部署在线推理服务（REST API）

%pip install "wedata3-mlflow-tcdeploy-plugin"

import mlflow_tcdeploy_plugin; print(mlflow_tcdeploy_plugin.__version__)  # 应为 0.1.1
from mlflow.deployments import get_deploy_client

# 推荐统一通过官方 API 获取客户端，target 固定为 "tcdeploy"
client = get_deploy_client("tcdeploy")

models:/<catalog_name>.<schema_name>.<model_name>/<version>

resp = client.create_deployment(
    name="churn-model-svc",                 # 服务名称（ServiceName）
    model_uri="models:/<catalog_name>.<schema_name>.<model_name>/<version>",
    config={
        "instance_type": "<instance_spec>",  # 必填：实例规格名，用 list_instance_types() 查询
        "replicas": 2,                       # 副本数，默认 1
        "charge_type": "POSTPAID_BY_HOUR",   # 计费模式，默认按量计费，可选 PREPAID
        "scale_mode": "MANUAL",              # 弹性模式，默认 MANUAL，可选 AUTO
        "authorization_enable": True,        # 是否开启鉴权 Token，生产建议开启
        "log_enable": True,                  # 是否开启日志投递，生产建议开启
        "service_description": "用户流失预测在线服务",
    },
)
service_id       = resp["data"]["service_id"]        # 形如 ms-xxxxxx
service_group_id = resp["data"]["service_group_id"]  # 形如 msg-xxxxxx

# 查询可用规格；spec_type 可选 "CPU" / "GPU"，不传则全部
cpu = client.list_instance_types(spec_type="CPU")
print(len(cpu), cpu[:2])
print(client.list_instance_types(spec_type="GPU")[:2])
print(client.get_instance_type_prices([s["spec_name"] for s in cpu if s["available"]][:2]))

import time

TERMINAL = {"RUNNING", "FAILED", "STOPPED"}
deadline = time.time() + 600           # 10 分钟超时上限，避免长时间阻塞

while True:
    detail = client.get_deployment(service_id)
    status = detail["data"].get("status")
    print("status =", status)
    if status in TERMINAL:
        break
    if time.time() > deadline:
        raise TimeoutError(f"service {service_id} not ready, status={status}")
    time.sleep(15)                     # 轮询间隔 ≥ 10 秒，避免触发频率限制

result = client.predict(
    deployment_name=service_group_id,          # 注意：是 service_group_id，不是 service_id
    inputs={"dataframe_records": [{"f1": 1.0, "f2": 2.0}]},
    config={"auth_token": "Bearer <token>"},    # 服务开启鉴权时必填
)
print(result)

# 1) 调整副本数（普通更新，不重建服务）
client.update_deployment(name=service_id, config={"replicas": 4})

# 2) 切换到新模型版本（自动滚动重启，无需先下线）
client.update_deployment(
    name=service_id,
    model_uri="models:/<catalog_name>.<schema_name>.<model_name>/<new_version>",
)

# 3) 停止 / 恢复 / 触发扩缩容（service_action 特殊语义）
client.update_deployment(name=service_id, config={"service_action": "STOP"})
client.update_deployment(name=service_id, config={"service_action": "RESUME"})
client.update_deployment(name=service_id, config={"service_action": "SCALE"})

# 列举当前工作空间下的服务组；传 endpoint 可按 ServiceGroupId 过滤
groups = client.list_deployments()
for g in groups:
    print(g["name"], g.get("service_group_name"), g.get("service_group_status"))

# 删除服务（幂等：服务不存在时不会报错）
client.delete_deployment(service_id)