Traditional infrastructure metrics are necessary but insufficient for real-time communications. A healthy CPU graph does not tell you whether callers are experiencing setup delay, packet loss or failed transfers.
Measure the customer journey
Start with service-level indicators that describe what the caller or application experiences: call setup success, post-dial delay, answer rate where applicable, media establishment success, transfer success and clean call completion.
Add media-quality telemetry
RTP quality should be visible by carrier, region, codec, media node and application path. Packet loss, jitter, round-trip time and late or discarded packets are more actionable when correlated with the exact call path.
- RTP packet loss and jitter.
- RTCP quality indicators when available.
- Codec and transcoding distribution.
- Media-node session count and headroom.
Use percentiles for latency
Average latency is rarely sufficient for interactive systems. Track P50, P95 and P99 for call setup, AI turn latency, external API calls and transfer operations. Long-tail latency is often the earliest signal of capacity or dependency problems.