This article is a note of a speech by Google Academician Jeff Dean at this year's socc conference. It mainly introduces some models used to design distributed systems, the video address of this speech (Silverlight needs to be installed ).
- Splitting a large and complex system into multiple services with as few dependencies as possible helps test, deployment, and independent development by small teams. Example: a Google search depends on more than 100 services. IKE: a mechanism is required to ensure the fault-tolerant of the service, so the failure or failure of a service cannot affect the global effect.
- You need protocol description language, such as protocol buffers. IKE: This reduces the amount of code for communication.
- Ability to predict performance based on system design before development. There are some important numbers at the bottom. IKE: remember the numbers.
- When designing a system, do not want to be comprehensive, but focus on it.
- For incremental design, but not unlimited design, for example, to design 5-50 times of incremental, but more than 1000 times, you need to rewrite and redesign.
- Backup requests are used to reduce latency. For example, if a processing involves 1000 machines, the backup request mechanism is used to prevent the processing from being delayed by a slow machine. IKE: this mechanism is very suitable for mapreduce.
- Use range to distribute data, rather than hash, because it is relatively simple in semantics and easy to control.
- A flexible system can be scaled as needed, and some features are disabled when required, for example, the spelling check is disabled.
- One interface and multiple implementations.
- Add enough observation and tuning hooks ).
Figure 1. Important Numbers