Glossary
Quality of Service (QoS)
Not all packets are equal. Quality of Service is how networks and applications make sure that a voice frame or a video packet gets through on time even when a file download is competing for the same link.
General definition
Quality of Service (QoS) has two layers. Network QoS is what routers, switches and Wi-Fi access points do: classify packets, queue them by priority and give delay-sensitive traffic a shorter line. Application QoS is what the software does to cope with whatever the network delivers: measure conditions, adapt the sending rate, buffer against variation and repair losses. Both exist because latency, jitter and packet loss are what users actually feel, and raw bandwidth alone does not control them.
On the network side the standard tool is Differentiated Services (RFC 2474). Each IP packet carries a six-bit DSCP field; a sender marks voice with Expedited Forwarding (EF, DSCP 46) and interactive video with an Assured Forwarding class such as AF41, and every hop that honours the marking queues those packets ahead of bulk traffic. RFC 8837 recommends the markings for WebRTC flows, and Wi-Fi maps them onto WMM access categories. The catch is scope: DSCP is respected inside networks you administer, typically an office, hospital or campus LAN and its uplink, and usually stripped or ignored once traffic crosses the public internet.
- Bitrate adaptation: WebRTC bandwidth estimation (transport-wide congestion control) raises and lowers the encoder bitrate to match what the path can carry, trading resolution for continuity
- Jitter buffers: the receiver holds a few tens of milliseconds of audio or video so packets that arrive unevenly play back smoothly, growing the buffer when jitter rises and shrinking it when the network calms
- Loss repair: retransmission, forward error correction and concealment cover gaps that prioritisation could not prevent
- Codec choice and layering: Opus adapts its rate on its own, and simulcast lets a media server serve each viewer a layer their link can sustain
QoS is measured against perceived quality. The Mean Opinion Score (MOS, ITU-T P.800) rates a call from 1 to 5, with 4 and above considered good and anything under 3.5 noticeably degraded; the E-model (ITU-T G.107) predicts it from delay, loss and codec. Chat degrades more gently than calls: a text message over TCP arrives late rather than broken, so QoS on a messaging platform is mostly about keeping the connection alive and delivering in order, while calls need the full toolkit.
In the Ethora ecosystem
Ethora runs chat over XMPP on WebSocket and calls over WebRTC through an SFU, so the application-level QoS mechanisms above are part of the stack: WebRTC bandwidth estimation, jitter buffering and loss repair on every call leg, and a server-side message archive with unread counts and last-message state so a chat client that suffers a bad stretch of network reconnects and catches up in order. Audio-only calls are available as an opt-in, which is the simplest QoS decision of all when a link cannot sustain video.
Network QoS is only as good as the network you control, which is one reason regulated organisations choose a dedicated server or self-hosted deployment: the media server and chat backend sit inside their own cloud account or on-premise network, where DSCP markings are honoured end to end on the LAN and the operations team can use the bundled Grafana and Prometheus monitoring to watch latency, jitter and loss rather than guessing. See the voice and video call API page for the calling stack.
