Performance on the Yahoo! Homepage


Published on

Overhauling one of the most visited web sites in the world is a major task, and add on top of it the pressure of keeping performance the same while adding a ton of new features, and you have quite a task. Learn how the Yahoo! homepage team achieved performance parity with the previous version even while adding a ton of new features.

Published in: Technology
1 Comment
  • Nice !! Download 100 % Free Ebooks, PPts, Study Notes, Novels, etc @
    Are you sure you want to  Yes  No
    Your message goes here
No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Performance on the Yahoo! Homepage

  1. 1. Performance on the Yahoo! Homepage Nicholas C. Zakas Principal Front End Engineer, Yahoo! Velocity, June 24 2010
  2. 2. (@slicknet on Twitter)
  3. 3. Principal Front End Engineer
  4. 4. Contributor
  5. 5. Author
  6. 6. The Challenge: Create a new Yahoo! homepage* *add a ton of new functionality **without sacrificing performance
  7. 7. By the Numbers 345 110 million million unique users per month unique users per month worldwide in United States (no pressure)
  8. 8. The Legacy
  9. 9. 1996
  10. 10. 1997
  11. 11. 1999
  12. 12. 2002
  13. 13. 2004
  14. 14. 2006
  15. 15. Today
  16. 16. We strapped ourselves in, believing we could make the fastest Yahoo! homepage yet
  17. 17. Performance is hard The best features for users aren't always the fastest
  18. 18. Content Optimization Engine determines which stories to display at request time
  19. 19. Sites can be completely customized by the user
  20. 20. Popular search topics are determined at request time to display up-to-date info
  21. 21. Random information about other parts of the Yahoo! network
  22. 22. Apps provide more info on-demand
  23. 23. The Cost of Customization • Spriting is difficult – Hard to know which images will be on the page together • Limited image caching – With content constantly changing, getting images into cache doesn't help much • A lot more JavaScript/CSS – And very different, depending on how the user has customized the page
  24. 24. Performance reboot Many of the optimizations made on the previous homepage won't work
  25. 25. Coming to peace with reality We can't optimize everything – so let's just focus on the parts we can
  26. 26. Areas of Focus • Time to interactivity • Ajax Responsiveness • Perceived performance
  27. 27. The time to interactivity is the time between the initial page request and when the user can complete an action
  28. 28. Time to Interactivity • For most pages, happens between DOMContentLoaded and onload – Can actually happen earlier • Links work, forms can be submitted even while the page is loading – As long as JavaScript isn't running • Difficult to measure
  29. 29. Net tab reveals some information Where DOMContentLoaded and onload occur
  30. 30. YSlow reports onload time Useful, but doesn't really determine time to interactivity
  31. 31. Goal: Ensure interactivity by DOMContentLoaded
  32. 32. Simple User Actions • Clicking a headline to read the story • Performing a search • Clicking on a favorite Wait a second! You don't need JavaScript for any of that!
  33. 33. Progressive Enhancement FTW! The more tasks that don't require JavaScript, the faster the user can complete an action
  34. 34. The page is very functional even without JavaScript
  35. 35. Not relying on JavaScript for everything allows us an opportunity to deliver what appears to be a faster experience
  36. 36. JavaScript Loading onto the page without pain
  37. 37. Traditional thinking was put scripts at the bottom
  38. 38. <html> <head> <!-- head contents --> </head> <body> <!-- body contents --> <script type="text/javascript" src="yourfile.js"> </script> <script type="text/javascript" src="yourfile2.js"> </script> </body> </html>
  39. 39. Our results were upsetting Putting scripts at the bottom actually caused other problems
  40. 40. Results • Page would fully render, but be frozen – User can't interact while JavaScript is being fetched/parsed/executed • Delayed onload to after 5s on fast connection • Time to interactivity tied to onload • Experience was especially bad over slower connections – Page was unresponsive for 30s or more
  41. 41. In order to fix things, we had to get lazy
  42. 42.
  43. 43.
