WebRTC Media Infrastructure Engineer Interview Questions

Option B is strongest because moving to an SFU with simulcast and active-speaker-aware forwarding scales upload cost and per-client bandwidth correctly as participant count grows, which is the standard, proven architecture for this exact scaling problem. Option A keeps a topology whose per-client bandwidth and CPU requirements grow with participant count, which becomes infeasible well before dozens of participants as stated in the question. Option C applies MCU transcoding, which adds server compute and latency cost, universally even to small calls where it is unnecessary overhead compared to a lighter SFU approach. Option D pushes the scaling problem onto users manually rather than solving it architecturally, and does not address the fundamental mesh bandwidth growth problem.

Option B is strongest because it addresses the actual technical cause, restrictive networks blocking or breaking STUN-based direct connectivity, with a properly configured TURN relay fallback including TCP 443 support, and adds telemetry to right-size TURN capacity. Option A abandons a solvable infrastructure problem and offers no real fix to affected users. Option C is exactly the known gap that fails on symmetric NAT and firewalled networks, matching the failure described in the question. Option D unnecessarily routes all traffic through relay servers, adding latency and relay infrastructure cost for the majority of users who could connect directly, when the fallback is only needed for the restrictive-network minority.

Option B is strongest because it investigates the actual likely causes, network path congestion, packet loss, and bitrate-adaptation behavior, and bandwidth-specific server metrics distinct from CPU, which together explain quality degradation that CPU monitoring alone would miss. Option A jumps to a client-side explanation without evidence and does not investigate the network and bandwidth-adaptation signals that are the more common cause of this specific symptom pattern. Option C assumes the bottleneck is compute when the question explicitly states CPU looks fine, which is very unlikely to fix a network-bandwidth-driven degradation. Option D provides temporary relief for one user's single call at best and does not investigate or fix the underlying peak-hour pattern affecting a subset of users repeatedly.

Option B is strongest because it deploys capacity where real usage data justifies it, elastically scales regional capacity to actual load, and optimizes region selection for the whole group's aggregate latency using continuously improving real telemetry. Option A accepts high latency for distant users and does not address the stated goal of minimizing latency globally. Option C over-provisions uniformly regardless of actual usage concentration, which is the exact over-provisioning the question asks to avoid. Option D optimizes each client's latency in isolation, which can produce a poor overall group experience if participants end up split across regions with no coordinated selection for the group as a whole.

Option B is strongest because it isolates recording as a decoupled, independently scalable pipeline fed from the SFU, ensuring recording load and failures cannot degrade or interrupt the live call experience for participants. Option A makes recording reliability and quality dependent on one user's device and connection, and directly competes with that user's own live call performance, degrading their experience. Option C couples recording load directly into the live forwarding path, creating exactly the degradation risk the question asks to avoid. Option D avoids the architecture problem rather than solving it, unnecessarily limiting a needed product capability instead of designing a properly isolated recording pipeline.