We have had a considerable drive with innovation that has spawned several megatrends that have affected how we manage and view our network infrastructure and the need for observability. In reality, we have seen the decomposition of everything, from one to many. Many services and dependencies in multiple locations need to be managed and operated instead of the monolithic where everything is generally housed internally. The megatrends have resulted in a dynamic infrastructure with new failure modes not seen in the monolithic forcing us to look a different system observability tools and practices.
There has also been a shift in point of control. We move towards new technologies, and many of these loosely coupled services or infrastructure your services lay upon are not directly under your control. The edge of control has been pushed, creating different types of network and security perimeters. These parameters are now closer to the workload than a central security stack. Therefore the workloads themselves are concerned with security. For a mode detailed explanation of these changes that drive the need for good observability and how they may effect you, a full 2-hour course I did for Pluralsight on DevOps Operational Strategies can be found here: DevOps: Operational Strategies.
How This Affects Failures
The major issue that I have seen with my clients is that application failures are no longer predictable, and the dynamic systems can fail in very creative ways challenging existing monitoring solutions. But, more importantly, the practices that support them. We have a lot of partial failures that are not just unexpected but not known or ever seen before. For example, if you recall, we have the network hero.
The Network Hero
It is someone that knows every part of the network and has seen every failure at least once before. They are no longer useful in today’s world, and you need proper Observability. When I was working as an Engineer, we would have plenty of failures, but more than likely, we would have seen them before. And there was a system in place to fix the error. Today’s environment is much different. We can no longer rely on simply seeing either a UP or Down and setting static thresholds and then alerting based on those thresholds. A key point to note at this stage is that none of these thresholds consider the customer’s perspective. If your POD is running at 80% CPU, does that mean the customer is unhappy? When looking to monitor, you should look from your customer’s perspective and what matters to them. Content Delivery Network (CDN) was one of the first to realize this game and measure what matters most to the customer.
The Different Demands
So the new modern and complex distributed systems place very different demands on your infrastructure and the people that manage the infrastructure. For example, in a microservices, there can be several problems with a particular microservice:
- The microservices could be running under high resource utilization and therefore slow to respond, causing a timeout
- The microservices could have crashed or been stopped and is therefore unavailable
- The microservices could be fine, but there could be slow-running database queries.
- So we have a lot of partial failures.
Therefore: We Can No Longer Predict
The big shift we see with software platforms is that they are evolving much quicker than products and paradigms that we are using, for example, to monitor them. As a result, we need to consider new practices and technologies with dedicated platform teams along with good system observability. We really can’t predict anything anymore, which puts the brakes on some traditional monitoring approaches, especially the metrics-based approach to monitoring. I’m not saying that these monitoring tools are not doing what you want them to do. But, they work in a siloed environment, and there is a lack of connectivity. So we have monitoring tools working in silos in different parts of the organization and more than likely managed by different people trying to monitor a very dispersed application with multiple components and services in various places.
Relying On Known Failures: Metric-Based Approach
A metrics-based monitoring approach relies on having encountered known failure modes in the past. The metric-based approach relies on known failures and predictable failure modes. So we have predictable thresholds that someone is considered to experience abnormal. Monitoring can detect when these systems are either over or under the predictable thresholds that are previously set. And then, we can set alerts and hope that these alerts are actionable. This is only useful for variants of predictable failure modes. Traditional metrics and monitoring tools can tell you any performance spikes or notice that a problem occurs. But they don’t let you dig into the source of problems and let us slice and dice or see correlations between errors. If the system is complex, this approach is harder to get to the root cause in a reasonable timeframe.
With the traditional style metrics systems, you had to define custom metrics, and these were always defined upfront. So with this approach, we can’t start to ask new questions about problems. So it would be best if you defined the questions to ask upfront. Then we set performance thresholds and pronounce them “good” or “bad.” and check and re-check those thresholds. We would tweak the thresholds over time, but that was about it. This monitoring style has been the de facto approach, but we don’t now want to predict how a system can fail. Always looking and always observing instead of waiting for problems, such as reaching a certain threshold before acting.
Metrics: Lack of Connective Event
Metrics did not retain the connective event. As a result, you cannot ask new questions in the existing dataset. These traditional system metrics could miss unexpected failure modes in complex distributed systems. Also, the condition detected via system metrics might be unrelated to what is happening. An example of this could be an abnormal number of running threads on one component that might indicate garbage collection is in progress. It might also indicate that slow response times might be imminent in an upstream service.
Users Experience: Static Thresholds
User experience means different things to different sets of users. We now have a model where different service users may be routed through the system in different ways, using different components, providing experiences that can vary widely. We also know now that the services no longer tend to break in the same few predictable ways over and over. We should have few alerts triggered by only focusing on symptoms that directly impact user experience and not because a threshold was reached.
The Challenge: Can’t Reliably Indicate Any Issues With User Experience
If you are using static thresholds, they can’t reliably indicate any issues with user experience. Alerts should be set up to detect failures that impact user experience. Traditional monitoring falls short in trying to do this. With the traditional metrics-based monitoring, we rely on static thresholds to define optimal system conditions, which has nothing to do with user experience. However, modern systems change shape dynamically under different workloads. Static thresholds for monitoring can’t reflect impacts on user experience. They lack context and are too coarse.
The Need For System Observability
System observability is a practice. Rather than just focusing on a tool that does logging, metrics, or altering, Observability is all about how you approach problems, and for this, you need to look at your culture. So you could say that Observability is a cultural practice and allows you to be proactive to findings instead of relying on a reactive approach that we are used to in the past. Nowadays, we need a different viewpoint, and we generally want to see everything from one place. You want to know how the application works and how it interacts with the other infrastructure components, such as the underlying servers, physical or server, the network, and how data transfer looks in a transfer and stale state. What level of observation do you need so you know that everything is performing as it should? And what should you be looking at to get this level of detail?
Monitoring is knowing the data points and the entities we are gathering from. On the other hand, Observability is like when you put all of the data together. So monitoring is the act of collecting data, and Observability is putting it all together in one single pane of glass. Observability is observing the different patterns and deviations from baseline; monitoring is getting the data and putting it into the systems.
The 3 Pillars of Observability
We have three pillars of System Observability. There are Metrics, Traces, and Logging. So it is an oversimplification to define or view Observability as just having these pillars. But for Observability, you need these in place. Observability is all about how you connect the dots from each of these pillars. If someone asked me which one I prefer, it would be distributed tracing. Distributed tracing allows you to visualize each step in service request executions. As a result, it doesn’t matter if services have complex dependencies. You could say that the complexity of the Dynamic systems is abstracted with distributed tracing.
Use Case: Challenges Without Tracing
For example, latency can stack up if a downstream database service experiences performance bottlenecks. As a result, the end-to-end latency is high. By the time that latency is detected three or four layers upstream, it can be incredibly difficult to identify which component of the system is the root of the problem because now that same latency is being seen in dozens of other services.
Distributed Tracing: A Winning Formula
Modern distributed systems tend to scale into a tangled knot of dependencies. Therefore, distributed tracing shows the relationships between various services and components in a distributed system. Traces help you understand system interdependencies. Unfortunately, those inter-dependencies can obscure problems and make them particularly difficult to debug unless their relationships are clearly understood.