[Serve] Add experimental CLI for serve deploy (#20371)

ci/travis/ci.sh

@@ -141,6 +141,7 @@ test_python() {
       python/ray/tests/...
       -python/ray/serve:conda_env # runtime_env unsupported on Windows
       -python/ray/serve:test_api # segfault on windows? https://github.com/ray-project/ray/issues/12541
+      -python/ray/serve:test_cli # cli
       -python/ray/serve:test_router # timeout
       -python/ray/serve:test_handle # "fatal error" (?) https://github.com/ray-project/ray/pull/13695
       -python/ray/serve:test_controller_crashes # timeout

doc/source/serve/deployment.rst

@@ -78,6 +78,7 @@ In general, **Option 2 is recommended for most users** because it allows you to
 
   my_func.deploy()
 
+
 Deploying on Kubernetes
 =======================
 

python/ray/serve/api.py

@@ -658,7 +658,7 @@ class Deployment:
             raise TypeError("version must be a string.")
         if not (prev_version is None or isinstance(prev_version, str)):
             raise TypeError("prev_version must be a string.")
-        if not (init_args is None or isinstance(init_args, tuple)):
+        if not (init_args is None or isinstance(init_args, (tuple, list))):
             raise TypeError("init_args must be a tuple.")
         if not (init_kwargs is None or isinstance(init_kwargs, dict)):
             raise TypeError("init_kwargs must be a dict.")

python/ray/serve/scripts.py

@@ -1,9 +1,13 @@
 #!/usr/bin/env python
+import json
+import os
 
 import click
-from ray.serve.config import DeploymentMode
 
 import ray
+from ray.serve.api import Deployment
+from ray.serve.config import DeploymentMode
+from ray._private.utils import import_attr
 from ray import serve
 from ray.serve.constants import (DEFAULT_CHECKPOINT_PATH, DEFAULT_HTTP_HOST,
                                  DEFAULT_HTTP_PORT)
@@ -14,7 +18,7 @@ from ray.serve.constants import (DEFAULT_CHECKPOINT_PATH, DEFAULT_HTTP_HOST,
 @click.option(
     "--address",
     "-a",
-    default="auto",
+    default=os.environ.get("RAY_ADDRESS", "auto"),
     required=False,
     type=str,
     help="Address of the running Ray cluster to connect to. "
@@ -26,8 +30,19 @@ from ray.serve.constants import (DEFAULT_CHECKPOINT_PATH, DEFAULT_HTTP_HOST,
     required=False,
     type=str,
     help="Ray namespace to connect to. Defaults to \"serve\".")
-def cli(address, namespace):
-    ray.init(address=address, namespace=namespace)
+@click.option(
+    "--runtime-env-json",
+    default=r"{}",
+    required=False,
+    type=str,
+    help=("Runtime environment dictionary to pass into ray.init. "
+          "Defaults to empty."))
+def cli(address, namespace, runtime_env_json):
+    ray.init(
+        address=address,
+        namespace=namespace,
+        runtime_env=json.loads(runtime_env_json),
+    )
 
 
 @cli.command(help="Start a detached Serve instance on the Ray cluster.")
@@ -73,3 +88,27 @@ def start(http_host, http_port, http_location, checkpoint_path):
 def shutdown():
     serve.api._connect()
     serve.shutdown()
+
+
+@cli.command(
+    help="""
+[Experimental]
+Create a deployment in running Serve instance. The required argument is the
+import path for the deployment: ``my_module.sub_module.file.MyClass``. The
+class may or may not be decorated with ``@serve.deployment``.
+""",
+    hidden=True,
+)
+@click.argument("deployment")
+@click.option(
+    "--options-json",
+    default=r"{}",
+    required=False,
+    type=str,
+    help="JSON string for the deployments options")
+def deploy(deployment: str, options_json: str):
+    deployment_cls = import_attr(deployment)
+    if not isinstance(deployment_cls, Deployment):
+        deployment_cls = serve.deployment(deployment_cls)
+    options = json.loads(options_json)
+    deployment_cls.options(**options).deploy()

python/ray/serve/tests/test_cli.py

@@ -1,7 +1,14 @@
+import json
+import os
+from pathlib import Path
 import subprocess
 import sys
 
 import pytest
+import requests
+
+from ray import serve
+from ray.tests.test_runtime_env_working_dir import tmp_working_dir  # noqa: F401, E501
 
 
 @pytest.fixture
@@ -27,5 +34,62 @@ def test_start_shutdown_in_namespace(ray_start_stop):
     subprocess.check_output(["serve", "-n", "test", "shutdown"])
 
 
+class A:
+    def __init__(self, value, increment=1):
+        self.value = value
+        self.increment = increment
+        self.decrement = 0
+        self.multiplier = int(os.environ["SERVE_TEST_MULTIPLIER"])
+
+        p = Path("hello")
+        assert p.exists()
+        with open(p) as f:
+            assert f.read() == "world"
+
+    def reconfigure(self, config):
+        self.decrement = config["decrement"]
+
+    def __call__(self, inp):
+        return (self.value + self.increment - self.decrement) * self.multiplier
+
+
+@serve.deployment
+class DecoratedA(A):
+    pass
+
+
+@pytest.mark.parametrize("class_name", ["A", "DecoratedA"])
+def test_deploy(ray_start_stop, tmp_working_dir, class_name):  # noqa: F811
+    subprocess.check_output(["serve", "start"])
+    subprocess.check_output([
+        "serve", "--runtime-env-json",
+        json.dumps({
+            "working_dir": tmp_working_dir,
+        }), "deploy", f"ray.serve.tests.test_cli.{class_name}",
+        "--options-json",
+        json.dumps({
+            "name": "B",
+            "init_args": [42],
+            "init_kwargs": {
+                "increment": 10
+            },
+            "num_replicas": 2,
+            "user_config": {
+                "decrement": 5
+            },
+            "ray_actor_options": {
+                "runtime_env": {
+                    "env_vars": {
+                        "SERVE_TEST_MULTIPLIER": "2",
+                    },
+                }
+            }
+        })
+    ])
+    resp = requests.get("http://127.0.0.1:8000/B")
+    resp.raise_for_status()
+    assert resp.text == "94", resp.text
+
+
 if __name__ == "__main__":
     sys.exit(pytest.main(["-v", "-s", __file__]))

python/ray/serve/tests/test_deploy.py

@@ -688,12 +688,6 @@ def test_deploy_handle_validation(serve_instance):
 
 
 def test_init_args(serve_instance):
-    with pytest.raises(TypeError):
-
-        @serve.deployment(init_args=[1, 2, 3])
-        class BadInitArgs:
-            pass
-
     @serve.deployment(init_args=(1, 2, 3))
     class D:
         def __init__(self, *args):
@@ -840,7 +834,7 @@ def test_input_validation():
 
     with pytest.raises(TypeError):
 
-        @serve.deployment(init_args=[1, 2, 3])
+        @serve.deployment(init_args={1, 2, 3})
         class BadInitArgs:
             pass