OSCON 2016 - Monitoring in Motion

We rely on our monitoring to tell us when our services, applications, or infrastructure diverge from “normal.” Containers have created a new world of dynamic infrastructure where normal is changing constantly, making it quite difficult to define. How do you check if a service is up when your scheduler or clustering tools are changing the hosts and ports it runs on? Ilan Rabinovitch will review techniques for successful monitoring workflows and explains how to instrument your code in your containers and track the performance and availability of your applications as they move around. The techniques discussed will apply regardless of the monitoring platform you choose.

Transcript

1. Monitoring In Motion Challenges in monitoring kubernetes, containers, and dynamic infrastructure. OSCON Open Container Day May 17, 2016 Ilan Rabinovitch Director, Technical Community
 Datadog
2. $ finger ilan@datadog [datadoghq.com] Name: Ilan Rabinovitch Role: Director, Technical Community 
 Interests: * Monitoring and Metrics * Large scale web operations * FL/OSS Community Events
3. • SaaS based infrastructure and app monitoring • Open Source Agent • Time series data (metrics and events) • Processing nearly a trillion data points per day • Intelligent Alerting • We’re hiring! (www.datadoghq.com/careers/) Datadog Overview
4. Operating Systems, Cloud Providers, Containers, Web Servers, Datastores, Caches, Queues and more... Monitor Everything
6. $ cat ~/.plan 1. Intro: The Importance of Monitoring 2. The Challenge: Monitoring Dynamic Infrastructure 3. Finding the Signal: How do we know what to monitor? 4. Implementation: Applying it to Containerized Workloads
7. Our Focus Area Culture Automation Metrics Sharing Damon Edwards and John Willis DevOps Day LA
8. Culture “organizations which design systems ... are constrained to produce designs which are copies of the communication structures of these organizations” - Melvin E. Conway
14. Follow @honest_update on Twitter
15. Collecting data is cheap;
 not having it when you need it can be expensive
16. Instrument all the things!
17. Sharing Looping Back on Culture Describe the problem as your “enemy” not each other Learn Together
18. Sharing Using and Sharing the same metrics and measurements across teams is key to avoiding misunderstandings.
19. Source: http://bit.ly/1SvvbuP
20. Source: http://bit.ly/1RQRsXW
21. Operational Complexity Increases with.. • Number of things to measure
 • Velocity of change
22. https://www.datadoghq.com/docker-adoption/
23. How much we measure? 1 instance • 10 metrics from cloud providers 1 operating system (e.g., Linux) • 100 metrics 50~ metrics per application
25. How much we measure? 1 instance • 10 metrics from cloud providers 1 operating system (e.g., Linux) • 100 metrics 50~ metrics per application
 N containers • 150*N metrics
26. Operational Complexity 100 instances 400 containers
27. Operational Complexity: Scale 160 metrics per host 640 metrics per host Assuming 4 containers per host
28. Operational Complexity: Scale 100 instances 64,000 metrics Assuming 4 containers per host
29. How much we measure? 1 instance • 10 metrics from cloud providers 1 operating system (e.g., Linux) • 100 metrics 50~ metrics per application
 N containers • 150*N metrics Metrics Overload!
30. Operational Complexity Increases with.. • Number of things to measure
 • Velocity of change
31. Source: Datadog
32. Source: http://bit.ly/1qFylWK
34. Operational Complexity Increases with.. • Number of things to measure
 • Velocity of change
36. Open Questions • Where is my container running? • What is the capacity of my cluster? • What port is my app running on? • What’s the total throughput of my app? • What’s its response time per tag? (app, version, region) • What’s the distribution of 5xx error per container?
37. Source: http://bit.ly/1YxJ7Jy
38. More Details at: http://www.datadoghq.com/blog/monitoring-101-alerting/
39. Monitoring 101
40. Finding Signal - Categorizing Your Metrics
45. Examples: NGINX - Metrics Work Metrics: 
 • Requests Per Second • Request Time • Error Rates (4xx or 5xx) • Success (2xx) Resource Metrics:
 • Disk I/O • Memory • CPU • Queue Length
46. Examples: NGINX - Events • Configuration Change • Code Deployment • Service Started / Stopped
47. Examples: Events
48. When to let a sleeping engineer lie?
49. When to alert?
50. Recurse until you find root cause
51. What to demand from our monitoring tooling?
52. Cryptic Alerts W H A T ?
53. EVERY ALERT MUST BE ACTIONABLE
54. Host Centric
55. Service Centric
56. Static configurations tracking dynamic infrastructure are not a recipe for success. Static vs Dynamic
59. Query Based Monitoring “What’s the average throughput of application:nginx per version ?” “Alert me when one of my pod from replication controller:foo is not behaving like the others?” “Show me rate of HTTP 500 responses from nginx” “… across all data centers” “… running my app version 2….”
60. Getting at the metrics…
61. Resource Metrics Utilization: • CPU (user + system) • memory • i/o • network traffic Saturation • throttling • swap Error • Network Errors 
 (receive vs transmit)
62. Container Events • Starting / Stopping Containers • Scaling Events for Underlying Instances • Deploying a new container build
64. How do we get at the upper layers?
65. Getting at the Metrics CPU METRICS MEMORY METRICS I/O METRICS NETWORK METRICS pseudo-files Yes Yes Some Yes, in 1.6.1+ stats command Basic Basic No Basic API Yes Yes Some Yes
66. Pseudo-files • Provide visibility into container metrics via the file system. • Generally under: 
 /cgroup/<resource>/docker/$CONTAINER_ID/ 
 or
 /sys/fs/cgroup/<resource>/docker/$CONTAINER_ID/

