As a Solutions Engineer at Percona, one of my responsibilities is to support our customers as they investigate new and emerging technologies. This affords me the opportunity to speak to many current and new customers who partner with Percona. The topic of Kubernetes is becoming more popular as companies are investigating and adopting this technology. The issue most companies are encountering is architecting a stateful database that doesn’t fall victim to an environment tuned for ephemeral workloads. This obviously introduces a level of complexity as to how to run a stateful database in an inherently stateless world, as databases are not natively designed for that.
To make your life easier, as a part of the Percona Cloud-Native Autonomous Database Initiative, our engineering teams have built two Kubernetes Operators: Percona Kubernetes Operator for Percona XtraDB Cluster and Percona Kubernetes Operator for Percona Server for MongoDB, which allows for Kubernetes Pods to be destroyed, moved, or created with no impact to the application. To see an overview of Kubernetes, you can read this previous blog of mine Introduction to Percona Kubernetes Operator for Percona XtraDB Cluster that covers this topic. It’s common for companies new to Kubernetes to attempt to run their databases in Kubernetes the same way they would in a traditional environment. But, this is not advised as it introduces the possibility of data loss and it is not recommended for production workloads. Why is this dangerous and how has Percona solved this?
Appropriate Workloads for Kubernetes
Kubernetes is not the answer for everyone. It’s even not the answer for most people. Do not be misled into thinking that moving a database into Kubernetes is going to solve any of your problems. Before you consider moving your database into Kubernetes, ensure the rest of your application is cloud-native and can be used with Kubernetes. Moving your database to Kubernetes should happen after you have started both elastic vertical and horizontal scale and need to orchestrate it to control costs.
As more companies are moving to Kubernetes something has to happen to the legacy workloads. Oftentimes we see a lift and shift mentality into Kubernetes, which can be dangerous or cause more work than expected. We have seen two primary ideal use cases for moving database workloads to Kubernetes: Microservices and Unified Abstraction Layer.
Monolithic, large datasets can prohibit some of Kubernetes’ strong points: self-healing and availability. This can be an issue due to the time it takes to physically transmit data to a new Pod instance as it joins the database cluster. If your dataset is too large, this process is slow due to physical limitations and prohibits performance and the availability of your database. Microservices are a great fit due to the relatively smaller datasets, which allows Kubernetes automation to work well with the dataset size.
Companies looking to take full advantage of cloud-native applications and databases can be a really good fit for Kubernetes as well. If you truly want the ability to deploy and run your databases anywhere utilizing the concept of a Unified Abstraction Layer, Kubernetes is a great option. You can move your databases to anywhere that is running Kubernetes and know it will work.
We talked about large unsharded datasets and the limitations Kubernetes presents when handling them, but we should mention a few more workloads better suited for traditional platforms. Applications with a throughput sensitivity may not do well on Kubernetes, or they may not be cost-effective to do so. Kubernetes is fundamentally designed for container orchestration and is not designed to handle highly performant databases that require low latency. This may be possible to achieve, but at what cost? This applies to highly performant distributed applications as well. Lowest latency across all nodes is not a core tenant of Kubernetes, so ensure you have planned and tested against this before you move everything over to Kubernetes.
Pods Are Cattle, Not Pets
If you’re not familiar with Pets vs Cattle, it’s a DevOps concept that differentiates deployment methodologies of unique servers that require attention when issues arise (pets) versus the ability to replace a server with a copy if issues arise (cattle). Due to the nature of how Kubernetes operates, Pods can be destroyed, spun up, and moved at any time due to factors outside of the application’s control, much like how cattle are treated. Kubernetes uses a scheduler, which by design, can destroy and recreate Pods to meet the configuration needs of your Kubernetes Cluster. This is great for stateless applications as any failure in the application will result in a Pod containing the application being destroyed and recreated, eliminating the need for human interaction, and greatly speeding up the process to a resolution. This isn’t ideal for databases as you don’t want your database to suddenly stop working, halt the application, and introduce the potential for lost or corrupted data. One of the tools Kubernetes can utilize to help combat this is called Stateful Sets. These help by keeping a Pod’s identity assigned to it as it is destroyed and re-created. This helps facilitate stateful workloads, but how does this come into play with high availability and utilizing the automation aspects of Kubernetes?
Databases Are Pets, Not Cattle
Databases by design need to keep their identity, information, and most importantly their data safe and accessible at all times. They are the backbones of the application as they are the source of truth an application relies on for normal processing. Any errors in their operations will quickly stop an application from functioning. They are important, to say the least. How can we safely run databases in Kubernetes and still ensure we have highly available database deployments? By using Stateful Sets and Persistent Volumes we can maintain data integrity, but we need an additional set of hands to take on database administrator tasks such as ensuring failover happens, database members are recovered, and re-join the highly available architecture, along with other technology-specific functions. Fortunately, Kubernetes is extensible and has Operators, which aims to automate the key task of a human operator who is managing a service or set of services.
Automation, Automation, Automation
We know the complexities of running a database (safely) in Kubernetes and some of the concepts used to help bridge the gaps between automation and traditional human functions. With the help of Percona’s Kubernetes Operators, we can safely run databases the way they were intended to run. Percona’s Kubernetes Operators are able to automate tasks that are usually done by a database administrator such as:
- Fully automated deployments with strict consistency and no single point of failure
- Automated scaling with the ability to change the size parameter to add or remove members of a Cluster or Replica-Set
- Fully automated backups and restores
- Fully automated self-healing by automatically recovering from the failure of a single Cluster or Replica-Set member.
- Automatically manage system users on password rotation
- Simplified updates
Always Use a Kubernetes Operator
With the complexities of running a highly available database environment and the inherent dangers introduced by using the dynamic Kubernetes environment, an Operator should always be used when deploying databases in Kubernetes. Fortunately, Percona has already solved this by providing Percona Kubernetes Operator for Percona XtraDB Cluster and Percona Kubernetes Operator for Percona Server for MongoDB. Percona provides full support for databases running in Kubernetes with the Percona Operators. If you are interested in learning more or obtaining support or professional services to maximize your database deployments, please reach out to us.