好玩的网站列表

1. Fail Smarter

• Average latency has a spike, availability down
• reasons
  • dependencies
  • cascading effect
  • client change peak

2. Load balancers

• clients - load balancers - servers
• hardware devices
• multi tenant - use VIP - for efficiency
• used for (features)
  • scaling
  • even traffic distribution
  • overload protection
• how
  • how to pick a server
    • random dice
    • round robin
    • least conns
  • desired algorithm - we use least conns
    • simple
    • reliable
    • even
  • max conns?
  • how to deal with overload?
    • attributes
      • cheap
      • local
      • buffering - having capacity soon
      • priority
    • reject requests - spillover - choose!
      • close the TCP connection
      • no buffer
    • hang on, wait in a queue - surge queue
      • may mkes it take longer
• some default we set - maxConns - perhost setting
  • little’s law
    • arrival rate * time shopping = people in the store
  • see load balancers how many services we have
  • fleet-wide concurrent requests / host count = avg conns

• metrics matter
  • latency netwrok latency + 25%
    • client side
    • server side
  • request rate
• coral server
  • concurrent requests - outstanding request - in one host
• 33% overhead room - dependency failures

3. Actual behave

• increase load to see average latency
• run actual test
• generate graph with outstanding requests
  • see the cross of client timeout and p99

4. Abnormal cases

• dependency latency, timeout
• network packet loss

• let server decide what’s the maxCon should be
• coral has
  • connection
  • worker work thread
• classify requests
  • importance
    • droppable
  • priority
    • order

---

转载请注明来源，欢迎对文章中的引用来源进行考证，欢迎指出任何有错误或不够清晰的表达。可以在下面评论区评论，也可以邮件至 stone2paul@gmail.com