[serve] Use environment variable for root_url from runtime env (#18269)

python/ray/serve/api.py

@@ -106,6 +106,7 @@ class Client:
         self._shutdown = False
         self._http_config: HTTPOptions = ray.get(
             controller.get_http_config.remote())
+        self._root_url = ray.get(self._controller.get_root_url.remote())
 
         # Each handle has the overhead of long poll client, therefore cached.
         self.handle_cache = WeakValueDictionary()
@@ -122,7 +123,7 @@ class Client:
 
     @property
     def root_url(self):
-        return self._http_config.root_url
+        return self._root_url
 
     def __del__(self):
         if not self._detached:
@@ -384,6 +385,10 @@ def start(
                 - "NoServer" or None: disable HTTP server.
             - num_cpus (int): The number of CPU cores to reserve for each
               internal Serve HTTP proxy actor.  Defaults to 0.
+            - root_url (str): The URL where the Deployment will be located at.
+              If no `root_url` is provided, Serve will check the environment
+              variable `RAY_SERVE_ROOT_URL` on the Ray cluster which is
+              serving the Deployment.
         dedicated_cpu (bool): Whether to reserve a CPU core for the internal
           Serve controller actor.  Defaults to False.
     """

python/ray/serve/config.py

@@ -1,14 +1,12 @@
 import inspect
 from enum import Enum
-import os
 from typing import Any, List, Optional
 
 import pydantic
 from pydantic import BaseModel, PositiveInt, validator, NonNegativeFloat
 
 from ray import cloudpickle as cloudpickle
-from ray.serve.constants import (DEFAULT_HTTP_HOST, DEFAULT_HTTP_PORT,
-                                 SERVE_ROOT_URL_ENV_KEY)
+from ray.serve.constants import DEFAULT_HTTP_HOST, DEFAULT_HTTP_PORT
 
 
 class BackendConfig(BaseModel):
@@ -167,15 +165,6 @@ class HTTPOptions(pydantic.BaseModel):
             return DeploymentMode.NoServer
         return v
 
-    @validator("root_url", always=True)
-    def fill_default_root_url(cls, v, values):
-        if v == "":
-            if SERVE_ROOT_URL_ENV_KEY in os.environ:
-                return os.environ[SERVE_ROOT_URL_ENV_KEY]
-            else:
-                return f"http://{values['host']}:{values['port']}"
-        return v
-
     class Config:
         validate_assignment = True
         extra = "forbid"

python/ray/serve/controller.py

@@ -2,6 +2,7 @@ import asyncio
 import json
 import time
 from collections import defaultdict
+import os
 from typing import Dict, List, Optional, Tuple, Any
 
 import ray
@@ -19,7 +20,7 @@ from ray.serve.common import (
     ReplicaTag,
 )
 from ray.serve.config import BackendConfig, HTTPOptions, ReplicaConfig
-from ray.serve.constants import CONTROL_LOOP_PERIOD_S
+from ray.serve.constants import CONTROL_LOOP_PERIOD_S, SERVE_ROOT_URL_ENV_KEY
 from ray.serve.endpoint_state import EndpointState
 from ray.serve.http_state import HTTPState
 from ray.serve.storage.kv_store import RayInternalKVStore
@@ -179,6 +180,16 @@ class ServeController:
         """Return the HTTP proxy configuration."""
         return self.http_state.get_config()
 
+    def get_root_url(self):
+        """Return the root url for the serve instance."""
+        http_config = self.get_http_config()
+        if http_config.root_url == "":
+            if SERVE_ROOT_URL_ENV_KEY in os.environ:
+                return os.environ[SERVE_ROOT_URL_ENV_KEY]
+            else:
+                return f"http://{http_config.host}:{http_config.port}"
+        return http_config.root_url
+
     async def shutdown(self) -> List[GoalId]:
         """Shuts down the serve instance completely."""
         async with self.write_lock:

python/ray/serve/tests/test_standalone.py

@@ -289,6 +289,7 @@ def test_http_root_url(ray_shutdown):
     f.deploy()
     assert f.url == root_url + "/f"
     serve.shutdown()
+    ray.shutdown()
     del os.environ[SERVE_ROOT_URL_ENV_KEY]
 
     port = new_port()
@@ -297,6 +298,15 @@ def test_http_root_url(ray_shutdown):
     assert f.url != root_url + "/f"
     assert f.url == f"http://127.0.0.1:{port}/f"
     serve.shutdown()
+    ray.shutdown()
+
+    ray.init(runtime_env={"env_vars": {SERVE_ROOT_URL_ENV_KEY: root_url}})
+    port = new_port()
+    serve.start(http_options=dict(port=port))
+    f.deploy()
+    assert f.url == root_url + "/f"
+    serve.shutdown()
+    ray.shutdown()
 
 
 @pytest.mark.skipif(sys.platform == "win32", reason="Failing on Windows")