Scaling Magento by COPIOUS Inc.

F6e77c80d980f4f8d2bafbfec5b9ded5?s=47 Kyle Terry
October 18, 2013

Scaling Magento by COPIOUS Inc.

Reid Parham, Aaron Edmonds, and Kyle Terry talk about what it takes to scale Magento for a large retailer.

Aaron talks about working on how to work around some core magento-isms. Kyle gives an overview of the hardware cluster powering the whole systems. Reid goes on to explain the people and team aspects of working on such a large and complex system.

F6e77c80d980f4f8d2bafbfec5b9ded5?s=128

Kyle Terry

October 18, 2013
Tweet

Transcript

  1. Scaling Magento Reid Parham, Aaron Edmonds, and Kyle Terry Public

    distribution: sensitive information omitted.
  2. COPIOUS • User-Centered Digital Experience Agency • Strategy • Experience

    • Engineering • http://copio.us/about
  3. Scale Your Code A.K.A. Magento is hard

  4. Code Management • Magento is big! ◦ Our project has

    over 820,000 lines of PHP • Multi-lingual, multi-currency, multi-store • Classes can have complex names ◦ *cough* Enterprise_Reward_Block_Adminhtml_Customer_Edit_T ab_Reward_History_Grid_Column_Renderer_Reason *cough*
  5. Code Management (cont.) • Configuration is driven by XML •

    The dreaded EAV • Magento Indices • Event-Observer
  6. Code Management (Tools) Good tools make the job easier! •

    A good IDE ◦ Magicento • Commerce Bug 2 • n98-magerun
  7. Code Management • NEVER modify core files ◦ Magento’s forum

    never helped • NEVER* add files to app/code/local/Mage ◦ Magento was built to be modular** • Test your code with flat catalog enabled and disabled • Before overwriting classes, check for events
  8. Code Optimization (Quick Wins) Caching Magento Blocks • DIY! Event

    to add cache data: core_block_abstract_to_html_before • OR use a module https://github. com/aligent/CacheObserver
  9. Code Optimization (Quick Wins) Mage::getModel(‘catalog/product’)->load($_product- >getId()); • This is bad

    in templates and when looping over product collections • Load with initial data select ◦ used_in_product_listing attribute option
  10. Code Optimization Make efficient use of Magento indices • Example:

    Catalog URL Rewrites ◦ Includes all products by default (including products marked as “Not Visible Individually”) ◦ Do you need SEO friendly URLs for products that will never be seen??? ◦ Reduce your index size by up to 95% ◦ Mage_Catalog_Model_Resource_Url::_getProducts
  11. Code Optimization (Quick Wins) Mage_Catalog_Model_Resource_Product_Typ e_Configurable_Product_Collection:: isEnabledFlat? FALSE

  12. Systems • Hardware Profile • Cluster Design • Scaling

  13. None
  14. Hardware Profile (overview) • 2 racks of hardware and dozens

    of servers • Top quality of available (and compatible) chipsets and memory • Buffered DDR3; 1 channel per CPU • 126 kW of stable, reliable, redundant, and backed up power • Minor kernel tweaks
  15. Hardware Profile (network) • NetScaler for load balancing ◦ Vserver

    pools ◦ Balances web, database, admin and endeca ◦ Monitors will remove downed hosts • Redundant Network Infrastructure ◦ Backplane uses LACP (link aggregation) for redundancy, load balancing and failover ◦ HA pairing of configurations
  16. Hardware Profile (network) Dynamic port forwarding for browsing: kyle@localhost $

    ssh -L 2221:127.0.0.1:2221 whitelistedhost.example.com kyle@whitelistedhost $ ssh -D 2221 cluster.example.com Static port forwarding for Navicat SSH tunneling (tunneling through a tunnel): kyle@localhost $ ssh -L 2222:127.0.0.1:2222 whitelistedhost.example.com kyle@whitelistedhost $ ssh -L 2222:127.0.0.1:22 cluster.example.com
  17. Hardware Profile (web) • Dual Intel Xeon E3-1230 @ 3.30GHz

    • 32 GB RAM • Dozens of servers • nginx and PHP5-FPM • 6:1 ratio of PHP processes to CPU cores
  18. Hardware Profile (database) • Redundant database hosts • MySQL 5.6

    chosen for scaling capability • tcmalloc further improves throughput • Master/slave replication • Standby hosts for warm failover • Failure point: > 4,000 checkouts/hour
  19. Hardware Profile (database) • Quad Intel Xeon E7-2860 ◦ 10

    cores + hyperthreading each totalling 80 threads • 128 GBs of RAM • RAID10 SSDs for data ◦ writeback cache; noatime,noexec mount options • RAID1 HDDs for OS
  20. Oops!

