Using Joyent Manta to Scale Event-based Data Collection and Analysis at Wanelo

Proprietary and Confidential

Presenter: Konstantin GredeskoulCTO, Wanelo.com

Based on work of Atasay Gökkaya and other engineers

"It's a Unix System! I know this!"

Using Manta to Scale Event-based Data Collection and Analysis

@kigster

■ Wanelo (“Wah-nee-lo” from Want, Need Love) is a global platform for all the world’s shopping

■ Users find products on online stores

■ They post these products to Wanelo via, a javascript “bookmarklet”

■ Others discover these products on Wanelo via feed, trending, search, etc

■ Users then save products they discovered to their own collections

How Wanelo Works

■ Users can follow other users. Following is bi-directional, like Twitter, and public

■ Besides following other users, you can follow individual stores on wanelo

■ Result is a personalized shopping feed, much like Twitter’s information feed

■ After seeing a product on Wanelo, users can buy the product on the original site

Wanelo is a Social Network

Mobile: iOS + Android 60K ratings

Backend Stack & Key Vendors

■ MRI Ruby 2.0 & Rails 3

■ PostgreSQL 9.2, solr, redis, memcached, twemproxy, nginx, haproxy

■ Joyent Cloud, SmartOSZFS, ARC Cache, raw IO performance, SMF, Zones, dTrace

■ Joyent Manta: Analytics and Backups

■ Chef, Opscode EnterpriseFull server automation, zero manual installs

■ Images: AWS S3 behind Fastly CDN

■ Circonus, NewRelic, statsd, Boundary

Final word about Wanelo...

We are slightly obsessed with cat pictures =)

Recording User Events: Why?

■ Let’s say user saves a product

■ Naturally we create a row in our main data store (PostgreSQL)

■ But we also want to record this event to an append-only log table, for future analysis

■ In the ideal world, this append-only table has every user-generated event of interest

Hey, What’s the Scale Here?

■ 10M users

■ 7M products saved over 1B times

■ 200K+ stores

■ Backend peaks at 200,000 RPMs

■ Generating between 5M and 20M user events per day

Recording Events: Stupidly

■ We are just starting: what’s the simplest thing we can do? Our traffic is still pretty low.

■ Let’s create a database table and append to that. Simple? Yes.

■ Scalable? Hell No.

■ One month after launch, we hit the wall.

Let’s Scale Data Collection

■ OK, so inserting 10M records into PostgreSQL per day is pretty stupid. Even I know that.

■ We looked around for various options. There were many. Flume, Fluentd, Scribe. Meh.

■ We chose rsyslog: clients can buffer records, send cheap UDP packets.

■ More than one log collector for redundancy

Scaling Event Data Collection

■ rsyslog rocks. We are now sending 20M events per day from 40+ hosts

■ rsyslog is dumping them into an ASCII pipe-delimited file

■ logadm rotates the file daily. We get 1GB+ file per day of activity

■ We have solved data collection problem for a long time, and very cheaply.

Now What?

■ So now we have 100s of files, closing in on 500GB of data

■ We want to ask some intelligent questions

■ For example: how many people who signed up four weeks ago are still active? (cohort retention)

■ How many products saved does it take for a user to become engaged?

Let’s Dive Deeper

■ Here is an example of our log file(spaces/alignment added for readability)

user_id platform action_type object object_id secondary_object sec_obj_id timestamp-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐8524264|ipad |SaveAction |Product|5757428|Collection |29399687|13683419427555287|android|SaveAction |Product|5758908|GiftsCollection|26680024|13683419423924118|iphone |SaveAction |Product|1979020|Collection |29463107|13683419421285811|ipad |SessionAction|User |1285811| | |13683419428246365|ipod |SaveAction |Product|7930662|Collection |28523544|13788951961233612|desktop|SessionAction|User |1233612| | |13788951969654098|desktop|PostAction |Product|7962904|Store |158163 |13788951979654098|desktop|SaveAction |Product|7962904|GiftsCollection|34407722|1378895197843456 |iphone |SessionAction|User |843456 | | |13788951979005146|android|SaveAction |Product|6389593|GiftsCollection|32117206|13788951976721497|desktop|CommentAction|Product|7930418|Comment |37304732|1378895197

Parsing ASCII files is simple

■ What we get with this file format is simplicity

■ grep, sort, uniq, comp, awk, wc

■ These UNIX tools have been optimized for four decades! I challenge you to write a faster grep!

Have YOU brushed up on your AWK skillz?

Let’s Ask Some Questions

cat user_actions_20130626.log | \awk -F'|' \ '{if ($2==“ipad” && $3==“FollowAction” ){ print $1 } }' | \sort | \uniq | \wc -l

■ How many unique users followed someone or something on iPad on 06/26/2013?

What About Registrations?

cat user_actions_20130626.log | \grep -F -e '|RegisterAction|’ | \wc -l

■ How many total user registrations happened across all platforms on the same day 06/26/2013?

How fast is it really?

■ It takes about 10 seconds to grep through a 1.5GB (single day of recorded events) file

> time gunzip -‐c user_actions.log.20130512.gz | \> /usr/bin/grep SaveAction | wc -‐l......

real 0m 9.584suser 0m 12.195ssys 0m 1.672s

Can we go back a whole year?

■ On one hand, we know how to do it...