67. Pseudo-files: CPU Metrics $ cat /sys/fs/cgroup/cpuacct/docker/$CONTAINER_ID/cpuacct.stat > user 2451 # time spent running processes since boot > system 966 # time spent executing system calls since boot $ cat /sys/fs/cgroup/cpu/docker/$CONTAINER_ID/cpu.stat > nr_periods 565 # Number of enforcement intervals that have elapsed > nr_throttled 559 # Number of times the group has been throttled > throttled_time 12119585961 # Total time that members of the group were throttled (12.12 seconds) Pseudo-files: CPU Throttling
68. Docker API • Detailed streaming metrics as JSON HTTP socket
 $ curl -v --unix-socket /var/run/docker.sock http://localhost/containers/ 28d7a95f468e/stats

69. STATS Command # Usage: docker stats CONTAINER [CONTAINER...] $ docker stats $CONTAINER_ID CONTAINER CPU % MEM USAGE/LIMIT MEM % NET I/O BLOCK I/O ecb37227ac84 0.12% 71.53 MiB/490 MiB 14.60% 900.2 MB/275.5 MB 266.8 MB/872.7 MB
70. Side Car Containers
71. Aren’t we still missing a layer?
72. Open Questions • What is the capacity of my cluster? • What’s the total throughput of my app? • What’s its response time per tag? (app, version, region) • What’s the distribution of 5xx error per container? • Where is my container running? what port?
73. Service Discovery Docker API Orchestrator Monitoring Agent Container A O A O Containers List & Metadata Additional Metadata (Tags, etc) Config Backend Integration Configurations Host Level Metrics
75. Custom Metrics • Instrument custom applications
 • You know your key transactions best.
 • Use async protocols like Etys’ STATSD or 
 DogstatsD
76. Source: http://bit.ly/1NoW6aj
77. Datadog at OSCON Today 11:30am Sweet deployment flows with Docker, Kubernetes, and OpenShift (Steven Pousty) Thursday 11:05am Detecting outliers and anomalies in real-time at Datadog - Homin Lee (Datadog)
78. Resources Monitoring 101: Alerting 
 https://www.datadoghq.com/blog/monitoring-101-alerting/ Monitoring 101: Collecting the Right Data https://www.datadoghq.com/blog/monitoring-101-collecting-data/ Monitoring 101: Investigating performance issues https://www.datadoghq.com/blog/monitoring-101-investigation/
 The Power of Tagged Metrics https://www.datadoghq.com/blog/the-docker-monitoring-problem/ How to Collect Docker Metrics https://www.datadoghq.com/blog/how-to-collect-docker-metrics/ 8 surprising facts about Docker Adoption https://www.datadoghq.com/docker-adoption/