Dask调度程序退出，输出为“ddf.persist()”上的“Killed”。

Question

我对DASK还是个新手，这一点很明显……我正在尝试运行一个分布式dask设置，其中1个节点用于调度器，并且有足够的工作节点来容纳内存中的数据--在本例中，我使用了15个工作节点。我可以很好地启动集群，还可以加载一些数据并对其进行分析。

我已经将数据复制到工作节点，但是在我的客户机上没有可用的数据，因此我像这样延迟了数据的加载：

import dask
import dask.dataframe as dd
from dask import delayed

def load_data(path):
    return dd.read_csv(path)

然后我可以做一些简单的分析：

taxi_df_2016 = delayed(load_data)('/tmp/2016/*.csv').compute()
taxi_df_2016['fare_amount'].mean().compute()

..。将返回一个值

但是当我想在内存中持久化文件时，调度器在控制台上打印Killed时就死机了……

taxi_df_2016_pers = taxi_df_2016.compute().persist()

将在调度器节点的控制台上显示以下内容：

distributed.scheduler - INFO - Register tcp://10.0.0.17:40385
distributed.scheduler - INFO - Register tcp://10.0.0.6:42847
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.17:40385
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.6:42847
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.9:44627
distributed.scheduler - INFO - Register tcp://10.0.0.7:44419
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.9:44627
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.7:44419
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.16:41907
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.16:41907
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.18:41879
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.18:41879
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.13:32993
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.13:32993
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.8:33265
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.8:33265
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.14:33851
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.14:33851
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.10:44653
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.10:44653
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.19:40201
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.19:40201
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.5:42207
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.5:42207
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.15:36087
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.15:36087
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.12:32827
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.12:32827
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Register tcp://10.0.0.11:35405
distributed.scheduler - INFO - Starting worker compute stream, tcp://10.0.0.11:35405
distributed.core - INFO - Starting established connection
distributed.scheduler - INFO - Clear task state
Killed

当我查看仪表板时，我看到所有173个分区都被成功加载，内存也被使用了，但是在那之后的某个时间，调度器就死了。

​

​

对如何调试这个有什么想法吗？

Answer 1

看起来你在延迟一个dask dataframe函数，这似乎很奇怪。

请参阅https://docs.dask.org/en/latest/delayed-best-practices.html#don-t-call-dask-delayed-on-other-dask-collections