ADR-0013 — Run MVP-1 on a single VPS via docker-compose

Status

Accepted — 2026-04-28

Context

MVP-1 traffic is < 100 messages/day from two known users. The bot must be reachable from Telegram (long-polling — ADR-0012 — so no inbound HTTPS endpoint required) and store data in PostgreSQL with local Prometheus + Grafana.

The operator is a solo developer who is also the primary user. There is no on-call rotation, no SLO commitment beyond “best effort”, and no compliance requirement.

Considered alternatives

Option	Summary	Pros	Cons	Outcome
A — Managed Kubernetes (GKE / EKS)	Each service as a Deployment	”Production-grade”; trivial scale-out	Months of ops setup; ~$70/mo idle cost; comically over-engineered for 100 msg/day; no scale story justifies it	rejected
B — Self-hosted k3s on the VPS	k3s + Helm chart per service	Lighter than full k8s; Helm portfolio sample	Still wildly more YAML than two docker-compose services; troubleshooting eats focus	rejected
C — Managed services (Heroku / Fly / Render)	Each container as a managed dyno	Zero ops	$$ at idle; Postgres add-ons cost; vendor lock-in for a bot that should be trivially portable	rejected
D (chosen) — Single VPS + docker-compose	One $5–10/mo VPS, all containers via `docker compose up`	Simplest possible; portable across providers; all logs/metrics local	Single point of failure (unimportant for personal bot); manual scale-up later if needed	selected

Decision

We will host the entire MVP-1 stack on one VPS managed by docker-compose. The compose file (compose.yml) defines four services: bot, postgres, prometheus, grafana. Volumes are named (not bind-mounts) so backup/restore is docker run --rm -v pgdata:/data ... tar-style.

All host ports bind to 127.0.0.1 only. External access is via SSH tunnel for Grafana and via Telegram itself for the bot. No reverse proxy, no public TLS, no firewall rules beyond default-deny inbound.

Consequences

Positive

Reproducible local dev: same compose.yml on laptop and on VPS.
Two-line deploy: docker compose pull && docker compose up -d.
~$10/mo all-in (VPS only).

Negative / trade-offs

VPS reboot kills the service for ~30 s. Acceptable: Telegram redelivers updates.
No automatic horizontal scaling. Re-evaluate at MVP-2 if user count grows past ~50.
Single-VPS backup story is “rsync pgdata volume to S3 nightly” — basic but sufficient. Will be formalized in Plan 8 deploy task.

Neutral / follow-ups

Plan 8: write a make deploy Makefile target that does docker compose pull && up -d --remove-orphans.
Plan 8: add nightly pg_dump cron + S3 upload.
Re-evaluate at MVP-2 when external onboarding starts (potential trigger to move to a small managed Postgres for backup off-host).

References

ADR-0004 (Modular monolith) — only one bot container is needed.
ADR-0012 (Long-polling) — no inbound HTTPS required.
compose.yml — current deployment definition.