Clean endpoint for gRPC usage (remove protocol, paths).

        Args:
            endpoint: Raw endpoint string

        Returns:
            Cleaned endpoint in format "host:port"
        """
        # Remove protocol prefixes
        endpoint = endpoint.replace("http://", "").replace("https://", "")

        # Remove paths (gRPC doesn't use paths)
        if "/" in endpoint:
            endpoint = endpoint.split("/")[0]

        # Ensure it has a port
        if ":" not in endpoint:
            endpoint = f"{endpoint}:4317"

        return endpoint

    @staticmethod
    def _extract_host(endpoint: str) -> str:
        """
        Extract host and convert to HTTP endpoint.

        Args:
            endpoint: Raw endpoint string

        Returns:
            HTTP endpoint without path (e.g., "http://host:4318")
        """
        # Remove protocol if present
        clean = endpoint.replace("http://", "").replace("https://", "")

        # Remove path if present
        if "/" in clean:
            clean = clean.split("/")[0]

        # Extract host without port
        if ":" in clean:
            host = clean.split(":")[0]
        else:
            host = clean

        return f"http://{host}:4318"

    @staticmethod
    def _grpc_to_http_endpoint(grpc_endpoint: str, path: str) -> str:
        """
        Convert gRPC endpoint to HTTP endpoint with path.

        Args:
            grpc_endpoint: gRPC endpoint (e.g., "host:4317")
            path: HTTP path (e.g., "/v1/metrics")

        Returns:
            HTTP endpoint (e.g., "http://host:4318/v1/metrics")
        """
        # Extract host from gRPC endpoint
        if ":" in grpc_endpoint:
            host = grpc_endpoint.split(":")[0]
        else:
            host = grpc_endpoint

        # Build HTTP endpoint with port 4318
        return f"http://{host}:4318{path}"

    @staticmethod
    def _ensure_http_endpoint(endpoint: str, path: str) -> str:
        """
        Ensure endpoint is in HTTP format with proper path.

        Args:
            endpoint: Raw endpoint string
            path: Required path (e.g., "/v1/logs")

        Returns:
            HTTP endpoint with protocol and path
        """
        # Add protocol if missing
        if not endpoint.startswith(("http://", "https://")):
            endpoint = f"http://{endpoint}"

        # Ensure it has the correct port for HTTP
        if ":4317" in endpoint:
            endpoint = endpoint.replace(":4317", ":4318")
        elif ":4318" not in endpoint and ":" in endpoint:
            # Has a port but not the right one, replace it
            parts = endpoint.split(":")
            if len(parts) >= 2:
                # Remove existing port and path
                base = ":".join(parts[:-1])
                endpoint = f"{base}:4318"
        elif ":" not in endpoint.replace("http://", "").replace("https://", ""):
            # No port at all, add it
            endpoint = f"{endpoint}:4318"

        # Ensure path is present
        if not endpoint.endswith(path):
            # Remove any existing path first
            if "/" in endpoint.split("://")[1]:
                base = endpoint.split("://")[0] + "://" + endpoint.split("://")[1].split("/")[0]
                endpoint = base
            endpoint = f"{endpoint}{path}"

        return endpoint

    @classmethod
    def get_resource_attributes(
        cls,
        service_name: str,
        service_version: str = "1.0.0"
    ) -> dict:
        """
        Get common resource attributes for all OTEL signals.

        Args:
            service_name: Name of the service
            service_version: Version of the service

        Returns:
            Dictionary of resource attributes
        """
        return {
            "service.name": service_name,
            "service.version": service_version,
            "deployment.environment": os.getenv("ENVIRONMENT", "development"),
            "k8s.namespace.name": os.getenv("K8S_NAMESPACE", "bakery-ia"),
            "k8s.pod.name": os.getenv("HOSTNAME", "unknown"),
            "k8s.cluster.name": os.getenv("K8S_CLUSTER_NAME", "bakery-ia-cluster"),
        }

    @classmethod
    def is_enabled(cls, signal: str) -> bool:
        """
        Check if a specific telemetry signal is enabled.

        Args:
            signal: One of "traces", "metrics", "logs"

        Returns:
            True if signal is enabled, False otherwise
        """
        signal = signal.lower()

        if signal == "traces":
            return os.getenv("ENABLE_TRACING", "true").lower() == "true"
        elif signal == "metrics":
            return os.getenv("ENABLE_OTEL_METRICS", "true").lower() == "true"
        elif signal == "logs":
            return os.getenv("OTEL_LOGS_EXPORTER", "").lower() == "otlp"
        else:
            return False

    @classmethod
    def get_protocol(cls, signal: str) -> str:
        """
        Get the preferred protocol for a signal.

        Args:
            signal: One of "traces", "metrics", "logs"

        Returns:
            Protocol name ("grpc" or "http")
        """
        protocol = os.getenv("OTEL_EXPORTER_OTLP_PROTOCOL", "grpc")

        # Signal-specific overrides
        if signal == "traces":
            return os.getenv("OTEL_EXPORTER_OTLP_TRACES_PROTOCOL", protocol)
        elif signal == "metrics":
            return os.getenv("OTEL_EXPORTER_OTLP_METRICS_PROTOCOL", protocol)
        elif signal == "logs":
            # Logs always use HTTP in our setup
            return "http"

        return protocol