Normal Accidents and Root Cause Analysis

From CitconWiki
Jump to navigationJump to search

Normal Accidents book:

Systems are categorized by Interactions that are Simple vs Complex, and Tightly Coupled vs Loosely Coupled.

There are a few different versions of the quadrant:

Douglas Squirrel talking about root-cause analysis:

Notes on Squirrel's talk:

Notes from John Bradshaw:

Normal accidents:

  • 3 Mile Island Accident - Blamed Operators
  • Any system can and will fail, and you should plan for it to fail
  • 2 Axis graph
    • Complexity -> Simple
    • Loose Coupling -> Tight Coupling
    • Complex & Tightly Coupled = Accident
  • Complex system that is Loosely coupled is the CITCON open space set up evening
    • We did not all rush to get food and beer
  • E.g had there been a Lion in there, 1 person could have warned rest
  • Chance to warn of danger
  • Simple but tightly coupled = Dam
    • Accident is water gets through the damn
    • Anything goes wrong with dam e.g. hole, no chance to resolve
    • Simple to reason about, wall of rock with a hole in
    • But is high risk
  • In nuclear plant accident, cooling system near radioactive rods
    • Operators can see there was a leak, but no context e.g. they can see the leak is leaking near/into the radioactive rod storage which would lead to an accident
  • Book to Read: Normal Accidents by Perrow
  • Are micro services tightly coupled and complex?
    • Depends
    • It's down to design and implementation
  • Always strive to be in the bottom right corner of the graph, low complexity loosely coupled
  • How do people plan for failure?
    • Rob - We go through a certification process to get into Retail
  • Each system that could fail is tested, e.g. chaos monkey style someone will manually go take down services
  • Internal team will run same tests internally before handing over to external certification team

How do you verify or even test your logging? Instance of a service that logged every time on failure, in a tight loop and filled the disks leading to further failure = Simple Tightly Coupled System

Root Cause Analysis

Scenario: Database deliberately down for maintenance. Instance of a service that logged every time on failure connecting to database, in a tight loop and filled the disks leading to further failure

  • Basic principals
    • Everybody who was affected comes to the meeting
  • To identity cultural or people problems
  • Not allowed to place blame
  • Ask/poll everyone what was the problem
    • Customer:
      • No system, was down, can't log on
    • Operations:
      • Confused by phone call
    • Customer Service:
      • Angry calls from customers, did not know what was going on
    • Developer:
      • Database down, no disk space
      • Then ask why:
    • Customer:
    • Operations:
    • Customer Service:
    • Developer:
      • Why: Maintenance on database, database down
      • Why: Analysed log files, saw huge files, checked code, logged with no delay
      • Why: Developer skills lacking
      • Why: No code review/inspection
      • Why: Test for this logging case lacking
      • When QA tested database was running
      • QA too busy to investigate database failures cases
      • No new blood in organisation
      • QA assigned/overbooked to too many projects
      • Action: Maintenance on DB, have redundant database to switch to
      • Action: QA involved earlier
  • Actions must be assigned and completed with a timeframe e.g. 1 week
  • When you hit that uncomfortable silence half way down, keep pushing

  • The root cause of failure is always the culture in an organisation
  • It’s always about people e.g.
    • The developer adding no delay to logging
    • Lack of testing
  • Create a RCA timeline of failure
    • At what time did system go down
    • At what time did customers complain
    • At what time did developers react
    • At what time was the system back up
    • Etc
  • Do as much technical investigation as possible before the RCA meeting
    • Eg this was the problem
    • We had these tests
      • But we didn’t have one for this scenario