23
Scalable File System In 14 Days Jeff Hoffer, Software Architect Alex Zherdev, Sr. Software Engineer

Scalable Text File Service with MongoDB (Intuit)

  • Upload
    mongodb

  • View
    278

  • Download
    4

Embed Size (px)

DESCRIPTION

Docstoc.com (founded in 2007, acquired by Intuit in 2013) is one of the largest online repositories of documents. A critical component of our product is our text file service, which delivers text documents to both humans and crawlers. In early 2013 this service, which was file system based, became a prohibitive bottleneck. To meet our scaling needs, we replaced it with one backed by a sharded MongoDB cluster. This talk will cover: Our traffic load (5:1 bots:humans ratio) How we implemented the system in our SOA environment How MongoDB fit our use case out of the box How we load tested peak time traffic before hardware purchase How we loaded the system and how we rolled it out live Performance metrics and gains in stability and reliability

Citation preview

  • Scalable File System In 14 Days Jeff Hoffer, Software Architect Alex Zherdev, Sr. Software Engineer
  • Our Background In the beginning... YouTube for Documents Today Make every small business better Professional Documents Custom Documents Business Licenses Jason Nazar Alon Shwartz The Team
  • Our Product www.docstoc.com
  • Initial Approach Pros: Existing libraries used Reliable storage Replication Cons: Hard to scale out Replication cant keep up Taxed all data SELECT `text_data` FROM `documents` WHERE `doc_id` = 8675309;
  • IIS HTTP Based Solution Pros: HTTP GET IIS Static Content Cache 5TB = Years of Growth Easy Setup & Deploy Cons: Not scalable NTFS & 30M small files Replication In-House HTTP GET http://docs.api/text/160717/8675309.txt
  • Importance of Performance IIS Source Failed early 2013 Page speed heavily influenced our traffic and SEO MongoDB solution implemented within 2 weeks and results immediately felt 0 5 10 15 20 25 Speed 0 1 2 3 4 Views
  • Requirements Sharded horizontal scale out of reads and writes Replication no single point of failure for core business data Doc Page Peak Read Load of 200 / second < 4s REST Interface switch only requires changing URL Easy to Maintain maintenance cost of no more than 1 FTE / day / month 99.9% uptime Can handle # of our current set of text files 43 M Production Rollout within 3 weeks
  • Requirements Sharded horizontal scale out of reads and writes Replication no single point of failure for core business data Doc Page Peak Read Load of 200 / second < 4s REST Interface switch only requires changing URL Easy to Maintain maintenance cost of no more than 1 FTE / day / month 99.9% uptime Can handle # of our current set of text files 43 M Production Rollout within 3 weeks
  • Requirements Sharded horizontal scale out of reads and writes Replication no single point of failure for core business data Doc Page Peak Read Load of 200 / second < 4s REST Interface switch only requires changing URL Easy to Maintain maintenance cost of no more than 1 FTE / day / month 99.9% uptime Can handle # of our current set of text files 43 M Production Rollout within 3 weeks
  • Requirements Sharded horizontal scale out of reads and writes Replication no single point of failure for core business data Doc Page Peak Read Load of 200 / second < 4s REST Interface switch only requires changing URL Easy to Maintain maintenance cost of no more than 1 FTE / day / month 99.9% uptime Can handle # of our current set of text files 43 M Production Rollout within 3 weeks
  • MongoDB FTW
  • Test Setup
  • { id : {document_id} body: {text_content} created: {date_time} } Simple Structure Object Size 50KB Shard on hashed id Rarely modified Heavy Reads Mongo Collection Structure
  • Tests Client Server MongoDB Duration Reads (100/sec) Writes (100/sec) Read+Writes (200/sec)** JMeter Ruby REST Server Empty Collection 20 min (3x) **10x peak load
  • Tests Client Server MongoDB Duration Reads (100/sec) Writes (100/sec) Read+Writes (200/sec)** JMeter Ruby REST Server Empty Collection 20 min (3x) **10x peak load
  • Test Setup
  • Tests Client Server MongoDB Duration Reads (100/sec) Writes (100/sec) Read+Writes (200/sec)** JMeter Ruby REST Server Empty Collection 20 min (3x) JMeter ASP.NET REST Server* Empty Collection 20 min (3x) *ASP.NET MVC 4 Web API **10x peak load
  • Tests Client Server MongoDB Duration Reads (100/sec) Writes (100/sec) Read+Writes (200/sec)** JMeter Ruby REST Server Empty Collection 20 min (3x) JMeter ASP.NET REST Server* Empty Collection 20 min (3x) Jmeter ASP.NET REST Server* Seeded Collection 2M 30 min (3x) *ASP.NET MVC 4 Web API **10x peak load
  • Test Setup
  • Tests Client Server MongoDB Duration Reads (100/sec) Writes (100/sec) Read+Writes (200/sec)** JMeter Ruby REST Server Empty Collection 20 min (3x) JMeter ASP.NET REST Server* Empty Collection 20 min (3x) Jmeter ASP.NET REST Server* Seeded Collection 2M 30 min (3x) .NET Console Loader ASP.NET REST Server* Seeded Collection 2M 1 hour (3x) *ASP.NET MVC 4 Web API **10x peak load
  • Tests Client Server MongoDB Duration Reads (100/sec) Writes (100/sec) Read+Writes (200/sec)** JMeter Ruby REST Server Empty Collection 20 min (3x) JMeter ASP.NET REST Server* Empty Collection 20 min (3x) Jmeter ASP.NET REST Server* Seeded Collection 2M 30 min (3x) .NET Console Loader ASP.NET REST Server* Seeded Collection 2M 1 hour (3x) .NET Console Loader ASP.NET REST Server* Seeded Collection 6M Overnight (10 hrs) *ASP.NET MVC 4 Web API **10x peak load
  • Production
  • In Conclusion Its Good Enough, Its Fast Enough, and Doggone It, Developers Like It! Fast Prototype Low Maintenance Quick Deployment Scale Out Stable Linux, Windows, Mac Excellent Support