nandi/frq · d0cf0eb

nandi/frqpublic⑂ Fork 0

⑂ d0cf0eb

This file is read twice, and the second time there is no environment

The deploy went green and the URL hung: it connected and then sent
nothing, for as long as anything waited.

Modal imports `app.py` again inside the container, to find `serve` by
name. So everything at module level runs in both places — and the
environment is not the same in both. `FRQ_WEB_IMAGE` is set by whoever
deploys and by nobody in the container, so the `raise` that was there to
stop a deploy without a tag killed every container on start instead. The
port never opened, the proxy waited, and the deploy that reported
success had deployed something that could not run.

`modal.is_local()` is the distinction, and is exported for exactly this.
The check stays where it was and now only fires where it means
something; in the container the tag is beside the point, since the image
is already the one that was deployed, so a placeholder stands in and is
never built.

This is the thing the loader avoided by pickling the function body
instead of importing it — `serialized=True`, with a comment about the
re-import — and the rewrite dropped the trick and then walked into what
it was for.

Deployed and checked rather than reasoned about: the URL answers 200 in
0.4s and serves the page, `frq_core.js` and `frq_host.js`.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

nandi committed 2026-09-20T10:06:48-07:00 Browse files

d0cf0eb parent: bd10e81

modified .modal/web/app.py +13 -4 Viewed

@@ -20,10 +20,19 @@ import subprocess
20	20
21	21	import modal
22	22
23		-# The tag CI just built and pushed. No default: unset, this fails here rather
24		-# than deploying whatever was current the last time somebody ran it.
	23	+# This file is read twice: here, by `modal deploy`, and again inside the
	24	+# container, where Modal imports it to find `serve` by name. Everything at
	25	+# module level therefore runs in both places — and the environment is not the
	26	+# same in both. `FRQ_WEB_IMAGE` is set by whoever deploys and by nobody in the
	27	+# container, so a bare `raise` here killed every container on start, the port
	28	+# never opened, and the URL hung while the deploy reported success.
	29	+#
	30	+# The tag CI just built and pushed. No default when deploying: unset, this
	31	+# fails rather than serving whatever was current the last time somebody ran
	32	+# it. In the container the tag is beside the point — the image is already the
	33	+# one that was deployed — so a placeholder stands in and is never built.
25	34	IMAGE = os.environ.get("FRQ_WEB_IMAGE", "")
26		-if not IMAGE:
	35	+if modal.is_local() and not IMAGE:
27	36	raise SystemExit(
28	37	"FRQ_WEB_IMAGE is unset. It is the image to serve, e.g.\n"
29	38	" FRQ_WEB_IMAGE=registry.rickub.com/nandi/frq-web:<sha> \\\n"
@@ -41,7 +50,7 @@ if not IMAGE:
41	50	# To go the other way, set the image private on rickub and pass
42	51	# `secret=modal.Secret.from_name("rickub-registry")` below, naming a Secret
43	52	# with REGISTRY_USERNAME / REGISTRY_PASSWORD for a pull-only token.
44		-image = modal.Image.from_registry(IMAGE)
	53	+image = modal.Image.from_registry(IMAGE or "python:3.13-slim")
45	54
46	55	# Named, and the name is what makes a second deploy replace the running one
47	56	# rather than stand another beside it.

  import modal
--# The tag CI just built and pushed. No default: unset, this fails here rather
++# This file is read twice: here, by `modal deploy`, and again inside the
--# than deploying whatever was current the last time somebody ran it.
++# container, where Modal imports it to find `serve` by name. Everything at
++# module level therefore runs in both places — and the environment is not the
++# same in both. `FRQ_WEB_IMAGE` is set by whoever deploys and by nobody in the
++# container, so a bare `raise` here killed every container on start, the port
++# never opened, and the URL hung while the deploy reported success.
++#
++# The tag CI just built and pushed. No default when deploying: unset, this
++# fails rather than serving whatever was current the last time somebody ran
++# it. In the container the tag is beside the point — the image is already the
++# one that was deployed — so a placeholder stands in and is never built.
  IMAGE = os.environ.get("FRQ_WEB_IMAGE", "")
--if not IMAGE:
++if modal.is_local() and not IMAGE:
      raise SystemExit(
          "FRQ_WEB_IMAGE is unset. It is the image to serve, e.g.\n"
          "  FRQ_WEB_IMAGE=registry.rickub.com/nandi/frq-web:<sha> \\\n"
  # To go the other way, set the image private on rickub and pass
  # `secret=modal.Secret.from_name("rickub-registry")` below, naming a Secret
  # with REGISTRY_USERNAME / REGISTRY_PASSWORD for a pull-only token.
--image = modal.Image.from_registry(IMAGE)
++image = modal.Image.from_registry(IMAGE or "python:3.13-slim")
  # Named, and the name is what makes a second deploy replace the running one
  # rather than stand another beside it.