Live education is unforgiving: when class goes live, latency and failure modes are visible to students and teachers immediately. In one chapter of my work we needed live chat in place by a hard deadline, then had to scale the system to 5k+ concurrent users while the broader product served on the order of 80k+ learners.
The stack combined real-time channels, careful database design, and deployment discipline. Integrations with external video providers taught me that vendor SLAs are part of your architecture — when quality dropped, we had to rip a provider out and fall back to other paths (including manual steps where automation hit platform limits).
If you’re building similar systems, plan for spiky load, not average load — and rehearse incident response before you need it at 9pm on a school night.