Mowing the DbaaS Weeds Without --Force

With the release of a recent paper that I had the pleasure to co-author, Building a Hybrid Database-as-a-Service Platform with Azure Stack HCI, I wanted to continue to author some additional interesting perspectives, and dive deep into the DbaaS technical weeds.

That recent paper was a refresh of a previous paper that we wrote roughly 16 months ago. That time frame is an eternity for technology changes. It was time to refresh with tech updates and some lessons learned.

The detail in the paper describes an end-to-end Database as a Service solution with Dell and Microsoft product offerings. The entire SysOps, DevOps, and DataOps teams will appreciate the detail in the paper. DbaaS actually realized.

One topic that was very interesting to me was our analysis and resource tuning of Kubernetes workloads. With K8s (what the cool kids say), we have the option to configure our pods with a very tightly defined resource allocation, both requested and limits, for both CPU and memory.

A little test harness history

First, a little history from previous papers with my V1 test harness. I first started working through an automated test harness constructed by using the HammerDB CLI, T-SQL, PowerShell, and even some batch files. Yeah, batch files… I am that old. The HammerDB side of the harness required a sizable virtual machine with regard to CPU and memory—along with the overhead and maintenance of a full-blown Windows OS, which itself requires a decent amount of resource to properly function. Let’s just say this was not the optimal way to go about end-to-end testing, especially with micro-services as an integral part of the harness.

Our previous test harness architecture is represented by this diagram:

The better answer would be to use micro-services for everything. We were ready and up for the task. This is where I burned some quality cycles with another awesome Dell teammate to move the test harness into a V2 configuration. We decided to use HammerDB within a containerized deployment.

Each HammerDB in a container would map to a separate SQL MI (referenced in the image below). I quickly saw some very timely configuration opportunities where I could dive into resource consolidation. Both the application and the database layers were deployed into their own Kubernetes namespaces. This allowed us a much better way to provide fine-grained resource reporting analysis.

There is a section within our paper regarding the testing we worked through, comparing Kubernetes requests and limits for CPU and memory. For an Azure Arc-enabled SQL managed instance, defining these attributes is required and there are minimums as defined by Microsoft here. But where do you start? How do we size a pod? There are a few pieces of the puzzle to consider:

• What is the resource allocation of each Kubernetes worker node (virtual or physical)— CPU and memory totals?
• How many of these worker nodes exist? Can we scale? Should we use anti-affinity rules? (Not that it is better to let the scheduler sort it out.) 
• Kubernetes does have its own overhead. A conservative resource allocation would allow for at least 20 percent overhead for each worker node.

For DbaaS CPU, how do we define our requests and limits?

We know a SQL managed instance is billed on CPU resource limits and not CPU resource requests. This consumption billing leaves us with an interesting paradigm. With any modern architecture, we want to maximize our investment with a very dense, but still performant, SQL Server workload environment. We need to strike a balance.

With microservices, we can finally achieve real consolidation for workloads.

Real… Consolidation...

What do we know about the Kubernetes scheduler around CPU?

• A request for a pod is a guaranteed baseline for the pod. When a pod is initially scheduled and deployed, the request attribute is used.
• Setting a CPU request is a solid best practice when using Kubernetes. This setting does help the scheduler allocate pods efficiently.
• The limit is the hard “limit” that that pod can consume. The CPU limit only affects how the spare CPU is distributed. This is good for a dense and highly consolidated SQL MI deployment.
• With Kubernetes, CPU represents compute processing time that is measured in cores. The minimum is 1m. My HammerDB pod YAML references 500m, or half a core.
• With a CPU limit, you are defining a period and a quota.
• CPU is a compressible resource, and it can be stretched. It can also be throttled if processes request too much.

Let go of the over-provisioning demons

It is time to let go of our physical and virtual machine sizing constructs, where most SQL Server deployments are vastly over-provisioned. I have analyzed and recommended better paths forward for over-provisioned machines for years.

