Telemetry & Observability Conventions

If the project has a dedicated telemetry/observability spec or conventions document, read it first — it takes precedence over the generic guidance below.

Log Levels

Level	When to Use	Example
error!	Unrecoverable failures requiring operator attention	External service call failed permanently
warn!	Recoverable issues, degraded operation	Retry triggered, fallback activated
info!	Lifecycle events, operational milestones	Server started, connection established
debug!	Internal state useful during development	State machine transition details
trace!	Per-frame, hot-path data (high volume)	Frame encoded, bytes serialized

Instrumentation Depth

Not all code deserves the same instrumentation level. Choose based on the crate's role:

• Hot-path code (serialization, framing, tight loops): trace! only, no spans. Spans add overhead that matters here.
• Type definition crates (pure data types, no logic): trace! only if anything at all.
• Core business logic (servers, handlers, state machines): Full instrumentation — #[instrument] + spans.
• Client-facing code (SDKs, CLI flows): Full instrumentation for debuggability.

#[instrument] Rules

• Use on async functions that represent logical operations or flow steps
• Skip on hot-path synchronous functions (serialization, encoding, tight loops)
• Always skip sensitive fields: tokens, keys, passwords, raw payloads
• Include identifying fields that aid correlation (IDs, resource names)
• Set appropriate level — default is INFO, use level = "debug" or level = "trace" for noisy functions

Structured Logging

Always use named fields, never string interpolation:

// Good: named fields — searchable, parseable
trace!(payload_bytes = payload.len(), frame_bytes = buf.len(), "frame encoded");

// Bad: string interpolation — opaque to log aggregators
trace!("frame encoded, payload={}, frame={}", payload.len(), buf.len());

Span Design

For request/message processing pipelines, use a three-tier span pattern:

• Inbound span: one per received message/request (message type, size, sender ID)
• Process span: business logic processing (operation type, affected resources)
• Outbound span: one per response/forwarded message (recipient, payload size)

This gives visibility into where time is spent and enables per-hop latency analysis.

General Principles

• Prefer tracing over log — structured spans enable distributed tracing
• Log at the point of decision, not at every intermediate step
• Include enough context to diagnose without reproducing: IDs, sizes, error details
• Never log secrets, tokens, keys, or raw user data — even at trace! level
• Use Display for user-facing context, Debug for developer diagnostics