ci: build image + validate nginx on PRs; fix duplicate proxy_read_timeout

.github/workflows/ci.yml

@@ -30,3 +30,41 @@ jobs:
 
       - name: Test
         run: uv run pytest -m 'not integration and not slow' -v
+
+  # Same Dockerfile deploy.yml pushes, built but not pushed — a PR that would
+  # break the prod image (Dockerfile, uv.lock, missing files) fails here instead.
+  build:
+    needs: test
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: docker/setup-buildx-action@v3
+
+      - name: Validate compose file
+        run: touch .env && docker compose config -q  # .env is gitignored
+
+      - name: Build image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          push: false
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+
+      # Renders the template with envsubst (the nginx image does it on start) and
+      # runs nginx -t, so a bad directive fails here, not as a crash-looping nginx
+      # on deploy. Throwaway cert + `app` host stand in for the real ones.
+      - name: Validate nginx config
+        run: |
+          d=$(mktemp -d); mkdir -p $d/live/ci.test
+          openssl req -x509 -newkey rsa:2048 -nodes -days 1 -subj /CN=ci.test \
+            -keyout $d/live/ci.test/privkey.pem -out $d/live/ci.test/fullchain.pem
+          docker run --rm --add-host app:127.0.0.1 -e DOMAIN=ci.test \
+            -v $PWD/nginx/nginx.conf:/etc/nginx/nginx.conf:ro \
+            -v $PWD/nginx/snippets:/etc/nginx/snippets:ro \
+            -v $PWD/nginx/templates:/etc/nginx/templates:ro \
+            -v /dev/null:/etc/nginx/conf.d/default.conf:ro \
+            -v $d:/etc/letsencrypt:ro \
+            nginx:alpine nginx -t

nginx/snippets/proxy-pass.conf

@@ -4,11 +4,8 @@
 set $upstream_app app:8000;
 proxy_pass http://$upstream_app;
 
-# No backend synthesis timeout exists today (ffmpeg.py only bounds the
-# encode step at 30s; TTS inference itself is unbounded) — pick a ceiling
-# safely above realistic worst-case rather than nginx's low defaults cutting
-# off legitimate slow requests.
-proxy_read_timeout 120s;
+# proxy_read_timeout is set per location (nginx rejects a directive repeated in
+# one block, so it can't live here and be overridden after the include).
 proxy_send_timeout 120s;
 proxy_connect_timeout 5s;
 

nginx/templates/20-speech-server.conf.template

@@ -74,9 +74,8 @@ server {
         include /etc/nginx/snippets/proxy-pass.conf;
 
         # TTS inference is CPU-bound and can be slow (24-layer models, long text,
-        # emulated arch) — give synthesis more than the shared 120s read timeout so
-        # a legitimately long generation isn't cut off with a 504. This override must
-        # come AFTER the include, whose 120s would otherwise win.
+        # emulated arch) — give synthesis more than the 120s default on other
+        # routes so a legitimately long generation isn't cut off with a 504.
         proxy_read_timeout 300s;
     }
 
@@ -91,5 +90,9 @@ server {
         limit_conn perip 20;
 
         include /etc/nginx/snippets/proxy-pass.conf;
+
+        # No backend timeout exists today (TTS inference is unbounded) — pick a
+        # ceiling safely above realistic worst-case rather than nginx's low default.
+        proxy_read_timeout 120s;
     }
 }