This talk was prepared for the November 2013 DataPhilly Meetup: Data in Practice ( http://www.meetup.com/DataPhilly/events/149515412/ ) Map Reduce: Beyond Word Count by Jeff Patti Have you ever wondered what map reduce can be used for beyond the word count example you see in all the introductory articles about map reduce? Using Python and mrjob, this talk will cover a few simple map reduce algorithms that in part power Monetate's information pipeline Bio: Jeff Patti is a backend engineer at Monetate with a passion for algorithms, big data, and long walks on the beach. Prior to working at Monetate he performed software R&D for Lockheed Martin, where he worked on projects ranging from social network analysis to robotics.

1. MapReduce:
Beyond Word Count
Jeff Patti
https://github.com/jepatti/mrjob_recipes

2. What is MapReduce?
“MapReduce is a programming model for processing large
data sets with a parallel, distributed algorithm on a cluster.”
- Wikipedia
Map - given a line of a file, yield key: value pairs
Reduce - given a key and all values with that key from the
prior map phase, yield key: value pairs

3. Word Count
Problem: count frequencies of words in
documents

4. Word Count Using mrjob
def mapper(self, key, line):
for word in line.split():
yield word, 1
def reducer(self, word, occurrences):
yield word, sum(occurrences)

5. Sample Output
"ligula" 4
"ligula." 2
"lorem" 5
"lorem." 4
"luctus" 3
"magna" 5
"magna," 3
"magnis" 1

6. Monetate Background
● Core products are merchandising,
personalization, testing, etc.
● A/B & Multivariate testing to determine
impact of experiments
● Involved with >20% of ecommerce spend
each holiday season for the past 2 years
running

7. Monetate Stack
● Distributed across multiple availability zones
and regions for redundancy, scaling, and
lower round trip times
● Real time decision engine using MySQL
● Nightly processing of each days data via
Hadoop using mrjob, a python library for
writing mapreduce jobs

8. Beyond Word Count
● Activity stream sessionization
● Product recommendations
● User behavior statistics

9. Activity Stream Sessionization
Goal: collate user activity, splitting into different
sessions if user inactive for more than 5
minutes
Input format: timestamp, user_id

10. Collate user activity
def mapper(self, key, line):
timestamp, user_id = line.split()
yield user_id, timestamp
def reducer(self, uid, timestamps):
yield uid, sorted(timestamps)

11. Sample Output
"998" ["1384389407", "1384389417", "1384389422",
"1384389425", "1384390407", "1384390417",
"1384391416", "1384392410", "1384392416",
"1384395420", "1384396405"]
"999" ["1384388414", "1384388425", "1384389419",
"1384389420", "1384390420", "1384391415",
"1384391418", "1384393413", "1384393425",
"1384394426", "1384395416", "1384396415",
"1384396422"]

12. Segment into Sessions
MAX_SESSION_INACTIVITY = 60 * 5
...
def reducer(self, uid, timestamps):
timestamps = sorted(timestamps)
start_index = 0
for index, timestamp in enumerate(timestamps):
if index > 0:
if timestamp - timestamps[index-1] >
MAX_SESSION_INACTIVITY:
yield uid, timestamps[start_index:index]
start_index = index
yield uid, timestamps[start_index:]

13. Sample Output
"999"[1384388414, 1384388425]
"999"[1384389419, 1384389420]
"999"[1384390420]
"999"[1384391415, 1384391418]
"999"[1384393413, 1384393425]
"999"[1384394426]
"999"[1384395416]
"999"[1384396415, 1384396422]

14. Product Recommendations
Goal: For each product a client sells, generate
a ‘people who bought this also bought this’
recommendation
Input: product_id_1, product_id_2, ...

