incident analysis - procedure and approach
Download as PPTX, PDF3 likes8,923 views
This document provides guidance on conducting post-mortem analyses of outages. It discusses establishing standard procedures and templates to minimize defensiveness and promote learning. Key recommendations include focusing analyses on process improvement rather than blame, having a third party lead investigations, and ensuring action is taken on findings through follow up. Templates are provided to structure data collection, including timelines, root cause analysis, and review of monitoring and logging. The overall goal is for teams to walk away from outages looking like heroes by continually improving based on lessons learned.
incident analysis - procedure and approach
- 1. How to walk away from your Outage looking like a HERO Teresa Dietrich, Vice President Technology Derek Chang, Director Site Reliability Engineering
- 2. Who we are and Why we are here…. Teresa Dietrich – VP of Technical Operations @ WebMD, previously with AOL, @teresadg (Twitter), www.teresadietrich.net Derek Chang – Director of Site Reliability Engineering aka SRE @WebMD, experience in Development, WebOps and CMS www.derekchang.me We are passionate about Outages, Process & Procedures and Always making new mistakes!! 2
- 3. About WebMD • Most Recognized & Trusted Brand of Health Information • Serves consumers, physicians, other healthcare professionals, employers and health plans. • 107 million visitors/month on both desktop and mobile platforms • 2.5 billion page views/month 3
- 4. What is An Outage? Service is unavailable to users or to a subset of users Service is unable to function as designed and implemented Degradation of service to the point the resource is unusable (Defined SLAs) 4
- 5. Why do Outages happen? Bugs in OS, middleware, and application Hardware failure Infrastructure failure (Network, SAN) Environment failures (Power, Cooling) Human Error Demand exceeds capacity Malicious attacks 5
- 6. How are Outages exacerbated? Too long for monitoring to catch the issue Monitoring does not catch the issue, humans eventually do Too long to alert appropriate people of issue Too long for people to respond to alerts Too long to find the cause or source of the issue To long to resolve the issue Lack of communication to Internal and External customers Multiple failure scenario 6
- 7. A different way to do a Post Mortem Focus on improving processes and systems for future, not assigning responsibility for the outage. Structure, structure, structure! Discover, Analyze and Review Analysis done by a third party engineer with DevOps experience @ WebMD. Data collected in a prescribed and orderly fashion, using a template. Recommendations for improvement owned, assigned and tracked through resolution. 7
- 8. Incident Analysis Template 1 You can download the template @ www.teresadietrich.net 8
- 9. Incident Analysis Template 2 You can download the template @ www.teresadietrich.net 9
- 10. Incident 1 – background info 10
- 11. Incident 1 – outage resolution 11
- 12. Incident 1 – timeline analysis 12
- 13. Incident 1 – timeline analysis 13
- 14. Incident 1 – recent application builds, changes and maintenance 14
- 15. Incident 1 – log analysis 15
- 16. Incident 1 – log analysis 16
- 17. Incident 1 – monitoring correlation 17
- 18. Incident 1 – monitoring correlation 18
- 19. Incident 1 – root cause analysis 19
- 20. Incident 1 – root cause analysis 20
- 21. Incident 1 – root cause analysis It's caused by a known Oracle bug 5181800 specifically on oracle version 10.2.0.2. About LNS: LNS (log-write network-server) and ARCH (archiver) processes running on the primary database select archived redo logs and send them to the standby database (IAD1) where the RFS (remote file server) background process within the Oracle instance performs the task of receiving archived redo-logs originating from the primary database (PHX1) 21
- 22. Incident 1 – review and recommendation # Type Review Description Recommendation Process no ON clear was sent after outage update 4 was the last 1. Better process for outage communication RR01 outage is cleared communication 2. firstaid NMS - notification management system Monitoring Currently oracle relies on home-grown detection script to monitor oracle event queue and We should look to third party monitoring tool at hand send email upon errors. The fact that IAD1 inadequate monitoring on (e.g. Zenoss) to monitor oracle components and RR03 RAC problem (which is the origin of oracle infrastructure implement oracle GRID control to provide additional control file lock in PHX1) didn't catch our monitoring attention made the troubleshooting a more difficult and longer process. Monitor alert inadequate monitoring on no alert was sent before/during outage RR04 We should set up alert from Gomez and Truesight. user experience from Gomez and Truesight. Development excessive errors in the request application log make it 1. review current logging implementation 15000 errors on 1/25, 28000 errors on extremely difficult to 2. log clean up RR05 1/26 and 10000 errors on 1/27 on a single troubleshoot by log and in 3. operations should review log and provide report tomcat server turn impact the recovery with engineering regularly (bi-weekly or monthly) time Ops request potential log rotation problem on tomcat server RR06 several logs are only 1 kilobytes in size review/correct log setting and rotation script. (Medscape www backend farm) 22
- 23. Investigation Procedures 23
- 24. Investigation Procedures 24
- 25. Investigation Procedures 25
- 26. Incident 2 – background information 26
- 27. Incident 2 – Timeline analysis and application profiling 27
