Python multiprocessing.Pool & memory

问题:

I'm using Pool.map for a scoring procedure:

1. "cursor" with millions of arrays from a data source
2. calculation
3. save the result in a data sink

The results are independent.

I'm just wondering if I can avoid the memory demand. At first it seems that every array goes into python and then the 2 and 3 are proceed. Anyway I have a speed improvement.

#data src and sink is in mongodb#
def scoring(some_arguments):
        ### some stuff  and finally persist  ###
    collection.update({uid:_uid},{'$set':res_profile},upsert=True)


cursor = tracking.find(timeout=False)
score_proc_pool = Pool(options.cores)    
#finaly I use a wrapper so I have only the document as input for map
score_proc_pool.map(scoring_wrapper,cursor,chunksize=10000)

Am I doing something wrong or is there a better way with python for this purpose?

回答1:

The map functions of a Pool internally convert the iterable to a list if it doesn't have a __len__ attribute. The relevant code is in Pool.map_async, as that is used by Pool.map (and starmap) to produce the result - which is also a list.

If you don't want to read all data into memory first, you should use Pool.imap or Pool.imap_unordered, which will produce an iterator that will yield the results as they come in.