15. Coincident Purchase Frequency
def mapper(self, key, line):
purchases = set(line.split(','))
for p1, p2 in permutations(purchases, 2):
yield (p1, p2), 1
def reducer(self, pair, occurrences):
p1, p2 = pair
yield p1, (p2, sum(occurrences))

16. Sample output
"8" ["5", 11]
"8" ["6", 19]
"8" ["7", 14]
"8" ["9", 11]
"9" ["1", 20]
"9" ["10", 22]
"9" ["11", 21]
"9" ["12", 13]

17. Top Recommendations
def reducer(self, purchase_pair, occurrences):
p1, p2 = purchase_pair
yield p1, (sum(occurrences), p2)
def reducer_find_best_recos(self, p1, p2_occurrences):
top_products = sorted(p2_occurrences, reverse=True)[:5]
top_products = [p2 for occurrences, p2 in top_products]
yield p1, top_products
def steps(self):
return [self.mr(mapper=self.mapper, reducer=self.reducer),
self.mr(reducer=self.reducer_find_best_recos)]

18. Sample Output
"7"
"8"
"9"
["15", "18", "17", "16", "3"]
["14", "15", "20", "6", "3"]
["15", "17", "19", "6", "3"]

19. Top Recommendations
Multi Account
def mapper(self, key, line):
account_id, purchases = line.split()
purchases = set(purchases.split(','))
for p1, p2 in permutations(purchases, 2):
yield (account_id, p1, p2), 1
def reducer(self, purchase_pair, occurrences):
account_id, p1, p2 = purchase_pair
yield (account_id, p1), (sum(occurrences), p2)
2nd step reducer unchanged

20. Sample Output
["9", "20"]
["9", "3"]
["9", "4"]
["9", "5"]
["9", "6"]
["9", "7"]
["9", "8"]
["9", "9"]
["8", "14", "13", "10", "1"]
["2", "4", "16", "11", "17"]
["3", "18", "11", "16", "15"]
["2", "1", "7", "18", "17"]
["12", "3", "2", "17", "16"]
["18", "5", "17", "1", "9"]
["20", "14", "13", "10", "4"]
["18", "7", "6", "5", "4"]

21. User Behavior Statistics
Goal: compute statistics about user behavior
(conversion rate & time on site) by account and
experiment in an efficient manner
Input:
account_id, campaigns_viewed, user_id, purchased?,
session_start_time, session_end_time

22. Statistics Primer
With sample count, mean, and variance for
each side of an experiment we can compute all
the statistics our analytics package displays

23. Statistics Primer (cont.)
y = a sessions metric value, ex: time on site
● Sample count: count the number of sessions
that viewed the experiment
○ sum(y^0)
● Mean: sum the metric / sample count
○ sum(y^1)/sum(y^0)

24. Statistics Primer (cont.)
● Variance:
○ Variance = mean of square minus square of mean
○ Variance = sum(y^2)/sum(y^0) - (sum(y^1)/sum(y^0)) ^ 2
For each side of an experiment we only need to
generate: sum(y^0), sum(y^1), sum(y^2)

25. Statistics by account
statistic_rollup/statistic_summarize.py

26. Sample Output
["8", "average session length"] [99, 24463, 7968891]
["8", "conversion rate"] [99, 45, 45]
["9", "average session length"] [115, 29515, 10071591]
["9", "conversion rate"] [115, 55, 55]

27. Statistics by experiment
statistic_rollup_by_experiment/statistic_summa
rize.py

28. Sample Output
["9", 0, "average session length"] [32, 8405, 3031009]
["9", 0, "conversion rate"] [32, 20, 20]
["9", 1, "average session length"] [23, 5405, 1770785]
["9", 1, "conversion rate"] [23, 14, 14]
["9", 2, "average session length"] [39, 9481, 2965651]
["9", 2, "conversion rate"] [39, 20, 20]
["9", 3, "average session length"] [25, 6276, 2151014]
["9", 3, "conversion rate"] [25, 13, 13]
["9", 4, "average session length"] [27, 5721, 1797715]
["9", 4, "conversion rate"] [27, 16, 16]

29. Questions?
?