Walmart pagespeed-slide

Real User Monitoring
@Walmarthttp://www.destructoid.com

Presentation Schedule
• Cliff - 10-20m on RUM
• Aaron – 10-20m on Monitoring in real-time
• Balaji – 10-20m on Correlating RUM and Business Analytics
• Final Q&A

Cliff Crocker
- Performance
- Reliability
- Platform & Site Analytics
Twitter: @cliffcrocker

shhhh…..
• We are not the
fastest retail site
on the internet
today.
IE 8 – Web Page Test Comp
Index – Item Page Performance
11s – Fully Loaded time
4

What gets measured, gets done…
• Before we start to optimize – see where our pain is
and prepare to measure for success.
• You never go on a diet without first stepping on the
scale…
5
Source: Flickr

Synthetic Monitoring
Pros
– Technology is great
– Real Browsers (IE, Chrome,
FF)
– Built in Alerting
– Charting/Trending
– Waterfall analysis
– Screen shots & headers
– Object level detail
Cons
– False Positives
– Impossible/impractical to
measure everything
– Fixed number of browser/OS
combinations
– Simulated bandwidth
constraints at best
– Too few data points for
statistical relevance
6

~2.2% of Global Internet
usage - Thanksgiving this year
7

• If only those users
could tell us about their
site experience…
8
Source: Flickr

Enter boomerang.js - https://github.com/yahoo/boomerang
boomerang.js with NavTiming (thanks Buddy & Phil)
• Doc Complete, Page Response, Page Processing, TTFB, DNS, Connect & more
• Cookie data (for parent domain)
• Location (geolookup on IP)
• Referrer
• User Agent
• Anything else you want to stuff into the beacon
+
9
Source: Flickr

Start with a url group (i.e. ‘Item Page’)
Find offenders that
have the most
impact

Identify usage patterns for browsers
(Chrome and Safari users more
active on weekends)
12

What we found scared us…
Home Page Performance – Jan 2012
14

Set some goals and SLAs
• Focus on ‘Page Processing Time’ First
– see Golden Rule – 80% of time spent here (more like 90% for Walmart)
• Look at backend & network response time SLAs later
• Use 95th
Percentile
• Set Achievable SLAs
• Revisit Monthly
• Celebrate Wins!
15

Case Study: Item Page
• Problem:
– Page takes ~24s for slowest 5% of users
• Too many elements
• Slow third party modules
• Several other pagespeed ‘no-nos’
• Goal: Meet SLA for February
– 20s (95th
percentile)
• Approach:
– Scrum team dedicated to perf optimization for 1 sprint
– Team pools resources and ideas - focuses on biggest bang
16

Success!
8s off the 95th
%tile
17

Success!
2.7s under February SLA
18

BEACON
SERVER
BEACON
SERVERCLIENTCLIENT INTERNETINTERNET
MAGIC SMOKE
How boomerang.js Works
1. client downloads JavaScript payload
2. client sends RUM payload to beacon
3. beacon server responds with HTTP 204
WHO CARES!WHO CARES!
20

Aaron Kulick
- Performance
- Founder: SF & SV Web Performance Group
http://www.sfwebperf.org
Twitter: @GoFastWeb
WHO AM I?

Configuring boomerang.js is EASY
Getting Actionable Data Is HARD
22

Where’s the data?
The initial incarnation of RUM @ WMT leveraged Akamai at the edge.
• data reliability issues
• data availability issues
• data “freshness” issues
It works… for certain values of work
Source: Flickr
Source: Flickr
23

HUBBLEHUBBLE
Enter Hubble
JETTYJETTY
DURABLE QUEUEDURABLE QUEUE
APACHE FLUME NGAPACHE FLUME NG
HDFSHDFS UDPUDP
HADOOPHADOOP
24
REAL TIMEREAL TIME

