20 lines
1.3 KiB
Markdown
20 lines
1.3 KiB
Markdown
# 0067 No nginx per app
|
|
|
|
Status: Accepted · Area: architecture · Date: 2026-10-16 · Decided by: user
|
|
|
|
## Context
|
|
Each app had its own `apps/<app>/nginx.conf` and `<app>-nginx` container in front of its release, so `tools/deploy.sh` could swap releases with no refused request. The user removed it.
|
|
|
|
## Decision
|
|
- No `nginx.conf` in this repo. Each app's container publishes `HTTP_PORT` on the host itself (3000 inside the container).
|
|
- `tools/deploy.sh <app>` checks the new release's `/healthz` in a container with no published port, then stops the old release and starts the new one on `HTTP_PORT`. A failed check leaves the old release serving; a failed start brings the old one back.
|
|
- The edge that routes phones to bull and desktops to bear (0001) points at each app's `HTTP_PORT`.
|
|
|
|
## Alternatives
|
|
- One nginx per app, reloaded onto the new release: no refused requests during a swap, but one more container and config per app. Removed by the user.
|
|
|
|
## Consequences
|
|
- Each deploy refuses connections for the old release's shutdown (in-flight requests finish, 30 s at most) plus the new one's start, about half a second when idle.
|
|
- `/healthz` answers on `HTTP_PORT`; the edge should hide it from visitors.
|
|
- The edge must set `Vary: User-Agent` (0001) and must not buffer streamed pages (the apps send `X-Accel-Buffering: no`).
|