[Train] Placement group reserved resources breaking changes

[MortalHappiness]

What happened + What you expected to happen

• In Ray 2.9.0, the reserved resources for CPUs are 1, 2, 2
• In Ray 2.34.0, the reserved resources for CPUs are 3, 2

Is this a breaking change that is not documented?

In KubeRay example like this https://docs.ray.io/en/releases-2.34.0/cluster/kubernetes/examples/mnist-training-example.html

Because each nodes only have 2 CPUs, this example causes them to fail.

Versions / Dependencies

2.34.0

Reproduction script

https://docs.ray.io/en/releases-2.34.0/cluster/kubernetes/examples/mnist-training-example.html

Run the above example with both Ray 2.9.0 and Ray 2.34.0 and see the differences.

Issue Severity

None

[MortalHappiness]

cc @kevin85421

[kevin85421]

cc @woshiyyya @justinvyu is this an expected breaking change? It breaks a KubeRay example. It's an easy fix for me, but not sure whether other Ray Train users will have the same issue or not.