Splunk announced the new Splunk Observability Cloud, the full-stack, analytics-powered and enterprise-grade Observability solution.
The following is an excerpt from my book, Mastering Kubernetes:
There is no such thing as a zero-downtime system. All systems fail and all software systems definitely fail. Sometimes the failure is serious enough that the system or some of its services will be down. Think about zero downtime as a best-effort distributed system design.
You design for zero downtime in the sense that you provide a lot of redundancy and mechanisms to address expected failures without bringing the system down. As always, remember that, even if there is a business case for zero downtime, it doesn't mean that every component must have zero downtime. Reliable (within reason) systems can be constructed from highly unreliable components.
The plan for zero downtime is as follows:
Redundancy at every level
This is a required condition. You can't have a single point of failure in your design because when it fails, your system is down.
Automated hot swapping of failed components
Redundancy is only as good as the ability of the redundant components to kick into action as soon as the original component has failed. Some components can share the load (for example, stateless web servers), so there is no need for explicit action. In other cases, such as the Kubernetes scheduler and controller manager, you need a leader election in place to make sure the cluster keeps humming along.
Tons of metrics, monitoring, and alerts to detect problems early
Even with careful design, you may miss something or some implicit assumption might invalidate your design. Often, such subtle issues creep up on you and with enough attention, you may discover it before it becomes an all-out system failure.
For example, suppose there is a mechanism in place to clean up old log files when disk space is over 90% full, but for some reason, it doesn't work. If you set an alert for when disk space is over 95% full, then you'll catch it and be able to prevent the system failure.
Tenacious testing before deployment to production
Comprehensive tests have proven themselves as a reliable way to improve quality. It is hard work to have comprehensive tests for something as complicated as a large Kubernetes cluster running a massive distributed system, but you need it.
What should you test? Everything. That's right. For zero downtime, you need to test both the application and the infrastructure together. Your 100% passing unit tests are a good start, but they don't provide much confidence that when you deploy your application on your production Kubernetes cluster, it will still run as expected.
The best tests are, of course, on your production cluster after a blue-green deployment or identical cluster. In lieu of a full-fledged identical cluster, consider a staging environment with as much fidelity as possible to your production environment. Here is a list of tests you should run. Each of these tests should be comprehensive because if you leave something untested, it might be broken:
• Unit tests
• Acceptance tests
• Performance tests
• Stress tests
• Rollback tests
• Data restore tests
• Penetration tests
Does that sound crazy? Good. Zero-downtime, large-scale systems are hard. There is a reason why Microsoft, Google, Amazon, Facebook, and other big companies have tens of thousands of software engineers (combined) just working on infrastructure, operations, and making sure things are up and running.
Keep the raw data
For many systems, the data is the most critical asset. If you keep the raw data, you can recover from any data corruption and processed data loss that happens later. This will not really help you with zero downtime because it can take a while to re-process the raw data, but it will help with zero data loss, which is often more important. The downside to this approach is that the raw data is often huge compared to the processed data. A good option may be to store the raw data in cheaper storage compared to the processed data.
Perceived uptime as a last resort
OK. Some part of the system is down. You may still be able to maintain some level of service. In many situations, you may have access to a slightly stale version of the data or can let the user access some other part of the system. It is not a great user experience, but technically the system is still available.