Seek - Elasticsearch: From Hack to Production

Brett Christensen, Reza Yousefzadeh and Nivantha Mandawala from Seek.com.au talk about migrating from a legacy search system to Elasticsearch for the Seek.com.au 'Talent Search' service.

Transcript

1. Elasticsearch: From Hack to Production
2. About us • Brett Christensen ◦ https://www.linkedin.com/in/brett-christensen-a73a4010 ◦ @brett_c • Reza Yousefzadeh ◦ https://au.linkedin.com/in/reza-yousefzadeh-7052b939 ◦ @reza_yz • Nivantha Mandawala ◦ https://au.linkedin.com/in/nivantha ◦ @nivazone
3. Agenda • Our journey • Continuous Delivery • SEEK Implementation of Elasticsearch • Questions
4. SEEK Search Team Mission: Provide world class search and matching • Highly customised search algorithms • Highly performant indexing/searching • World class consumer experience
5. SEEK Search Team Continuous Delivery as a culture • Cloud first • Faster feedback cycles • Easy to diagnose/resolve issues • Zero downtime deployments • Speed to market++ Credit to Łukasz Górnicki / @derberq
6. The Products
7. Job Search
8. Job Search
9. Talent Search
10. Talent Search
11. The problem: Hitting roadblocks
12. The problem: Hitting roadblocks • Proprietary search engine • Complex queries were too slow • Not horizontally scalable • Not cloud ready • Limited set of API’s, limited SDKs • Very small community
13. Pick a Target Why Talent Search? • Decoupled/Continuously Delivered • All the components already in Cloud • Less high profile compared to jobsearch
14. Pick a Target Why Elasticsearch? • Free and easily available • Doco/community++ • It’s Elastic! • Highly customizable • High team engagement
15. Spiking it out: A hackathon project • 3 days for vertical slice of functionality • No experience with Elasticsearch as core search • Started out with Amazon hosted service
16. Hackathon learnings about Elastic • Good documentation • Large community • Feature rich • Customisable • Highly performant • High confidence in estimating switching costs
17. Next step: Get it on the roadmap Obstacles • Issues with current system known but... • Previous experience with migrating painful
18. Next step: Get it on the roadmap Approach • Start small and demonstrate value quickly • Get people on board early • Build it fast and show it off.... • ...but not just a hack • Test with a small percentage of traffic
19. Outcome • (Most) of Talent Search is now powered by Elasticsearch! • Job Search to follow...
20. Continuous Delivery “In software, when something is painful, the way to reduce the pain is to do it more frequently, not less.” ― David Farley, Continuous Delivery: Reliable Software Releases through Build, Test, and Deployment Automation
21. Continuous Delivery • Architecture • Immutable Infrastructure • Automated Testing • Deployment Pipeline • Technologies
22. Search architecture
23. Old vs. New side by side...
24. Immutable Infrastructure • No updates to running environments • Infrastructure as code • Blue/Green Deployments
25. Immutable Infrastructure • No updates to running environments • Infrastructure as code • Blue/Green Deployments
26. Automated Testing
27. Automated Testing • Smoke test • Stop the pipeline on failed tests • Performance test before switching
28. Deployment Pipeline
29. Immediate feedback
30. Technologies
31. SEEK Implementation of Elasticsearch
32. Elasticsearch - High Level View
33. Going custom...
34. Elasticsearch - Indexing Pipeline
35. Elasticsearch - Indexing Pipeline
36. Elasticsearch - Indexing Pipeline
37. Elasticsearch - Indexing Pipeline
38. Elasticsearch - Indexing Pipeline
39. Elasticsearch - Indexing Pipeline
40. Elasticsearch - Indexing Pipeline
41. Elasticsearch - Customizing Similarity Algorithm { "settings": { "similarity": { "algorithm_1_bm25": { "type": "BM25", "k1": 0.2, "b": 0.35 }, "algorithm_2_bm25": { "type": "BM25", "k1": 0.001, "b": 0 } } } }
42. { "mappings": { "profile": { "properties": { "jobTitle": { "type": "string", "similarity": "algorithm_1_bm25" }, "jobDescription": { "type": "string", "similarity": "algorithm_2_bm25" } } } } } Elasticsearch - Customizing Similarity Algorithm...
43. Elasticsearch - Tuning Relevance • Challenging and time consuming • Quick feedback cycle • Weightings/Boosting • Avoid re-indexing when we can
44. Elasticsearch - Tuning Performance • Memory intensive, Java heap size • Enable replicas after re-indexing • SSD instance storage over EBS • HTTP Compression • Cluster rebuilt/re-indexed in 2 hours!
45. Prod vs. Dark, Champ vs. Challenger • Testing/tuning relevance in dummy data is quite hard • Why not create “dark cluster” with prod data? • Site override, Split traffic 50/50...
46. Summary • Elasticsearch is easy to work with, customizable and feature rich • Continuous delivery made a huge difference • Sell your idea to fellow team members, gather momentum • Start small but build it as if you are going to productionize it
47. Keep in touch... • Brett Christensen ◦ https://www.linkedin.com/in/brett-christensen-a73a4010 ◦ @brett_c • Reza Yousefzadeh ◦ https://au.linkedin.com/in/reza-yousefzadeh-7052b939 ◦ @reza_yz • Nivantha Mandawala ◦ https://au.linkedin.com/in/nivantha ◦ @nivazone