  44. 44. function loadScript(url, callback){ var script = document.createElement("script") script.type = "text/javascript"; if (script.readyState){ //IE script.onreadystatechange = function(){ if (script.readyState == "loaded" || script.readyState == "complete"){ script.onreadystatechange = null; callback(); } }; } else { //Others script.onload = function(){ callback(); }; Dynamically loaded scripts } don't block page load script.src = url; document.getElementsByTagName("head")[0].appendChild(script); }
  45. 45. <html> <head> <!-- head contents --> </head> <body> <!-- body contents --> <script type="text/javascript" src="smallfile.js"> </script> <script type="text/javascript"> loadScript(filename, function(){ //initialization }); </script> </body> </html>
  46. 46. Y.Get.script(YUI.presentation.lazyScriptList, { onSuccess: function() { Y.use("*"); Y.ModulePlatform.init(Y.dali.config, true); }});
  47. 47. First script file Everything else
  48. 48. Results • Page is interactive as soon as each section is rendered • Reduced onload time to ~2.5s on fast connections • Slow connection experience vastly improved
  49. 49. JavaScript Loads • Small amount on page load • Larger amount loaded in non-blocking manner – Everything necessary for core JavaScript interactivity • Ajax responses can specify more JavaScript is needed to handle the response – True, on-demand loading
  50. 50. Page Flushing Getting data out to the browser fast
  51. 51. Traditional thinking is to flush after <head>
  52. 52. Flushing after <head> ensures CSS starts to download as quickly as possible But the user still sees a blank screen until the rest of the HTML is rendered to the browser Solution: flush after major sections of the page have been output
  53. 53. <div class="doc"> <div class="hd"> </div> <!-- flushing here does nothing --> <div class=”bd”> </div> </div> The browser won't render a block-level element inside of <body> until the closing tag has been received
  54. 54. <div class="hd"> </div> <!-- flushing here causes the head to render --> <div class=”bd”> </div> Removing page-level wrapper <div> allows the head to render as quickly as possible
  55. 55. Flush Flush
  56. 56. (Actually, we flush a bunch of times as the page is output) Even when the browser can't render yet, it can still start to download other resources such as images
  57. 57. Areas of Focus • Time to interactivity • Ajax Responsiveness • Perceived performance
  58. 58. The biggest area of concern regarding Ajax performance was around the apps. For our very first test, it sometimes took as long as 7 seconds to load a single app.
  59. 59. start stop What exactly is taking 7 seconds? The measurement itself was a huge black box – before doing anything, we had to figure out exactly what was happening in that time
  60. 60. roundtrip start stop Roundtrip Time The first step is the amount of time between when the browser sends the request and the time it receives the response
  61. 61. roundtrip parse start stop Parse Time Next, the JSON response returned from the server has to be parsed
  62. 62. roundtrip parse download start stop JavaScript/CSS Download Time Each response can indicate it needs more JavaScript/CSS before the content can be used
  63. 63. roundtrip parse download render start stop Render Time The amount of time it takes to actually change the display via innerHTML
  64. 64. Where We Started
  65. 65. Fixing Roundtrip Time What's taking so damn long?
  66. 66. The right-side ads were a roundtrip issue The server-side ad call took extra time plus the ad markup represented 50-60% of the returned markup
  67. 67. “Fixing” the ad Entire right column now renders in an iframe. This defers the ad call until after the app has been loaded in the browser, saving both server time for app rendering and bytes in the response.
  68. 68. Fixing Parse Time What's taking so freakin' long??
