Return

Pattern:  Observability

Cloud native distributed systems require constant insight into the behavior of all running services in order to understand the system’s behavior and to predict potential problems or incidents

Teams are moving to MS, and there are more and more pieces—the number of components is growing. Traditional responsive monitoring cannot recognize service failures.

In This Context

Traditional systems practice assumes the goal is for every system to be 100% up and running, so monitoring is reactive—i.e., aiming to ensure nothing has happened to any of these components. It alerts when a problem occurs. In traditional monitoring if a server fails, you will have an event; response, even if automatic, is manually triggered. This assumption is not valid for distributed systems.
• A distributed system is by definition not 100% stable—when you have so many pieces, some of them will go up and down at random times.
• Resilience is built into the system to handle the assumption that eventually everything in the system will go down at some point.
• The number of components is always increasing while the number of people working on the application remains reasonably stable.
• Always a cost: if you get something, you must pay for it in some way. Here it is complexity, and you must manage it.
• Manual response is never fast enough in a cloud native system.

Therefore

Put in logging, tracing, alerting, and metrics to always collect information about all the running services in a system.
• Switch from a centrally planned system to distributed self-governing system.
• Continually analyze availability of services and behavioral trends of the individual services and entire system.
• Instead of focusing on specific pieces of hardware, focus on functional behavior of components.
• Consistently collect as many metrics as possible.
• Analyze the trends.
• Create an interface to make system state accessible to all stakeholders: anyone involved in system maintenance or development who needs to understand the system at any given time must be able to observe the system’s behavior.

Consequently
There is a continuous overview of the state of the system, visible to anyone for whom this is relevant information.
+ Analytics and proactive monitoring can be used to discover trends, which can be used to predict failure.
- Any response to a single specific failure is extremely difficult.