我试图将数据库表读入一个dask数据帧，然后保持数据帧。我尝试了一些变体，它们要么导致内存不足，要么出错。
我使用的是具有8 GB内存的Windows 10笔记本电脑。当我尝试读取大型MySQL或Oracle数据库表时，问题就出现了。我可以用SQLite重现这个问题。
以下是设置700 MB SQLite表以重现该问题的代码。 （请原谅python代码中的任何笨拙 - 我已经成为SAS数据分析师10年了。我正在寻找一种更便宜的替代方案，所以我是python，numpy，pandas和dask的新手。注意SAS可以阅读在SQLite表中，将其写入磁盘，并在90秒内创建索引，而无需锁定笔记本电脑。）
import numpy as np
from sqlalchemy import create_engine, MetaData, Table, Column, Integer
import sqlite3
# function to create a list of dicts with chunkSize rows by 3columns
# except for the first column, the columns are filled with random integers
def getChunk(chunkSize, prevCount):
x = np.random.randint(low=0, high=2**32, size=(chunkSize,3), dtype='int64')
y = x.ravel().view(dtype=[('a', 'i8'), ('b', 'i8'), ('c', 'i8')])
y['a'] = [k + prevCount for k in range(chunkSize)]
names = y.dtype.names
listOfDicts = [dict(zip(names, row)) for row in y]
return listOfDicts
# set up a SQLAlchemy engine to a sqlite DB
dbPath = "C:\\temp2\\test.db"
connString = "sqlite:///{}".format(dbPath)
engine = create_engine(connString)
# create a table with 3 Integer columns
metadata = MetaData()
testTable = Table('testTbl', metadata,
Column('a', Integer, primary_key='True'),
Column('b', Integer),
Column('c', Integer)
)
metadata.create_all(engine)
conn = engine.connect()
chunkSize = 25000
numberChunks = 1400
sqlite3.register_adapter(np.int64, lambda x: int(x))
# use the SQLAlchemy table insert method to load list of dicts into the table, one chunk at a time
prevCount = 0
with conn.begin():
for i in range(0, numberChunks) :
listOfDicts = getChunk(chunkSize, prevCount)
conn.execute(testTable.insert(), listOfDicts)
prevCount = prevCount + chunkSize
conn.close()
我在dask调度程序上尝试了4种变体：
每个变体的代码如下。我怎样才能做到这一点？
1.
# default scheduler -- OOM
import dask.dataframe as ddf
from dask.distributed import Client
import dask
import chest
cache = chest.Chest(path='c:\\temp2', available_memory=8e9)
dask.set_options(cache=cache)
dbPath = "C:\\temp2\\test.db"
connString = "sqlite:///{}".format(dbPath)
df = ddf.read_sql_table('testTbl', connString, index_col = 'a')
df = df.persist()
import dask.dataframe as ddf
from dask.distributed import Client
import dask
import chest
cache = chest.Chest(path='c:\\temp2', available_memory=8e9)
dask.set_options(cache=cache)
client = Client()
dbPath = "C:\\temp2\\test.db"
connString = "sqlite:///{}".format(dbPath)
df = ddf.read_sql_table('testTbl', connString, index_col = 'a')
df = client.persist(df)
例外开头是这样的：
>>> tornado.application - ERROR - Exception in callback <bound method Nanny.memory_monitor of <Nanny: tcp://127.0.0.1:57522, threads: 1>>
Traceback (most recent call last):
File "C:\Program Files\Python36\lib\site-packages\psutil\_pswindows.py", line 635, in wrapper
return fun(self, *args, **kwargs)
File "C:\Program Files\Python36\lib\site-packages\psutil\_pswindows.py", line 821, in create_time
return cext.proc_create_time(self.pid)
ProcessLookupError: [Errno 3] No such process
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "C:\Program Files\Python36\lib\site-packages\psutil\__init__.py", line 368, in _init
self.create_time()
File "C:\Program Files\Python36\lib\site-packages\psutil\__init__.py", line 699, in create_time
self._create_time = self._proc.create_time()
File "C:\Program Files\Python36\lib\site-packages\psutil\_pswindows.py", line 640, in wrapper
raise NoSuchProcess(self.pid, self._name)
psutil._exceptions.NoSuchProcess: psutil.NoSuchProcess process no longer exists (pid=14212)
import dask.dataframe as ddf
from dask.distributed import Client
import dask
import chest
cache = chest.Chest(path='c:\\temp2', available_memory=8e9)
dask.set_options(cache=cache, get=dask.get)
client = Client(processes=False)
dbPath = "C:\\temp2\\test.db"
connString = "sqlite:///{}".format(dbPath)
df = ddf.read_sql_table('testTbl', connString, index_col = 'a')
df = client.persist(df, get=dask.get)
一个命令行：c：> dask-scheduler --host 127.0.0.1
另一个命令行：c：> dask-worker 127.0.0.1:8786 --nprocs 1 --nthreads 1 --name worker-1 --memory-limit 3GB --local-directory c：\ temp2
import dask.dataframe as ddf
from dask.distributed import Client
import dask
import chest
cache = chest.Chest(path='c:\\temp2', available_memory=8e9)
dask.set_options(cache=cache)
client = Client(address="127.0.0.1:8786")
dbPath = "C:\\temp2\\test.db"
connString = "sqlite:///{}".format(dbPath)
df = ddf.read_sql_table('testTbl', connString, index_col = 'a')
df = client.persist(df)
这些消息一遍又一遍地杀死了这名工人：
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.12 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.16 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.24 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.31 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.39 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Worker is at 81% memory usage. Pausing worker. Process memory: 2.46 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.47 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.54 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.61 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.66 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.73 GB -- Worker memory limit: 3.00 GB
distributed.worker - WARNING - Memory use is high but worker has no data to store to disk. Perhaps some other process is leaking memory? Process memory: 2.81 GB -- Worker memory limit: 3.00 GB
distributed.nanny - WARNING - Worker exceeded 95% memory budget. Restarting
distributed.nanny - WARNING - Worker process 17916 was killed by signal 15
distributed.nanny - WARNING - Restarting worker
我不相信你在列'a'上有一个索引，这意味着每个分区访问可能在扫描表时在sqlite中使用了大量内存。无论如何，大熊猫通过sqlalchemy访问数据库并不是特别节省内存，因此在访问期间出现内存峰值时我并不感到惊讶。
但是，您可以增加能够访问数据的分区数。例如：
df = ddf.read_sql_table('testTbl', connString, index_col = 'a', npartitions=20)
或者可能减少可用的线程/进程数，以便每个线程有更多的内存。
请注意，chest在这里根本没有帮助，它只能保存已完成的结果，并且在加载数据期间会发生内存峰值（此外，分布式工作程序应该溢出到光盘而不显式提供缓存）。