up -d hangs indefinitely with depends_on: condition: service_healthy, even when the dependency becomes healthy

[lautou]

Describe the bug

podman-compose -f compose.yaml up -d (no --wait) hangs indefinitely when a service uses depends_on: <service>: condition: service_healthy — confirmed stuck 40+ minutes across two separate CI runs before being cancelled. The hang starts immediately after the dependency's image finishes pulling ("Writing manifest to image destination"), before any container-creation message appears, with zero further output.

The dependency's own healthcheck is not the problem: a separate CI run with the condition: service_healthy clause removed entirely shows postgres reporting healthy within ~10s. So this isn't the dependency actually failing its healthcheck — podman-compose's own dependency-condition polling appears to hang rather than the underlying health status ever resolving.

Versions

• podman-compose: 1.6.0 (installed via pip install podman-compose in CI)
• podman: 4.9.3 (Ubuntu-bundled)
• Environment: GitHub Actions ubuntu-latest runner

To Reproduce

compose.yaml:

services:
  postgres:
    image: docker.io/library/postgres:16-alpine@sha256:57c72fd2a128e416c7fcc499958864df5301e940bca0a56f58fddf30ffc07777
    environment:
      POSTGRES_USER: pie
      POSTGRES_PASSWORD: pie_password
      POSTGRES_DB: pie_db
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U pie -d pie_db"]
      interval: 5s
      timeout: 5s
      retries: 10

  backend:
    image: <any image with a long-running server process>
    depends_on:
      postgres:
        condition: service_healthy

podman-compose -f compose.yaml up -d

Expected: backend starts once postgres reports healthy (a few seconds).
Actual: the command hangs indefinitely with no output, well past the point postgres would already be healthy.

What we ruled out

• Not a "2 services polling the same condition" issue. Our real compose file has two services depending on postgres's health (backend and a second worker service). A first fix attempt removed the condition from only the second service (leaving it depend on backend instead), keeping backend's own postgres: condition: service_healthy intact. The hang recurred at the identical point — proving a single service alone using the condition is sufficient to trigger it, not concurrent polling from multiple services.
• Not Podman Compose 1.3.0 up -d command never returns/finishes/ends #1178 / podman-compose fails to enforce the service_healthy condition #1183. Those were fixed by Fix service_healthy condition enforcing #1184, merged into podman-compose 1.4.0 (2025-05-10). Our CI already runs 1.6.0, which includes that fix — this is a different, still-open issue.
• Not the dependency's healthcheck config itself — see above, postgres becomes healthy quickly on its own once the condition is removed.

Workaround

Removing condition: service_healthy entirely (plain depends_on: [postgres], no condition) resolved the hang immediately — verified across multiple CI runs, an isolated local reproduction, and a real production-style release pipeline. We now rely on the app's own crash/restart resilience (restart: unless-stopped) plus an external, independent health check (a reverse proxy's active healthcheck) instead of podman-compose's own dependency-condition gate.

Full write-up, reproduction data, and the fix: lautou/pie-manager#71