The Bridge Between Then and Now
• Etsy/StatsD - https://github.com/etsy/statsd
– Timers
– Counters
– Simple Aggregation
• Min, Max, Mean, Median, 95th
Percentile
• Whisper, Carbon & Graphite - https://launchpad.net/graphite
– Scalable Realtime Graphing
– RRD-like Storage Requirements + Federated Option
– Supports Irregular Updates

The Now Bits
HUBBLEHUBBLE
UDPUDP
StatsDStatsD
AGGREGATION
Graphite
+
Whisper
Graphite
+
Whisper
UDP/TCPUDP/TCP
STORAGE
+
DISPLAY
• Fixed some calculation problems in Etsy/StatsD
• Added median calculation for timers
• Incorporated multiple flush interval patches

Pretty Pictures
RAW DATA FORMAT
uswmt.all.t_page.upper_95(9289.0),1329259510,1329260710,10|15904.0,9184.0,9125.0,12736.0,11735.0,16776.0,8484.0,
10839.0,14620.0,7579.0,8871.0,8240.0,12390.0,5211.0,10301.0,24784.0,9410.0,16554.0,9609.0,11871.0,12751.0,9797.0,
11003.0,15962.0,7953.0,7707.0,4181.0,11616.0,11746.0,12814.0,10566.0,24782.0,18303.0,20904.0,7718.0,8531.0,7312.0,
9614.0,8749.0,11671.0,5989.0,9832.0,10592.0,11611.0,16946.0,18858.0,14360.0,15927.0,10470.0,10140.0,11307.0,9739.0,
9772.0,9875.0,13641.0,11626.0,14758.0,6529.0,11727.0,10194.0,8003.0,10639.0,7297.0,9891.0,10312.0,12497.0,11557.0,
11406.0,12456.0,12939.0,11029.0,10813.0,11737.0,10618.0,14128.0,16879.0,15865.0,6255.0,14605.0,8861.0,27425.0,
10948.0,19666.0,7185.0,13266.0,13156.0,15111.0,13110.0,15151.0,8666.0,16775.0,10110.0,10387.0,17274.0,22183.0,
8937.0,13168.0,12267.0,11891.0,9635.0,10446.0,8129.0,9550.0,9229.0,8375.0,8657.0,11119.0,6799.0,9094.0,21952.0,
14989.0,16828.0,9001.0,13444.0,10332.0,13609.0,9266.0,13349.0,11546.0,9289.0
uswmt.all.t_page.median(1844.0),1329259510,1329260710,10|4165.0,2333.5,2073.5,2584.0,2547.0,2627.5,2401.0,1575.0,
2170.0,1169.0,1970.0,1838.0,2083.0,5211.0,2496.0,3242.5,1541.0,1437.5,1928.0,1971.0,1776.0,3108.0,2010.5,2044.0,2325.5,
2640.0,1733.0,3924.0,2629.0,1867.5,1782.0,2370.5,2921.0,4783.0,2260.0,1340.0,3256.0,2297.0,2565.0,1874.0,2000.0,2483.5,
2705.5,2432.0,1809.0,2826.0,2204.0,2695.0,1045.0,1615.5,2250.0,2387.0,1562.5,1998.0,2512.0,2139.0,1482.0,2138.5,2100.0,
2583.0,2652.0,3277.0,2549.0,1755.0,2196.5,2766.0,2989.5,3638.0,3034.0,3615.5,2650.5,5207.5,3023.0,1941.0,1918.5,1768.0,
3048.0,1522.5,2710.0,1392.0,2402.0,2005.0,3246.0,1383.0,1880.0,2398.0,1833.0,2579.0,2052.0,2622.0,2089.0,1102.0,1296.0,
3339.0,2132.5,2831.0,3466.0,2131.0,2026.0,2754.0,3228.5,1000.0,2075.0,2011.5,2428.0,4019.5,2788.0,1665.0,1968.0,2695.5,
2873.0,1752.0,2314.5,1766.0,2971.0,3091.5,2205.5,3033.0,2476.0,1844.0
27

