Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Etsy on Migrating to Kafka (in three short years)
Search
Hakka Labs
January 22, 2015
Programming
4
6k
Etsy on Migrating to Kafka (in three short years)
Full post with video here:
Hakka Labs
January 22, 2015
Tweet
Share
More Decks by Hakka Labs
See All by Hakka Labs
New Workflows for Building Data Pipelines
hakka_labs
0
2.9k
Collaborative Topic Models for Users and Texts
hakka_labs
0
2.8k
Groupcache with Evan Owen
hakka_labs
2
5.4k
Testing Android at Spotify
hakka_labs
1
4.5k
It's Not a Bug, It's a Feature!
hakka_labs
0
3.2k
K-means Clustering to Understand Your Users
hakka_labs
0
2k
Building Amy: The Email-based Virtual Assistant by x.ai
hakka_labs
0
5k
Deep Learning and NLP Applications
hakka_labs
3
13k
Go and the Gophers
hakka_labs
2
11k
Other Decks in Programming
See All in Programming
250830 IaCの選定~AWS SAMのLambdaをECSに乗り換えたときの備忘録~
east_takumi
0
400
MCPでVibe Working。そして、結局はContext Eng(略)/ Working with Vibe on MCP And Context Eng
rkaga
5
2.3k
testingを眺める
matumoto
1
140
JSONataを使ってみよう Step Functionsが楽しくなる実践テクニック #devio2025
dafujii
1
650
RDoc meets YARD
okuramasafumi
4
170
Performance for Conversion! 分散トレーシングでボトルネックを 特定せよ
inetand
0
3.4k
AI Agents: How Do They Work and How to Build Them @ Shift 2025
slobodan
0
110
パッケージ設計の黒魔術/Kyoto.go#63
lufia
3
440
より安全で効率的な Go コードへ: Protocol Buffers Opaque API の導入
shwatanap
2
820
Updates on MLS on Ruby (and maybe more)
sylph01
1
180
「待たせ上手」なスケルトンスクリーン、 そのUXの裏側
teamlab
PRO
0
570
print("Hello, World")
eddie
2
530
Featured
See All Featured
Visualization
eitanlees
148
16k
Creating an realtime collaboration tool: Agile Flush - .NET Oxford
marcduiker
31
2.2k
No one is an island. Learnings from fostering a developers community.
thoeni
21
3.4k
Into the Great Unknown - MozCon
thekraken
40
2k
Agile that works and the tools we love
rasmusluckow
330
21k
How STYLIGHT went responsive
nonsquared
100
5.8k
Facilitating Awesome Meetings
lara
55
6.5k
What’s in a name? Adding method to the madness
productmarketing
PRO
23
3.7k
Building Better People: How to give real-time feedback that sticks.
wjessup
368
19k
Designing for humans not robots
tammielis
253
25k
Sharpening the Axe: The Primacy of Toolmaking
bcantrill
44
2.5k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
33
2.4k
Transcript
Migrating to Kafka in Three Short Years A look at
the choices that defined the Etsy analytics stack
None
Path Dependence
Decisions made in the past limit options in the present,
even if the circumstances under which those past decisions were made are no longer relevant.
In other words, we can’t upgrade the Hadoop cluster until
we port all of the Cascading.jruby jobs to Scalding.
Sneak Preview ! 1. How Etsy built its original analytics
stack 2. Handling changes prepared us to rebuild our data pipeline 3. Kafka!
Starting from scratch
Choice #1 ! Acquire Adtuitive
None
None
Before you can work on search, you need real analytics
Choice #2 ! Build a zero-impact analytics stack
Etsy is not a cloud company but the first analytics
stack was cloud-based
(illustration here) browser CDN EMR S3 mysql FTP
Legacy effects: ! 24 hour latency on events 48 hour
latency on visits
Choice #3 ! Cascading.jruby
Hadoop Cascading Cascading.jruby
Choice #4 ! Use GA _utma cookie to define visits
Benefits: ! •Simpler ETL •Visits computed on the client side
•Easy to reconcile against Google Analytics
Choice #5 ! Using existing feature library for A/B tests
Leveraged existing experience with operational ramp-ups
Low impact: just required a logging change
Choice #6 ! Build analytics stack around visit-level metrics
Great for search and ads, less great for measuring engagement
Changing the tires without stopping the car
How do we instrument the iOS app? Summer 2012
1. Native app visits should have the same structure as
Web visits
2. Native app events should use the existing data pipeline
3. The native app should buffer events and send them
when convenient
Solution: ! 1. App uploads bundles of events to API
endpoint 2. Backend event logger curls the beacon for every event
Side effect: ! We have a backend event logger that
is now used all over the place
CDN diversification project Fall 2012
None
Migrated to our own beacon infrastructure
Data pipeline based on Apache, PHP, logrotate, and cron
We built our own Hadoop cluster: Etsydoop Fall 2012
We hired the Scalding guy Fall 2012
Hadoop Cascading Cascading.jruby Scalding
None
Uh oh, the Google Analytics JS hurts performance Fall 2012
The event logger’s GA dependency precluded async loading, hurting performance
First idea: duplicate the _utma functionality in our own code
The trouble with backend events
Visit Time Logger Event Type 1 12:01 frontend home 1
12:03 backend login 1 12:03 frontend view listing 1 1:31 backend logout 2 1:31 frontend view listing 2 1:32 frontend search 2 1:33 frontend view listing wrong visit
Complete rewrite of our ETL jobs Spring/Summer 2013
None
Backend page-view events Fall 2013
None
2014: the next phase
EventPipe goals
Use POST rather than multiple GET requests to prevent data
loss
Use JSON rather than query strings for comprehensibility
Validate beacon data before it enters the data pipeline
Use a binary serialization format for long-term storage
Use Kafka for data transfer to escape the batch paradigm
Eliminate individual beacon servers as points of failure
How do we handle the impedance mismatch between Apache/PHP and
Kafka?
Wrote a server in Go to serialize beacons in Thrift
and send them to Kafka
Use Apache for SSL termination
Still to come
Real-ish time ETL
Streaming infrastructure
Offline processing for more products
Other Kafka applications
Takeaways
Every choice you make has long-term implications
Fixing stuff creates new opportunities
@rafeco http://rc3.org