  69. 69. { "msg": "Hello world!", "day": 6, "found": true, } JSON is super-efficient for transporting numbers, Booleans, simple strings, objects, and arrays
  70. 70. { "html":"<div id="default-p_26583360" class="mod view_default"> <div id="default-p_26583360-bd" class="bd type_pacontainer type_pacontainer_default"><div class=" pa-app-col1 y-pa-ln-open-dk "><div class="hd pa-app-hd"><h2 class="x-large"><a href="_ylt=ArPtckll5ytFOZy32_Tg07qbvZx4/SIG=10u61l0b2/**http %3A//">Finance</a></h2> t<div class="pa- menu-options">n tt<a role="button" class="pa-menu-optionsbtn small pa-app-header" href="#" data- b="_ylt=AhzOmRGiUKjiPuGRaW8LrGabvZx4">Options<span class="y-fp-pg- controls arrow"></span></a>ntttt<ul id="p_26583360-settings-menu" class="y-menu medium">ntttttnttttt<li><a href="_ylt=AtqN.M70D5mHiPrcLvHF9vibvZx4/SIG=1254msah3/**http %3A//" class="y-link-1 help-option"><span class="y-fp-pg- controls"></span>Help</a></li>ntttttnttt </ul>ntt </div></div><div id="default-u_93109" class="mod view_default"> <div id="default-u_93109-bd" class="bd type_finance type_finance_default"> <div class="finance-nav clearfix">n <a href="_ylt=AvKZuIwh_mvmWInFE6c7Zc.bvZx4/SIG=10u61l0b2/**http %3A//" class="small text-color goto-link"><span class="goto">Go to:</span> <span class="property"><img src="" alt="Finance"> Finance</span></a>n <ul class="y-tablist med-small" id="u_93109-tabs"> <li class="selected"><a href="#" data- b="_ylt=AhW8HKKgyZxBNcux07hCVxGbvZx4">Overview</a></li> <li class=""><a href="#" data-b="_ylt=AuEzZyDTq.4N_vTGBXpu2VubvZx4">My Portfolios</a></li> </ul>n </div>n <div class="y-tabpanels y-tp- default">n <div class="tabpanel selected">n <div class="menu menu- empty y-glbl-mod-grad"></div>n <div class="market-overview">n <div class="holder">n <p class="x-small date text-color">Sat, Jun 12, 2010 Very inefficient for large HTML strings 10:10am PDT</p>n <table class="index-update">n class="med-large">n textindent">&nbsp;</td>n" <tbody>n <td class="hide-contents hide- <tr } Escapement adds a lot of extra bytes
  71. 71. The larger the JSON string, the longer it took to parse Keep in mind there was no native browser JSON parsing when we began testing the new page The regular expression validation in the YUI JSON utility (based on json2.js) could take up to 40% of the parse time
  72. 72. Shrinking the size of the HTML by deferring the ad helped But we still wanted to see if we could eek out better gains
  73. 73. [{ "msg": "Hello world!", "day": 6, "found": true, }] ===== <div class="foo"><a href="">Yahoo! </a></div> We experimented with an alternate response format where meta data was in JSON form but HTML was included afterward in plain text
  74. 74. Results were good But then native JSON parsing hit and a lot of problems went away
  75. 75. Fixing Download Time What's taking so (*&#$Q@! long???
  76. 76. On-demand JavaScript/CSS downloading hurt app loading time Intended to decrease page load time, and did – but left us with this side effect
  77. 77. Waiting until the user takes an action ensures paying the cost of download What if you knew which apps the user was going to use? Solution: predictive fetch of JavaScript/CSS before you need it
  78. 78. After page load, we start to download JavaScript/CSS for the apps on the page
  79. 79. After onload onload
  80. 80. Fixing Render Time WTF is taking so (*&#$Q@! long?!?!?
  81. 81. Actually, render time was okay
  82. 82. Results
  83. 83. Areas of Focus • Time to interactivity • Ajax Responsiveness • Perceived performance
  84. 84. Don't underestimate the power of perception
  85. 85. Initially, the new page was actually slower to load than the previous To be expected – a lot more JavaScript and CSS
  86. 86. Blank space is bad Makes it seem like nothing is happening
  87. 87. Adjusting Perception • Frequent page flushing – Progressive rendering avoids a lot of blank space • JavaScript at the bottom – Ensure it doesn't block rendering • Lazy load JavaScript – Decrease time to interactivity
  88. 88. Initially, apps were completely blank while loading (seemed slow)
  89. 89. We changed it to a pseudo- loaded state, which made loading seem faster
  90. 90. In the end, user testing showed that perceived performance of the new page was the same as on the old page
  91. 91. Wrap Up
  92. 92. Lessons Learned • Time to interactivity is a big deal • Progressive enhancement creates a better experience – Allows for delayed load of JavaScript • Load as much JavaScript as possible in a non- blocking manner • Ajax performance is a macro measurement – Get more insight by looking at the parts • Perceived performance is important
  93. 93. Achievements • Reduced time to onload from ~5s to ~2.5s – Actually better than previous version • Very short time to interactivity • Reduced time to open apps from ~7s to ~2s • Maintained perception of speed from previous version
  94. 94. Questions?
  95. 95. Etcetera • My blog: • My email: • Twitter: @slicknet