The Work Tomorrow…
The Good
•Metric Throughput
•Commodity Storage
•Commodity Infrastructure
The Bad
•More Metric Throughput
•Calculation Complexity
•Web Sockets (pretty)
•Metric Fan-out
28

Is Page Performance a Factor of
Site Conversion? And how big is
it?
February, 2012
v
s

31
Walmart.com - Fun Facts
• Reach
–Millions of Shoppers/week.
–Billions of page requests/year - Spikes up to 1500%
–Billions of internal product search volumes/year
• Scale
–Millions of active product SKUs + Market Place
–Millions of pages indexed in search engines
• Complexity
–1/4th
of page contents served by partners, affiliates and Marketplace
–Multiple departments, 10+ checkout paths
Page Performance & Site Conversion – Feb 2012

32
So, how do you monitor?...

33
Few Industry Benchmarks…
• Factoid 1: Large eCommerce site extensively A/B
tested page performance and published a study
showing 100 millisecond delay = 1% drop in revenue
• Factoid 2: Search Engines A/B tested performance
and found that a 500 millisecond delay caused a 20%
drop in traffic.
• Factoid 3: In an experiment across multiple retailers,
a 1 second delay caused a 7% decline in conversion

34
So, how big is it for Walmart.com?

35
Agenda
• Phase 1 – Baseline Measurement - Impact of Site
Performance on Conversion, Bounce rates &
Revenue
• Phase 2 - Targets for Page Performance
• Phase 3 – Optimization Results
• Key Highlights & Takeaways

36
Agenda
Revenue
• Phase 2 – Targets for Page Performance

37
Impact of site performance on overall site conversion rate….
Baseline – 1 in 2 site visits had response time > 4 seconds
* Sharp decline in conversion rate as average site load time increases from 1 to 4 seconds
* Overall average site load time is lower for the converted population (3.22 Seconds) than the non-
converted population (6.03 Seconds)
Note: Load Time here is the time taken from head of the page to page ready (T_Page)

38
@ Page level….
Page load time is lower for Buyers compared to Non-Buyers
* The Page load time is highest for certain pages - 6.38 secs when there was a conversion and
8.06 where there was no conversion.

39
@ Department level….
Department load time is lower for Buyers compared to Non-Buyers
* Key Categories has 2-3 seconds difference b/w buyer Vs non-buyer

40
What about bounce?
Page Bounce Rate Vs Response Time
* Key pages have high bounce rates which correlates with high T_Page as well
* Significant difference ( up to 9secs) in T_Page between bounced and non-bounced for landing
pages.

41
Bounce rates @ department level….
Department Bounce Rate Vs Response Time
* High T_Page for key pages (up to 19.82s) and key department making Bounce rate significantly
higher

42
Agenda
Revenue

43
Phase 2 – Targets for Page Performance……
Conversion & Bounce Rate Impacts Drives Prioritization

44
Agenda
Revenue

45
Phase 3 – Success Story….
First Win….and yes conversion had positive improvements…

46
• Page speed matters for site conversion!
• Monitor real user performance in a “Big Data” way!!
• Every 1s improvement = Up to 2% increase in CVR
• 100ms improvement = Up to 1% incremental revenue
• SEO benefits for entry pages and reduce bounces
• Test & Learn - Target segments and run A/B Tests focused on
improving page performance
Key Highlights

We’re Hiring for Everything!!!!
47
Source: Flickr
http://www.walmartlabs.com/open-positions/
https://walmartstores.com/careers/apply/?ba=eCom
@cliffcrocker or @GoFastWeb

Walmart pagespeed-slide

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Walmart pagespeed-slide

Similar to Walmart pagespeed-slide (20)

More from Bitsytask

More from Bitsytask (20)

Recently uploaded

Recently uploaded (20)

Walmart pagespeed-slide