python-observability

Configure structlog with a processor chain including merge_contextvars, add_log_level, TimeStamper, and JSONRenderer, then call structlog.get_logger() to emit JSON logs. Initialize the configuration once at application startup with your desired log level.

Store the correlation ID in a ContextVar and bind it with structlog.contextvars so it appears in every log entry. Accept or generate the ID in middleware via the X-Correlation-ID header, then forward that header on outbound httpx requests to downstream services.

Track the four golden signals: request latency as a Histogram, traffic as a Counter, errors as a Counter labeled by error type, and saturation as a Gauge such as database connection pool usage. Label metrics by method, endpoint, and status.

User IDs have unbounded cardinality, creating millions of time series that explode storage costs and degrade query performance. Use bounded labels like endpoint or user tier instead, and record user IDs in structured logs where high cardinality is safe.

Yes, OpenTelemetry supports async Python through tracer.start_as_current_span context managers around awaited operations. Configure a TracerProvider with a BatchSpanProcessor and OTLP exporter, and check the official docs since the API evolves actively.