■ The problem is: 10 seconds x 360 files

■ Sounds like a data warehouse! /run query; /come back the next day

■ Now we are talking hours of parsing!

Map/Reduce

■ Google published this model in 2004

■ It describes a way to parallelize algorithms across huge data sets

Map/Reduce

■ Decidedly, Map/Reduce requires a new way of thinking

■ Today we have many related projects, such as Hadoop, HDFS, Spark, Hive, Pig

■ Which means that it also requires learning these (somewhat) new tools

On Demand or Permanent?

■ With Hadoop, one practical question is that of infrastructure lifecycle:■ One can create an “on-demand” Hadoop

cluster to run analytics

■ But “on-demand” solution is cheap. Once queried, Hadoop cluster can be killed

■ This requires copying lots of (TBs) of data from storage (typically S3) and takes time

Static Hadoop Cluster

■ With a continuously running Hadoop cluster, the biggest issue is cost

■ It’s very expensive to keep a large cluster around, sitting on top of a copy of a giant dataset

Enter Joyent’s Manta

■ Distributed Object Store, sort of like S3

■ UNIX-like file system semantics for objects, and supports directories (YES!!!!)

■ Native compute on top of objects!

■ Strongly consistent instead of eventual consistency

Detailed look at Manta later at Surge2013

Mark Cavage and David Pacheco (Joyent) will discuss building Manta in “Scaling the Unix Philosophy to Big Data” talk on Friday @ 10am

User Events → Joyent Manta

■ Instead of saving daily event logs to NFS, we now push them as objects to Manta

■ One object = one file = one day of events

■ Let’s look at an example...

Uploading and Downloading

> mput -‐f user_actions.20130911 \ /wanelo/stor/user_actions/20130911

> mget \ /wanelo/stor/user_actions/20130911 > user_actions.20130911

> mmkdir /wanelo/stor/user_actions

Listing Uploaded User Events

> mls /wanelo/stor/user_actions .... 20130909 20130910 20130911 20130912

Beyond Object Store

■ What makes Manta unique is native compute on top of our objects

■ We submit a compute job to Manta

■ Manta creates many virtual instances in seconds (or even milliseconds)

■ We even get root access!

■ We parse our event objects in parallel

Manta’s “Map/Reduce”

■ Streams objects into initial phase

■ Pipes output of initial phase into the input of the next phase (like UNIX!)

■ Each phase is either one-to-one (map phase), or many-to-one (reduce)

Manta’s “Map/Reduce”

input object filtered object

combined resultinput object filtered object

input object filtered object

map phase 1 map phase 2 reduce phase

It’s very familiar, because it’s so similar to piping on a single machine

Real Example

■ Let’s ask a more computationally expensive question:

■ How many times a store was followed in the last three months?

Aggegating Store Follows

■ Map phase:

■ Reduce phase (sum up all the numbers):

grep -F -e '|FollowAction|’ | \grep -F -e '|Store|’ | \wc -l

awk ' { total += $1 } END { print total } '

Cohort Retention Analysis

■ We can save output of map/reduce jobs in another stored object

■ “Cohort” is a set of unique users sharing a particular property

■ Let’s save a unique set of users who registered between 21 and 28 days ago into a temporary object

Cohort Retention Analysis, ctd

awk -F '|' '{ if ($3 == “RegisterAction”) { print $1 } }'

■ Map Phase runs only on 7 days for the given week

■ Reduce phase saves the result into a temporary object

sort | \uniq | \mtee /wanelo/stor/tmp/cohort_user_ids

Cohort Retention Analysis, ctd

■ Now we just need to get unique users active this week, and intersect them with the temporary object

awk -F'|' '{ print $1 }'

sort | \uniq > period_uniq_ids && \ comm -12 period_uniq_ids \ /assets/wanelo/stor/tmp/cohort_user_ids | \wc -l

■ Map Phase runs on last 7 days

■ Reduce phase intersects

Other Uses of Manta @ Wanelo

■ We can migrate user images to Manta instead of S3, and serve them via CDN

■ If we need to create new image format, we submit a job to use CLI tools to generate new format, or thumbnail size

■ We can (and do!) push database backups and PostgreSQL archive logs to Manta

Conclusion■ We were able to create a very cost-efficient

way to store massive amount of events

■ Manta allows us to perform complex algebraic queries on our event data, very fast and also cheap

And we are just scratching the surface of what’s possible with Manta...

Thanks!apidocs.joyent.com/mantagithub.com/wanelogithub.com/wanelo-chef

Wanelo’s technical blog:building.wanelo.com

@kigster

Using Joyent Manta to Scale Event-based Data Collection and Analysis at Wanelo

questions proprietary

user events

users nd products

wanelo result

gb of data

wanelo works

unique users

usergenerated event

Technology

Manta - Panamá

Brendan Gregg Lead Performance Engineer, Joyent -

Joyent Scale in the Cloud GT

Yahoo! + Joyent Quickstart

Manta Trust - Manta Ray Gill Plates ID Guide

€¦ · stoz/or/oz . Portoviejo Portoviejo Portoviejo...

Best Practices for Shopify Stores on...

Manta Termica

"Closing the Gap Between Developers & Customers," Wanelo >>....

Manta bags catalogue / Catálogo Manta carteras

Wanelo Stores Best...

Rimadesio MANTA

Joyent Cloud Advantages

Manta rays

Performance Document - Joyent, Amazon, Rackspace -

Canton Manta