Velocity NY - Crafting Performance Alerting Tools

The performance team at Etsy has recently developed a number of new alerting tools to help us discover and dig into performance regressions across the site. We built these new tools on top of existing technology (Nagios and Nagios Herald) to bring added context to site slowdowns and help us fix regressions more quickly.

These new alerts change the conversation with our coworkers. Before, we’d ask people about regressions long after they occurred. Now, we discover regressions almost immediately, and are able to share context such as graphs and recent site changes, as we work with other teams to track down and fix regressions.

Transcript

1. Velocity New York – October 13, 2015 @aemcknig Allison McKnight Crafting Performance Alerting Tools
2. Agenda Before monitoring Adding monitoring Iterating on alerting & tools Here’s how it is What’s next? Questions
3. Allison McKnight | #performance Performance at Etsy Allison McKnight | @aemcknig Lara Natalya Kristyn Allison Mike 3
4. Graph everything. 4
5. Allison McKnight | @aemcknig Graphing performance :  320  ms web logs phptime
6. Allison McKnight | @aemcknig 6 Graphing performance github.com/etsy/logster Logster web logs phptime:320 phptime:520 phptime:600 phptime:410 phptime:380 phptime:804 count:          6
 average:      505
 median:        410
 95th  perc:  804
7. Allison McKnight | @aemcknig Graphing performance 7 backend time (ms)
10. Allison McKnight | @aemcknig We needed monitoring 10
11. Allison McKnight | @aemcknig 11
12. Allison McKnight | @aemcknig 12
13. Allison McKnight | @aemcknig Regression report 13 • Didn’t catch small or slow-creep regressions • Difficult to tune • Additional investigation was required to verify and understand regressions • Alert fatigue
14. Allison McKnight | @aemcknig What could we do here? 14 • Enforce better graph-watching during deploys • Change alerting mechanism • Create tools to help investigate regressions • Change alert format
15. Allison McKnight | @aemcknig What could we do here? 15 • Enforce better graph-watching during deploys • Change alerting mechanism • Create tools to help investigate regressions • Change alert format
16. Changing the alerting mechanism
17. Allison McKnight | @aemcknig Monitoring page performance with Nagios 17
18. Allison McKnight | @aemcknig fast and fine-tuned alerting 18 Nagios
19. Allison McKnight | @aemcknig check graphite data script 19 Nagios github.com/etsy/nagios_tools
20. Allison McKnight | @aemcknig Individual check for each service
 you’d like to monitor 20 Nagios
21. Allison McKnight | @aemcknig Individual thresholds for each page
 you’d like to monitor 21 Nagios
22. Allison McKnight | @aemcknig How do you
 choose thresholds for 40 pages? 22
23. Creating tools
24. Allison McKnight | @aemcknig 24
25. Allison McKnight | @aemcknig 25 recommended thresholds
26. Allison McKnight | @aemcknig define  service  {
    use                                          graphite-­‐service
    host_name                              Performance
    service_description          Test  home  Performance
    check_command                      check_graphite_data!700!900!300!'http://...'
    check_interval                    1
    retry_interval                    5
    max_check_attempts            10
    notification_interval      1440
    contact_groups                    performance
 } 26
27. Allison McKnight | @aemcknig define  service  {
    use                                          graphite-­‐service
    host_name                              Performance
    service_description          Test  home  Performance
    check_command                      check_graphite_data!700!900!300!'http://...'
    check_interval                    1
    retry_interval                    5
    max_check_attempts            10
    notification_interval      1440
    contact_groups                    performance
 } 27
28. Allison McKnight | @aemcknig define  service  {
    use                                          graphite-­‐service
    host_name                              Performance
    service_description          Test  home  Performance
    check_command                      check_graphite_data!700!900!300!'http://...'
    check_interval                    1
    retry_interval                    5
    max_check_attempts            10
    notification_interval      1440
    contact_groups                    performance
 } 28
29. Allison McKnight | @aemcknig 29
30. Allison McKnight | @aemcknig define  service  {
    use                                          graphite-­‐service
    host_name                              Performance
    service_description          Test  home  Performance
    check_command                      check_graphite_data!700!900!300!'http://...'
    check_interval                    1
    retry_interval                    5
    max_check_attempts            10
    notification_interval      1440
    contact_groups                    performance
 } 30
31. Allison McKnight | @aemcknig Creating a tool to visualize our performance alerts
 helped us develop well-tuned alerts. 31
32. Allison McKnight | @aemcknig 32
33. Allison McKnight | @aemcknig 33 Current value: 768.5,
 warn threshold: 700.0,
 critical threshold: 1000.0
34. Allison McKnight | @aemcknig We needed alerts that helped us
 understand the problem. 34
35. Changing the alert format
36. Allison McKnight | @aemcknig a tool for adding context to Nagios alerts 36 Nagios Herald github.com/etsy/nagios-herald
37. Allison McKnight | @aemcknig 37 Allison McKnight | @aemcknig
38. Allison McKnight | @aemcknig 38 Allison McKnight | @aemcknig
39. Allison McKnight | @aemcknig 39 Allison McKnight | @aemcknig
40. Allison McKnight | @aemcknig Allison McKnight | @aemcknig 40
41. Allison McKnight | @aemcknig Allison McKnight | @aemcknig 41
42. Allison McKnight | @aemcknig With dependencies, we receive
 only actionable alerts 42
43. Improving sleuthing tools
44. Allison McKnight | @aemcknig top 44 Allison McKnight | @aemcknig 44 actual thresholds
45. Allison McKnight | @aemcknig 45 Allison McKnight | @aemcknig 45 hello
46. Allison McKnight | @aemcknig Allison McKnight | @aemcknig 46
47. Allison McKnight | @aemcknig Allison McKnight | @aemcknig 47
48. Allison McKnight | @aemcknig Allison McKnight | @aemcknig 48
49. Allison McKnight | @aemcknig Allison McKnight | @aemcknig 49
50. Allison McKnight | @aemcknig led to easier and faster investigation 50 Improved context and alerting tools
51. #performance
52. #performance
53. #performance
54. #performance
56. #performance
58. #performance
60. #payments
61. Allison McKnight | @aemcknig improved cross-team collaboration 61 Improved context and alerting tools
62. What’s next?
63. Allison McKnight | @aemcknig 63 Alerting on improvements
64. Allison McKnight | @aemcknig Adding more context
 (teams, alert history) 64
65. Allison McKnight | @aemcknig Better alert integration
 (alert other teams, alert recent deployers) 65
66. Allison McKnight | @aemcknig More comprehensive alerting
 (front-end, mobile, API) 66
67. Use context to improve your tools
68. 68 68 Questions? www.etsy.com/shop/cateanevski Resources Open-source tools:
 github.com/etsy/logster
 github.com/etsy/nagios_tools
 github.com/etsy/nagios-herald
 Icons: www.endlessicons.com @aemcknig Allison McKnight