What you gain
No egress meter, no per-minute transcode bill, no per-session analytics licence. The marginal cost of one more viewer approaches the cost of the bandwidth itself.
Everything inside your walls
One virtualised estate carries the whole platform. The application tier lives on a single node; live and video-on-demand are served from separate edge machines so a match-night surge cannot starve the back catalogue. Nothing about this design depends on a public cloud being reachable.
How it fits together
Every tier named, every boundary drawn. Blue is capacity you own; grey is capacity you rent.
Untrusted — the public internet
- Web player
- Mobile apps
- TV apps
- Contribution
TLS terminates here and nowhere else
- traefik-edge
- CrowdSec
- Live path
- VOD path
Reachable only through the edge, or through the bastion for operators
- app-core
- content-api
- search-api
- session-authz
- billing
- key-delivery
- recommendation
- bastion
A match-night surge cannot evict the back catalogue
- live-transcode
- live-origin
- vod-transcode + GPU
- vod-origin-cache
- upload-api
- job-scheduler
- RabbitMQ
- MinIO
State lives inside the perimeter
- MariaDB
- Redis
- Elasticsearch
- ClickHouse
- Redpanda
- OpenBao
Declared entirely in code — the estate rebuilds from the repository
- Proxmox VE
- Terraform
- Ansible
- Pipeline → registry
- Prometheus
- SigNoz
- Grafana
- Loki
Tier by tier
| Tier | Runs on | Scales by | Fails over to |
|---|---|---|---|
| Application | Isolated guests on Proxmox | Adding guests, then nodes | Restart on a peer node |
| Live edge | Dedicated host | Adding edge hosts behind DNS | VOD edge, degraded |
| VOD edge | Dedicated host | Adding edge hosts behind DNS | Origin direct, degraded |
| Origin | MinIO on local disk | Adding drives, then nodes | Erasure-coded set |
| Encoder | GPU passthrough guest | Queue depth, then GPUs | Queued, not dropped |
| Control plane | Terraform + Git | Not applicable | State is in Git |
What it buys and what it costs
Every architecture pays for something. Here is the bill for this one.
What you take on
Hardware, power, and a peering or transit relationship. This pattern rewards predictable regional traffic and punishes unpredictable global traffic.
Where it stops
Serving a distant continent from one region is a physics problem, not a software one. At that point the answer is the hybrid pattern, not a bigger box.
When this is the right pattern
You already have rack space, your traffic is regional, and metered egress is the line item you want to delete entirely.
The other two patterns
Most evaluations end up comparing all three. They are all delivered, and none of them is the default answer.
Which one fits your estate?
Send what you have — hardware, cloud account, current bill, launch date. The answer is usually one of these three, and occasionally it is that you should change nothing.