operations #

things that bit while running this on Modal.

• a cold client gets a 303. while a container starts, Modal answers long-waiting requests with a redirect, and typesafe-sdk doesn't follow it. warm the app with just health, which follows redirects, before pointing a client at it.

• old containers keep serving after a deploy until they scale down. to test new code right away, stop them by id:

  uv run modal container list
  uv run modal container stop -y <id>
  

• modal app stop and container stop abort without -y when not run interactively.

• an out-of-memory error leaves the container holding ~20 GB until it's replaced. stop it by id. with the kernel loading and rows grouped by length, 150-tool chunks fit on an L4.

• modal app logs shows a recent window, not a container's startup. the startup summary (region, patched-row drift, warm-up time) is on /health instead.

• check where the container landed. /health reports region. without a region pin, an H100 landed in asia-south2.

• modal container exec <id> -- cmd does not forward stdin. pass files base64-encoded in the command instead.

• modal billing: uv run modal billing report --for today --show-resources.

• auth is Modal proxy auth (requires_proxy_auth=True). unauthenticated requests get a 401 from modal-http in ~110 ms, and no container starts (checked with zero containers before and after). tokens are workspace-wide and are revoked only in the dashboard; the CLI can only create them.