Distributed Events

Slides of the presentation I gave at the NCrafts conference in Paris on 12 May (www.ncrafts.io)

Transcript

1. Stijn Volders Freelance .net consultant Loves complex problems (that can be solved with software) @one75 https://www.linkedin.com/in/stijnvolders stijn@one75.be
2. Distributed Events A case study from the trenches
3. Agenda • ReTiBo • The “Green List” concept • Synchronization with events • Scaling with HTTP • Questions
4. Problem space
5. Problem space
6. Problem space
7. Problem space
8. Problem space ReTiBo (Registration, Ticketing, Board computer) • User registration • Travel information (trip duration, available connections,...) • Traffic jams, deviations,...
9. Problem space 4500 driver consoles
10. Problem space 12000 validators
11. Green List
12. Problem space Synchronization
13. Problem space over 3G
14. Problem space
15. Solution
16. Solution • Inspired by ATOM • Based on immutable events • ASP.NET Web API • nginx
17. Solution • No power = no connection • During (school) holidays: 5 to 20% not in use
18. Solution
19. Synchronisation GET api/streams/greenlist/current now past Green separator: event Black separator: page
20. Synchronisation { “links”: [ { “uri”: “api/streams/greenlist/current”, “Rel”: “self” }, { “uri”: “api/streams/greenlist/2015/4/14?page=2”, “Rel”: “previous” } ], “items”: [{...},{...}], }
21. Synchronisation GET api/streams/greenlist/2015/4/14?page=2 now past
22. Synchronisation GET api/streams/greenlist/2015/4/14?page=1 now past
23. Synchronisation GET api/streams/greenlist/2015/4/13?page=9 now past Colored box: already in local datastore
24. Synchronisation past now
25. Synchronisation past now
26. Synchronisation Why backwards?
27. Synchronisation when([SomePatternMatch], function(state, event) { return new state; }); (http://docs.geteventstore.com/introduction/event-sourcing-basics/)
28. Scaling
29. Scaling
30. Scaling Archived events • Cache-Control: max-age=31363200 (somewhere near a year)
31. Scaling Current window • Cache-Control: max-age=3000
32. Scaling
33. Scaling Given • Nothing is cached • A 5 minute interval • 16500 devices
34. Scaling 198 000 connections / hour 198 000 database hits / hour
35. Scaling
36. Scaling Given • A 5 minute cache • A 5 minute interval • 16500 devices
37. Scaling 6 connections 6 database hits
38. Scaling 33 000 times less...
39. Scaling
40. Scaling
41. Bandwith 3G, remember?
42. Bandwidth HTTP to the rescue (again) • ETags • Cache validation • Conditional requests
43. Bandwidth
44. Bandwidth Number time… • 600kb payload for a page of 1000 items • 16h of activity (6am - 10pm) • Sync every 5 minutes
45. Bandwidth Number time… • 7.2 mb every hour (per device) • 110 mb after 16 hours (per device) • 1.8 TB for all devices each day
46. Bandwidth
47. Bandwidth Number time… (conditional requests) • Most greenlist events are created by night (batch) • 1 request in the morning = almost always updated (HTTP 200 OK) • 1 update per hour
48. Bandwidth Number time… (conditional requests) • 600 kb every hour (per device) • ~10 mb after 16 hours (per device) • 165 GB for all devices each day
49. Measure!
50. Wrapping it up • Expose immutable events • Start at the end and follow “previous” • Replay the events chronologically • Conditional requests
51. Questions? @one75 https://www.linkedin.com/in/stijnvolders stijn@one75.be
52. Thank you! @one75 https://www.linkedin.com/in/stijnvolders stijn@one75.be