Question

.I’m trying to speed up an insert into Cassandra using pycassa.
I heard that using multithreading and opening multiple connections speeds it up a lot. I’m inserting a load of tweets in json format.
My code here works, for a bit, and then the threads start throwing exceptions and it stops, it seems the more threads I have the faster it stops working… I’m guessing that the problem is the connections to cassandra, something to do with connection pooling. Any ideas?

EDIT: all threads throw “Exception in thread Thread-3 (most likely raised during interpreter shutdown):”

import time
import pycassa
from pycassa.pool import ConnectionPool
from pycassa.columnfamily import ColumnFamily
from datetime import datetime 
import json
import threadpool
pool = threadpool.ThreadPool(4)
kspool = ConnectionPool('TweetsKS',use_threadlocal = True)

def process_tasks(lines):

    #let threadpool format your requests into a list
    requests = threadpool.makeRequests(insert_into_cfs, lines)

    #insert the requests into the threadpool
    for req in requests:
        pool.putRequest(req) 


def read(file):
    bench = open("bench.txt", "w")
    bench.write(str(datetime.now())+"\n")
    """read data from json and insert into keyspace"""
    json_data=open(file)
    lines = []
    for line in json_data:
        lines.append(line)
    process_tasks(lines)


def insert_into_cfs(line):

    user_tweet_cf = pycassa.ColumnFamily(kspool, 'UserTweet')
    user_name_cf = pycassa.ColumnFamily(kspool, 'UserName')
    tweet_cf = pycassa.ColumnFamily(kspool, 'Tweet')
    user_follower_cf = pycassa.ColumnFamily(kspool, 'UserFollower')

    tweet_data = json.loads(line)
    """Format the tweet time as an epoch seconds int value"""
    tweet_time = time.strptime(tweet_data['created_at'],"%a, %d %b %Y %H:%M:%S +0000")
    tweet_time  = int(time.mktime(tweet_time))

    new_user_tweet(user_tweet_cf,tweet_data['from_user_id'],tweet_time,tweet_data['id'])
    new_user_name(user_name_cf,tweet_data['from_user_id'],tweet_data['from_user_name'])
    new_tweet(tweet_cf,tweet_data['id'],tweet_data['text'],tweet_data['to_user_id'])

    if tweet_data['to_user_id'] != 0:
        new_user_follower(user_follower_cf,tweet_data['from_user_id'],tweet_data['to_user_id'])


"""4 functions below carry out the inserts into specific column families"""     
def new_user_tweet(user_tweet_cf,from_user_id,tweet_time,id):
    user_tweet_cf.insert(from_user_id,{(tweet_time): id})

def new_user_name(user_name_cf,from_user_id,user_name):
    user_name_cf.insert(from_user_id,{'username': user_name})

def new_tweet(tweet_cf,id,text,to_user_id):
    tweet_cf.insert(id,{
    'text': text
    ,'to_user_id': to_user_id
    })  

def new_user_follower(user_follower_cf,from_user_id,to_user_id):
    user_follower_cf.insert(from_user_id,{to_user_id: 0})   

if __name__ == '__main__':
    read('tweets.json')

Answer 1

Ok, the issue here was with my use of threadpool.
I needed pool.wait after pool.putRequest(req) (outside the loop)
My main thread was finishing before the rest of them and they were not daemons.

With 2 threads my Cassandra insert is about twice as fast… but guess what!? It is still slower than MySQL!!
With 6 threads it is about the same… more tinkering is needed I guess!