  21. Hardware Profile (cache) • Powering discrete instances of Redis ◦

    Sessions ◦ Full page cache ◦ Magento back end cache ◦ Background processing queues • Discrete instances are for threading, differing memory limits, differing backup rules, and multi-db deprecation
  22. Hardware Profile (cache) • Content is compressed with LZF ◦

    Compression and decompression with LZF is faster than gzip so it’s an ideal solution • Decreased utilization of network capacity • Sentinel for failover (soon) • RDB BGSAVE: prime number intervals
  23. Compression Outcomes

  24. Hardware Profile (cache) • Quad Intel Xeon E5-2620 @ 2.00GHz

    • 128 GBs of RAM • 4 bonded network interfaces ◦ Prevents saturation of private network ◦ 4 Gb/s ◦ Bonding mode 5 (balance-tlb) ▪ No special switch support ▪ Nice when the colo manages the switch
  25. Hardware Profile (utility) • Cron and systems jobs • Scripts

    • Deploys • Chef Server 10 for deploy and configuration • Tests ◦ Database test suite in Perl (Test::DatabaseRow) • Backups (and copies)
  26. Cluster Overview • Production ◦ Most hardware serves production •

    Staging ◦ Some data promoted to production nightly • Preview{1..n} ◦ Instances for testing and previewing new features, bug fixes and design changes.
  27. • Aggregate hardware availability exceeds six nines (99.9999%) • Software

    availability is ~99.999% • Software, including deployments: 99.98% • Software, including maintenance: 99.9% • Non-recoverable human errors: 98% Production Uptime
  28. Scale Your Team

  29. Team Profile • 16 committers; 8.25 FTE • 4 Project

    Managers • 5 departments • 31 vendors • 5 time zones
  30. Team Values • State your needs; respect others’ • Respect

    is given, then adjusted • Process can always change and improve • Work/life balance • Mature and non-aggressive; mediate conflict • Honesty and transparency
  31. Team Mantras • Trust (relevant) data; make things visible •

    Measurable, repeatable, falsifiable (scientific method) • Redundancy reduces risks (if documented) • Set expectations (timing, contents, formats) and deliver on them
  32. Team Mantras • Automate what is repeated • Use known

    patterns and proven architectures • Grow talent from within • Compartmentalization of some data, code, and knowledge
  33. None
  34. 10 Integrated Vendors Adobe, Akamai, tax calculation, legacy software, Ebay,

    gift cards, ERP (fulfillment and inventory), Oracle, Tierpoint (Dallas, Seattle, Spokane), Endeca provider
  35. advertising, application analytics, email, hardware analysis and functionality, maps, offsite

    storage, promotions, payment gateways, remarketing, shipping estimates, SMS, social networks, uptime 21 Accessory Vendors
  36. • Group emails: avoid general questions, assign actions to people,

    minimize distribution lists • Identify urgency of requests • Use email filters • Coach and mentor Effective Communication
  37. • Daily phone calls: only while needed • Set an

    agenda; keep to a schedule • Encourage people to skip calls or to leave early • End the call when completed Effective Communication
  38. Tools • GitHub • Google Docs • Pivotal Tracker •

    Conference calls, Skype, and IM • BugHerd
  39. None
  40. None
  41. Launch Day

  42. Release Day

  43. QA preparation

  44. Productive!

  45. Off-hours chaos

  46. Build Knowledge • Document the “obvious” • 1000-line README •

    Capture failures and solutions • What happens when? • Which database and server?
  47. Automation Schedule

  48. “This is how we work.”

  49. Example Git Workflow

  50. Learn from previous failures.

  51. Code Review • Standardize pull request structures • Constructive feedback;

    ask questions • emoji-cheat-sheet.com
  52. Code Review Pull requests can also be workspaces

  53. Releases and Git flow: rhythm, ownership, and pride.

  54. None
  55. Deployments • Monday through Thursday only! • Communication: tickets, cross

    references, pull requests, QA status, and releases • Set expectations: timings for outages, maintenance, and degraded functionality • Are we done, yet? • Explain outcomes and options
  56. Community Participation • Patches submitted ◦ Redis ◦ Cm_RedisSession ◦

    Cm_Cache_Backend_Redis ◦ https://github.com/magento/magento2 • Modules improved ◦ CacheObserver ◦ VF_CustomMenu
  57. Community Participation • http://magento.stackexchange.com/ • http://stackoverflow.com/ • phpredis bug(s)

  58. • Spence, Muneera U. Collaborative Processes lecture. 13 Apr. 2006.

    • Marks, Andrea. "The Role of Writing in a Design Curriculum." AIGA: Design Education (2004). • Katzenbach, Jon R., and Douglas K. Smith. The Wisdom of Teams. HarperCollins, 2003. Collaboration Texts
  59. • Bennis, Warren, and Patricia W. Biederman. Organizing Genius. Perseus,

    1997. • Marcum, James W. After the Information Age. Peter Lang, 2006. • https://en.wikipedia.org/wiki/Collaboration (and collaborative method) Collaboration Texts
  60. See Also GitHub (and Gist) @parhamr @kyleterry @aedmonds

  61. Questions?