Allow field path env variables

Summary:
# Context
Currently the kubernetes scheduler translate all the env of the role into something like this.
```
- name: FOO
  value: bar
```
In some cases we may also need env variables of type:
```
- name: FOO
  valueFrom:
    fieldRef:
      fieldPath: bar.path
```
This diff adds the ability to specify env in the role that will be translated to the second option.

# Changes
- Create constant placeholder to insert in the value of env in roles if we want them to be translated to the second option
- If placeholder is present in the value, convert to code below stripping the placeholder
```
V1EnvVar(
  name=key,
    value_from=V1EnvVarSource(
      field_ref=V1ObjectFieldSelector(field_path=value)
    ),
)
```
if not, translate to
```
V1EnvVar(name=key, value=value)
```
- update tests

Differential Revision: D59865500

torchx/schedulers/kubernetes_scheduler.py

@@ -169,6 +169,8 @@
 
 LABEL_INSTANCE_TYPE = "node.kubernetes.io/instance-type"
 
+PLACEHOLDER_FIELD_PATH = "[FIELD_PATH]"
+
 
 def sanitize_for_serialization(obj: object) -> object:
     from kubernetes import client
@@ -183,7 +185,9 @@ def role_to_pod(name: str, role: Role, service_account: Optional[str]) -> "V1Pod
         V1ContainerPort,
         V1EmptyDirVolumeSource,
         V1EnvVar,
+        V1EnvVarSource,
         V1HostPathVolumeSource,
+        V1ObjectFieldSelector,
         V1ObjectMeta,
         V1PersistentVolumeClaimVolumeSource,
         V1Pod,
@@ -303,9 +307,20 @@ def role_to_pod(name: str, role: Role, service_account: Optional[str]) -> "V1Pod
         image=role.image,
         name=name,
         env=[
-            V1EnvVar(
-                name=name,
-                value=value,
+            (
+                V1EnvVar(
+                    name=name,
+                    value_from=V1EnvVarSource(
+                        field_ref=V1ObjectFieldSelector(
+                            field_path=value.strip(PLACEHOLDER_FIELD_PATH)
+                        )
+                    ),
+                )
+                if value.startswith(PLACEHOLDER_FIELD_PATH)
+                else V1EnvVar(
+                    name=name,
+                    value=value,
+                )
             )
             for name, value in role.env.items()
         ],

torchx/schedulers/test/kubernetes_scheduler_test.py

@@ -28,6 +28,7 @@
     KubernetesOpts,
     KubernetesScheduler,
     LABEL_INSTANCE_TYPE,
+    PLACEHOLDER_FIELD_PATH,
     role_to_pod,
 )
 from torchx.specs import AppState
@@ -75,7 +76,7 @@ def _test_app(num_replicas: int = 1) -> specs.AppDef:
             "--rank0-env",
             specs.macros.rank0_env,
         ],
-        env={"FOO": "bar"},
+        env={"FOO": "bar", "FOO_FIELD_PATH": f"{PLACEHOLDER_FIELD_PATH}bar"},
         resource=specs.Resource(
             cpu=2,
             memMB=3000,
@@ -149,7 +150,9 @@ def test_role_to_pod(self) -> None:
             V1ContainerPort,
             V1EmptyDirVolumeSource,
             V1EnvVar,
+            V1EnvVarSource,
             V1HostPathVolumeSource,
+            V1ObjectFieldSelector,
             V1ObjectMeta,
             V1Pod,
             V1PodSpec,
@@ -188,7 +191,15 @@ def test_role_to_pod(self) -> None:
             ],
             image="pytorch/torchx:latest",
             name="name",
-            env=[V1EnvVar(name="FOO", value="bar")],
+            env=[
+                V1EnvVar(name="FOO", value="bar"),
+                V1EnvVar(
+                    name="FOO_FIELD_PATH",
+                    value_from=V1EnvVarSource(
+                        field_ref=V1ObjectFieldSelector(field_path="bar")
+                    ),
+                ),
+            ],
             resources=resources,
             ports=[V1ContainerPort(name="foo", container_port=1234)],
             security_context=V1SecurityContext(),
@@ -303,6 +314,10 @@ def test_submit_dryrun(self) -> None:
           env:
           - name: FOO
             value: bar
+          - name: FOO_FIELD_PATH
+            valueFrom:
+              fieldRef:
+                fieldPath: bar
           - name: TORCHX_RANK0_HOST
             value: localhost
           image: pytorch/torchx:latest