Continuous deployment is core to Etsy, and we push public facing code over 30 times per day. Evaluating an experiment amidst this shifting landscape is a difficult task as our traditional methods of monitoring operational metrics don’t provide enough information to make product-level decisions. To this end, we have developed internal tooling for deep analytics that enables us to systematically analyze our experimental results in a continuously changing environment.
This talk will focus on the analysis framework that we have built – from the raw logging data, to our elastic mapreduce-based data transformations, to the final dashboards and underlying statistics that drive decision making.
15. Most Times it’s Not
Cart Page
Restyled
Monday, November 5, 2012
16. Most Times it’s Not
Etsy Mentioned
on Oprah Cart Page
Restyled
Justin Beiber Internet Goes
Sneezes Down in Brooklyn
New Cat Meme
Monday, November 5, 2012
31. 3 Types of Metrics
•Co-occurence
•visits that did X (searched)
•average number of X (searches)
•Funnels
•Special Cases*
Monday, November 5, 2012
32. Co-Occurence
•Co-occurence of all actions with the variant
•i.e. - how many people saw this variant and
then purchased
•i.e. - average number of purchases per variant
Monday, November 5, 2012
33. Funnels
Search Listing Buy
Buy
Monday, November 5, 2012
34. Funnels
Search Listing Buy
Buy
Home Search Listing BuyBuy
Monday, November 5, 2012
35. Funnels
Search Listing Buy
Buy
Home Search Listing Buy
BuyBuy
Monday, November 5, 2012
36. Special Cases
•Exiting pages
•Bounce rate
•Visit length
Monday, November 5, 2012
37. Pipeline
Logs A/B
COLLECTION
Client HDFS Events Visits MySql
PROCESSING
DB
Monday, November 5, 2012
38. Bucketing
Home Search Listing Buy
Search A/B Test
Monday, November 5, 2012
39. Bucketing
Home Search Listing Buy
Search A/B Test
Monday, November 5, 2012
40. Process
•Idea
•Hypothesis
•Deploy
•Wait
•Analysis
Monday, November 5, 2012
41. Process
•Idea
•Hypothesis
•Deploy
•Wait
•Analysis
Monday, November 5, 2012
62. Final Thoughts
•Bet driven development - be explicit about your
changes
•Easy experimental setup and quick analysis
makes for more experiments
•An experiment is successful if you learned
something, not if you launched
Monday, November 5, 2012