Enforce port range for Apps on localhost (similar to cloudspace)

[hummuscience]

Description & Motivation

I am trying to deploy Pose App on a local network as I don't have access to the machine in-person.

In localhost-mode, the App spawns multiple Lightning Works that occupy random ports. This makes it hard to open the needed ports in the firewall. Opening up all ports is obviously not wanted. So simply setting the app URL to 0.0.0.0 would also not help in this case.

Example of running lsof -i -P -n | grep LISTEN

python    83652 abdelhaym    9u  IPv4 458580      0t0  TCP 127.0.0.1:58537 (LISTEN)
python    83652 abdelhaym   10u  IPv6 458581      0t0  TCP [::1]:58537 (LISTEN)
python    83653 abdelhaym    9u  IPv4 488642      0t0  TCP 127.0.0.1:35337 (LISTEN)
python    83653 abdelhaym   10u  IPv6 488643      0t0  TCP [::1]:35337 (LISTEN)
python    83654 abdelhaym    9u  IPv4 465625      0t0  TCP 127.0.0.1:33571 (LISTEN)
python    83654 abdelhaym   10u  IPv6 465626      0t0  TCP [::1]:33571 (LISTEN)
python    83655 abdelhaym    9u  IPv4 485462      0t0  TCP 127.0.0.1:37229 (LISTEN)
python    83655 abdelhaym   10u  IPv6 485463      0t0  TCP [::1]:37229 (LISTEN)
pt_main_t 83657 abdelhaym   31u  IPv4 480899      0t0  TCP 127.0.0.1:7501 (LISTEN)
label-stu 83775 abdelhaym    5u  IPv4 467897      0t0  TCP 127.0.0.1:56903 (LISTEN)
python3.1 84173 abdelhaym    7u  IPv4 488698      0t0  TCP 127.0.0.1:49879 (LISTEN)

When I set LIGHTNING_CLOUDSPACE_HOST=0 and LIGHTNING_CLOUDSPACE_EXPOSED_PORT_COUNT=15 the app runs through 0.0.0.0 and the ports follow span the range from 7501 till 7515.

python    85048 abdelhaym    9u  IPv4 466716      0t0  TCP *:7502 (LISTEN)
python    85049 abdelhaym    9u  IPv4 465756      0t0  TCP *:7503 (LISTEN)
python    85050 abdelhaym    9u  IPv4 481099      0t0  TCP *:7504 (LISTEN)
python    85051 abdelhaym    9u  IPv4 475878      0t0  TCP *:7505 (LISTEN)
pt_main_t 85053 abdelhaym   31u  IPv4 484872      0t0  TCP *:7501 (LISTEN)
label-stu 85120 abdelhaym    5u  IPv4 475897      0t0  TCP *:7508 (LISTEN)
python3.1 85422 abdelhaym    7u  IPv4 486706      0t0  TCP *:7506 (LISTEN)

This seems to work. But the app doesn't load. One of the instances (the project_ui instance) creates a weird URL that looks like so: https://7503-0/root.project_ui

I tried to trace this down and got to some like in the multiprocess code. This one:

pytorch-lightning/src/lightning/app/runners/backends/mp_process.py

Line 84 in ac3f1ee

work._future_url = f"https://{work.port}-{constants.LIGHTNING_CLOUDSPACE_HOST}"

And this one:

pytorch-lightning/src/lightning/app/runners/multiprocess.py

Line 74 in ac3f1ee

f"https://{port}-{constants.LIGHTNING_CLOUDSPACE_HOST}"

This seems to be the url that is generated when 'LIGHTNING_CLOUDSPACE_HOST' is set. I tried to set it to "any", "", or "0.0.0.0" with the same issue.

I got it working by editing those to lines to this: f"http://<machine_hostname>:{port}"

Maybe I am missing something, but I couldn't find a better workaround.

Pitch

Setting 'LIGHTNING_CLOUDSPACE_HOST' seems to trigger the app and works to use consecutive ports to run.

Is it possible to allow the same mechanism when 'LIGHTNING_CLOUDSPACE_HOST' is unset?

This way, we would only have to set LIGHTNING_APP_STATE_URL to 0.0.0.0 without messing with the cloud deployment loagic.

Alternatives

No response

Additional context

Version 2.2.5

cc @Borda

[hummuscience]

I also raised this issue in the Pose App repo where you can follow the things I tried: Lightning-Universe/Pose-app#119