Playwright in a podman container: page.goto net::ERR_SOCKET_NOT_CONNECTED at http://localhost:<port> for a server published on the host; use host.containers.internal
Setup: a web app runs in one rootless podman container published as -p 9347:8765. A second container, mcr.microsoft.com/playwright, runs an end-to-end script against it with podman run --rm --network host -v .:/workspace:Z -w /workspace <playwright image> bash -c "npm install playwright@<image version> && node test.js".
Symptom: Test failed: page.goto: net::ERR_SOCKET_NOT_CONNECTED at http://localhost:9347/ (Call log: navigating to "http://localhost:9347/", waiting until "networkidle"). Meanwhile curl -s http://127.0.0.1:9347/api/models from the host worked fine.
Likely cause: same issue as plain curl on the host. With rootless pasta, localhost resolves to ::1 first, the IPv6 connection to the published port is forwarded into a server that only listens on IPv4 (0.0.0.0), and the connection drops. Chromium reports that as ERR_SOCKET_NOT_CONNECTED instead of falling back to 127.0.0.1.
What worked: drop --network host and address the host by name:
podman run --rm --add-host=host.containers.internal:host-gateway \
-v .:/workspace:Z -w /workspace mcr.microsoft.com/playwright:<tag> \
bash -c "npm ci && node test.js"
# test.js
await page.goto('http://host.containers.internal:9347', { waitUntil: 'networkidle' });The page then loaded and its console logs showed up in the test. (Recent podman adds host.containers.internal to /etc/hosts by default, but passing --add-host makes it explicit.) Making the app server listen on :: (dual-stack) should also fix localhost URLs. Separately, if the page uses Leaflet tiles or other CDN assets, waitUntil: 'networkidle' plus a waitForFunction on map state can time out at 30000 ms. Wait for a specific DOM element, such as .leaflet-marker-icon, rather than global idle.