- 28. Incident 2 – root cause 28
- 29. Incident 2 - resolution 29
- 30. Incident 2 – Resolution rollout • Research: Further research revealed the Jsp compilation meta data are only stored in JVM when the Tomcat Jasper engine runs at development mode • Potential business impact: Teams agreed the solution to turn-off development mode under the assumption that there is no business impact – PJSP update will still function properly • POC: A brief POC test showed non-development mode does reduce memory footprint (memory usage dropped from 196.2Mb to 61.3Mb and total objects in memory dropped from 2.6m to 876k) and all PJSP updates are recompiled and ready to serve in a short moment. • Deployment: Zenoss JMX chart showed the memory dropped back close to initial consumption (0.2-0.3Gb) after each GC cycle while with development mode, the memory inflated to 1G in a couple days and GC could not reclaim memory space and tomcat needed to be restarted. 30
- 31. Incident 2 – Resolution rollout Fix verification: The fix was applied to the whole farm in production. Since then, the result is good - no more restart due to out of memory space and view article performance is more than 30% better in Truesight (avg. 109.5ms compared to 155.9ms before) 31
- 32. Incident 2 – review and recommendation 32
- 33. Change people’s reaction to “Post Mortem” Removing the emotion and blame from the Post Mortem process help minimize the dread and lack of participation. Standard procedures and templates shape people’s expectations and perceptions of the Post Mortem process. With the lead engineer of the investigation having no day to day responsibility with regards to product in question, we can greatly reduced the defensiveness and political stances by those involved. 33
- 34. Ensure the lessons are learned Publishing the results to first to the teams involved and then to the entire technology organization helps with education, openness about the process and accountability for the changes recommended. Take the recommendations, once agreed and approved, and turn them into actionable items: Dev Change Requests, Ops Tickets, Process Update and Communication, Monitoring Change. A single person should own the recommendations becoming action items and responsibility for seeing them through completion. Don’t let them fall by the wayside. During the next outage, try and highlight how the previous lessons improved the next outage, do your own PR for your process. 34
- 35. Questions Time permitting OR Office hours Tuesday June 26 @ 1pm 35
- 36. Appendix - Investigation Procedures 1. Collect background information – Scope of impact – Information about the product(s) impacted – Interview personnel involved 2. Initial interpretation – Type of incident – outage, service degradation – Expectation from senior management – Depth and scope of investigation – Resource planning 36
- 37. Appendix - Investigation Procedures 3. In-depth analysis – Timeline analysis – Change analysis – Log analysis – Monitoring data correlation 4. Research – Vendor documentation and white paper – Architecture review – Code review and application profiling – Infrastructure review 5. Resolution and recommendation 37
Editor's Notes
- #24: Request investigation: initiated by business or senior managementPlanning - initial interpretation (Derek); Scope/depth of investigation and resourceCommunicate – to stakeholders and potential contributors (investigation can be expensive), coordinationDATA – discover, analyze, reviewRecommendation – dev request, ops request, process improvementCommunicate – investigation resultsImplement changes – monitoring improvement
- #25: Quantitative approach/analysis Know existing systems. Products and infrastructureData sourcesInterviews (listen to all parties) advantage and disadvantage – we don’t know the product. Be justTimeline/events: build deployments, changes and maintenanceArchitecture and vendor documentationServer/app Log analysisCode and config review – application profilingData correlationRecurring cycles of discover, analyze and review
- #26: Planning – inioutline investigation (scope/depth),
- #27: Gradual memory degradation 5 weeks4 weeks2 weeks. Restart every 2 weeks is where we stand when SRE is engagedMemory consumption (old gen) quickly built up within 48 hours after restartOverall performance (host latency) improve by 30-40%
- #28: Figure 1 – JVM memory consumption trend – JMX (Java Management Extension) export to Zenoss80% of JVM heap occupied by jspsevlet compilation hash maps
- #29: Jasper stores jsp compilation meta data for developers to review when error occursJasper checks JSP timestamp update in EVERY page request. Easier/faster for developers to verify jsp changes.Because of 1, meta data cannot be GC’ed
- #37: Here I am going to share with you about the approach and the procedures my team take during investigation.Certainly the first step is to collect information to estimate the scope of impact and details about the products that are impacted.source of information are documentation, email communication and then most important of all, through interview people who are involved. (listen to all parties before you set the direction and expectation)Secondly, we will classify the incident, whether it’s an outage or it’s a gradual service degradation (which can possibly turn into an outage) Then we meet with senior management about our findings and decide the scope and the depth of the investigation:we’re not alone in investigation.Investigation can be expensive
- #38: Quantitative approach/analysis – choose the right tools. Splunk, expolog, dynatrace, Data collection, correlation and interpretationLearn, research and review