如何配置Ray以在多进程中使用标准Python Logger？

尝试用Ray提升进程运行速度，希望将日志消息传递给标准Python Logger，实现日志的格式化、过滤与存储。但使用Ray时，日志未按自定义Logger配置格式化，也未传递至根Logger。已在ray.init()中设置log_to_driver=True和configure_logging=True，问题仍存在。

复现代码 #

from ray.util.multiprocessing import Pool
import pathlib
import logging
import json

def setup_logging(config_file: pathlib.Path):
    with open(config_file) as f_in:
        config = json.load(f_in)
    logging.config.dictConfig(config)

logger = logging.getLogger(__name__)
config_file = pathlib.Path(__file__).parent / "log_setup/config_logging.json"
setup_logging(config_file=config_file)

def f(index):
    logger.warning(f"index: {index}")
    return (index, "model")

if __name__ == "__main__":
    logger.warning("Starting")
    pool = Pool(1)
    results = pool.map(f, range(10))
    print(list(results))

Logger配置文件（log_setup/config_logging.json） #

{
    "version": 1,
    "disable_existing_loggers": false,
    "formatters": {
      "detailed": {
        "format": "[%(levelname)s|%(name)s|%(module)s|L%(lineno)d] %(asctime)s: %(message)s",
        "datefmt": "%Y-%m-%dT%H:%M:%S%z"
      }
    },
    "handlers": {
      "stdout": {
        "class": "logging.StreamHandler",
        "level": "INFO",
        "formatter": "detailed"
      }
    },
    "loggers": {
      "root": {
        "level": "DEBUG",
        "handlers": [
          "stdout"
        ]
      }
    }
  }

输出对比 #

Python原生map输出 #

[WARNING|__main__|ray_trial|L28] 2024-03-27T15:14:21+0100: Starting
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 0
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 1
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 2
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 3
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 4
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 5
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 6
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 7
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 8
[WARNING|__main__|ray_trial|L22] 2024-03-27T15:14:21+0100: index: 9

Ray Pool输出 #

2024-03-27 14:54:55,064 INFO worker.py:1743 -- Started a local Ray instance. View the dashboard at 127.0.0.1:8265 
(PoolActor pid=43261) index: 0
(PoolActor pid=43261) index: 1
(PoolActor pid=43261) index: 2
(PoolActor pid=43261) index: 3
(PoolActor pid=43261) index: 4
(PoolActor pid=43261) index: 5
(PoolActor pid=43261) index: 6
(PoolActor pid=43261) index: 7
(PoolActor pid=43261) index: 8
(PoolActor pid=43261) index: 9

Ray的进程（包括Pool中的Actor）是独立的Python进程，主进程的日志配置不会自动传递到子进程，且Ray默认会覆盖部分日志设置，需从以下几个方面调整：

1. 在Ray子进程中重新初始化日志配置 #

Ray启动的子进程不会继承主进程的日志配置，需在任务函数内部重新加载日志配置：
修改任务函数f：

def f(index):
    # 子进程中重新初始化日志
    config_file = pathlib.Path(__file__).resolve().parent / "log_setup/config_logging.json"
    setup_logging(config_file=config_file)
    logger = logging.getLogger(__name__)  # 重新获取Logger实例
    logger.warning(f"index: {index}")
    return (index, "model")

2. 禁用Ray的默认日志配置 #

在初始化Ray时，关闭自动日志配置，避免Ray覆盖自定义设置：

if __name__ == "__main__":
    import ray
    # 禁用Ray默认日志配置，避免干扰自定义Logger
    ray.init(configure_logging=False, log_to_driver=False)
    
    logger.warning("Starting")
    pool = Pool(1)
    results = pool.map(f, range(10))
    print(list(results))

log_to_driver=False会关闭Ray默认的日志转发，确保日志完全由自定义Logger处理。

3. 使用Ray日志适配器保留进程标识 #

如果需要保留Ray的进程标识（如(PoolActor pid=xxx)），同时应用自定义格式，可以使用RayLoggerAdapter包装Logger：

from ray.util.log import RayLoggerAdapter

def f(index):
    config_file = pathlib.Path(__file__).resolve().parent / "log_setup/config_logging.json"
    setup_logging(config_file=config_file)
    logger = logging.getLogger(__name__)
    # 包装成Ray日志适配器，保留进程信息
    ray_logger = RayLoggerAdapter(logger, {"actor_name": "PoolActor"})
    ray_logger.warning(f"index: {index}")
    return (index, "model")

内容的提问来源于stack exchange，提问作者gabrown86