Daskdistributed.scheduler - 错误 - 无法收集密钥

Question

import joblib

from sklearn.externals.joblib import parallel_backend
with joblib.parallel_backend('dask'):
 
    from dask_ml.model_selection import GridSearchCV
    import xgboost
    from xgboost import XGBRegressor
    grid_search = GridSearchCV(estimator= XGBRegressor(), param_grid = param_grid, cv = 3, n_jobs = -1)
    grid_search.fit(df2,df3)

我使用两台本地机器创建了一个 dask 集群

client = dask.distributed.client('tcp://191.xxx.xx.xxx:8786')

我正在尝试使用 dask gridsearchcv 找到最佳参数。我面临以下错误。

istributed.scheduler - ERROR - Couldn't gather keys {"('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 1202, 2)": ['tcp://127.0.0.1:3738']} state: ['processing'] workers: ['tcp://127.0.0.1:3738']
NoneType: None
distributed.scheduler - ERROR - Workers don't have promised key: ['tcp://127.0.0.1:3738'], ('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 1202, 2)
NoneType: None
distributed.client - WARNING - Couldn't gather 1 keys, rescheduling {"('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 1202, 2)": ('tcp://127.0.0.1:3738',)}
distributed.nanny - WARNING - Restarting worker
distributed.scheduler - ERROR - Couldn't gather keys {"('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 1, 2)": ['tcp://127.0.0.1:3730']} state: ['processing'] workers: ['tcp://127.0.0.1:3730']
NoneType: None
distributed.scheduler - ERROR - Couldn't gather keys {"('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 0, 1)": ['tcp://127.0.0.1:3730'], "('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 5, 1)": ['tcp://127.0.0.1:3729'], "('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 4, 2)": ['tcp://127.0.0.1:3729'], "('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 2, 1)": ['tcp://127.0.0.1:3730']} state: ['processing', 'processing', 'processing', 'processing'] workers: ['tcp://127.0.0.1:3730', 'tcp://127.0.0.1:3729']
NoneType: None
distributed.scheduler - ERROR - Couldn't gather keys {'cv-n-samples-7cb7087b3aff75a31f487cfe5a9cedb0': ['tcp://127.0.0.1:3729']} state: ['processing'] workers: ['tcp://127.0.0.1:3729']
NoneType: None
distributed.scheduler - ERROR - Couldn't gather keys {"('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 4, 0)": ['tcp://127.0.0.1:3729'], "('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 2, 0)": ['tcp://127.0.0.1:3729'], "('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 0, 0)": ['tcp://127.0.0.1:3729']} state: ['processing', 'processing', 'processing'] workers: ['tcp://127.0.0.1:3729']
NoneType: None
distributed.scheduler - ERROR - Couldn't gather keys {"('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 0, 2)": ['tcp://127.0.0.1:3729'], "('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 2, 2)": ['tcp://127.0.0.1:3729']} state: ['processing', 'processing'] workers: ['tcp://127.0.0.1:3729']
NoneType: None
distributed.scheduler - ERROR - Workers don't have promised key: ['tcp://127.0.0.1:3730'], ('xgbregressor-fit-score-7cb7087b3aff75a31f487cfe5a9cedb0', 1, 2)
NoneType: None

我希望有人帮助解决这个问题。提前致谢。

Answer 1

我也遇到了同样的问题，我发现这可能是防火墙引起的。

假设我们有两台机器，191.168.1.1 用于调度程序，191.168.1.2 用于工作程序。

当我们启动调度程序时，我们可能会得到以下信息：

distributed.scheduler - INFO - -----------------------------------------------
distributed.http.proxy - INFO - To route to workers diagnostics web server please install jupyter-server-proxy: python -m pip install jupyter-server-proxy
distributed.scheduler - INFO - -----------------------------------------------
distributed.scheduler - INFO - Clear task state
distributed.scheduler - INFO -   Scheduler at:  tcp://191.168.1.1:8786
distributed.scheduler - INFO -   dashboard at:                   :8787

所以对于调度程序，我们应该确认port 8786和port 8786可以访问。

同样，我们可以查看工人的信息：

istributed.nanny - INFO -         Start Nanny at: 'tcp://191.168.1.2:39042'
distributed.diskutils - INFO - Found stale lock file and directory '/root/dask-worker-space/worker-39rf_n28', purging
distributed.worker - INFO -       Start worker at:  tcp://191.168.1.2:39040
distributed.worker - INFO -          Listening to:  tcp://191.168.1.2:39040
distributed.worker - INFO -          dashboard at:        191.168.1.2:39041
distributed.worker - INFO - Waiting to connect to:   tcp://191.168.1.1:8786
distributed.worker - INFO - -------------------------------------------------

nanny 端口是39042，worker 端口是39040，dashboard 端口是39041。

将这些端口设置为 191.168.1.1 和 191.168.1.2 打开：

firewall-cmd --permanent --add-port=8786/tcp
firewall-cmd --permanent --add-port=8787/tcp
firewall-cmd --permanent --add-port=39040/tcp
firewall-cmd --permanent --add-port=39041/tcp
firewall-cmd --permanent --add-port=39042/tcp
firewall-cmd --reload

并且任务可以成功运行。

最后，Dask将为worker随机选择端口，我们也可以使用自定义端口启动worker：

dask-worker 191.168.1.1:8786 --worker-port 39040 --dashboard-address 39041 --nanny-port 39042

更多参数可以参考这里。