Cross-reference

ICE (incl. Trickle & mDNS)

Where it enters the lineage

This concept is first developed in § m3-l4 — ICE: The State Machine That Saved P2P. The historical problem there matters: WebRTC components are not arbitrary layers; each is a repair for a specific limit in the system before it.

1What it is

The complete RFC 8445 machine: candidates, priorities, checklists, roles, nomination — plus Trickle ICE, mDNS privacy candidates, consent freshness, and ICE restarts.

ICE gathers host, server-reflexive, peer-reflexive, and relay candidates, forms pairs, runs ordered connectivity checks, and nominates a working pair. Trickle sends candidates incrementally; a restart changes credentials and repeats path selection.

2How to reason about it

ICE candidates describe possible transport addresses, while connectivity checks prove which pairs actually work. The selected pair—not the most attractive-looking candidate—determines whether media is direct, server-reflexive, or relayed.

Separate the control plane from the data plane, then name the clock, identifier, and unit attached to each observation. Ask what is negotiated, what is measured live, and what is merely configured. A robust explanation predicts both a successful trace and the characteristic failure trace.

Concept plate — Follow the mechanism to an observation that can falsify your model.

3Its role in WebRTC

In an application, this concept does not stand alone. It participates in a chain of negotiation, transport, media processing, and feedback. The practical boundary is the API, SDP attribute, RTP/RTCP field, or stats record where the browser exposes it. Use that boundary in tests: feed controlled input, observe the named output, and verify fallback behavior when the preferred path is unavailable.

A useful study method is to draw three columns: configured, negotiated, and observed. Put application preferences and constraints in the first, the answer's accepted parameters in the second, and live packet, state, or stats evidence in the third. Disagreements between columns are diagnostic information. They reveal fallback, unsupported capability, stale state, or a mistaken assumption about which endpoint controls the behavior. This method scales from a single codec preference to an ICE restart or a multi-layer SFU route.

Diagnostic discipline

Do not infer success from configuration alone. A codec in capabilities is not necessarily negotiated; a candidate in SDP is not necessarily selected; a connected peer connection is not necessarily receiving decodable frames. Prefer the narrowest live evidence.

4Failure questions

Question	Evidence
Was it negotiated or selected?	Inspect the answer and the live stats graph.
Did the state transition complete?	Record ordered events with timestamps.
Are counters moving in the expected direction?	Compute deltas; never compare unrelated cumulative samples.
What happens beyond the latency or capacity budget?	Inject delay, loss, reordering, or constrained bandwidth.

5Related concepts

JSEP (incl. Plan B/Unified Plan & ORTC) — RFC 8829's philosophy — the browser does media, you do signaling — its signaling state machine and rollback, and the standards fights (Plan B vs Unified Plan, the ORTC rebellion) fossilized in the API.
GCC & Bandwidth Estimation — Delay-gradient congestion sensing plus loss-based backstop: arrival-time filtering, trendline estimation, AIMD rate control, probing and pacing — and why it must act before loss appears.
PCM & Digital Sampling — Sampling theorem, quantization, bit depth, μ-law/A-law companding and G.711, plus psychoacoustic context — why 8 kHz was chosen in 1962 and why WebRTC runs at 48 kHz internally.
Perfect Negotiation & Glare — Glare mechanics across a century of telephony, polite/impolite roles, the makingOffer/ignoreOffer guards, rollback, and parameterless setLocalDescription — the canonical pattern that makes renegotiation safe, reused in Modules 7-8 and the final Live Lab.

6Further reading

WebRTC: From First Principles · Concept reference