• For SQL Server, do we always consume the limit, or max CPU, 100 percent of the time? I doubt it. Our workloads almost always go up and down—consuming CPU cycles, then pulling back and waiting for more work. 
• For workload placement, the scheduler—by the transitive property—therein defines our efficiency and consolidation automatically. However, as mentioned, we do need to reference a CPU limit because it is required.
• There is a great deal of Kubernetes CPU sizing guidance to not use limits; however, for a database workload, this is a good thing, not to mention a requirement and good fundamental database best practice.
• Monitor your workloads with real production-like work to derive the average CPU utilization. If CPU consumption percentages remain low, throttle back the CPU requests and limits.
• Make sure that your requests are accurate for SQL Server. We should not over-provision resource “just because” we may need them.
• Start with half the CPU you had allocated for the same SQL Server running in a virtual machine, then monitor. If still over-provisioned, decrease by half again.

Kubernetes also exists in part to terminate pods that are no longer needed or no longer consuming resources. In fact, I had to fake out the HammerDB container with a “keep-alive” within my YAML file to make sure that the pod remained active long enough to be called upon to run a workload. Notice the command:sleep attribute in this YAML file:

apiVersion: v1
kind: Pod
metadata:
  name: <hammerpod>
  namespace: <hammernamespace>
spec:
  containers:
  - name: <hammerpod>
    image: dbaasregistry/hammerdb:latest
    command:
      - "sleep"
      - "7200"
    resources:
      requests:
        memory: "500M"
        cpu: "500m"
      limits:
        memory: "500M"
        cpu: "500m"   
  imagePullPolicy: IfNotPresent

Proving out the new architecture

Our new fully deployed architecture is depicted below, with a separation of applications, in this case HammerDB from SQL Server, deployed into separate namespaces. This allows for tighter resource utilization, reporting, and tuning.

It's also important to note that setting appropriate resources and limits is just one aspect of optimizing your Azure Arc-enabled data services deployment. You should also consider other factors, such as storage configuration, network configuration, and workload characteristics, to ensure that your microservice architecture runs smoothly and efficiently.

Scheduled CPU lessons learned

The tests we conducted and described in the paper gave me some enlightenment regarding proper database microservice sizing. Considering our dense SQL MI workload, we again wanted to maximize the amount of SQL instances we could deploy and keep performance at an acceptable level. I also was very mindful of our consumption-based billing based on CPU limit. For all my tests, I did keep memory as a constant, as it is a finite resource for Kubernetes.

What I found is that performance was identical, and it was even better in some cases when:

• I set CPU requested to half the limit, letting the Kubernetes scheduler do what it is best at—managing resources.
• I monitored the tests and watched resource consumption, tightening up allocation where I could.

My conclusion is this: It is time to let go and not burn my own thinking analysis on trying to outsmart the scheduler. I have better squirrels to go chase and rabbit holes to dive into. 😊

Embrace the IT polyglot mindset

To properly engage and place the best practice stake in the ground, I needed to continue to embrace my polyglot persona. Use all the tools while containerizing all the things! I wrote about this previously here.

I was presenting on the topic of Azure Arc-enabled data services at a recent conference. I have a conversation slide that has a substantial list of tools that I use in my test engineering life. The question was asked, “Do you think all GUI will go away and scripting will again become the norm?” I explained that I think that all tools have their place, depending on the problem or deployment you are working through. For me, scripting is vital for repeatable testing success. You can’t check-in a point-and-click deployment.

There are many GUI tools for Linux and Kubernetes and others. They all have their place, especially when managing very large environments. I do also believe that honing your script skills first is best. Then you understand and appreciate the GUI. 

Being an IT polyglot means that you have a broad understanding of various technologies and how they can be used to solve different problems. It also means that you can communicate effectively with developers and other stakeholders, from tin to “C-level” who may have expertise in different areas.

For most everything I do with Azure Arc, I first turn to command line tools, CLI or kubectl to name a few. I love the fact that I can script, check in my work, or feed into a GitOps pipe, and forget about it. It always works on my machine. 😉

To continue developing your skills as an IT polyglot, it's important to stay up to date with the latest industry trends and technologies. This can be done by attending conferences, reading industry blogs and publications, participating in online communities, and experimenting with new tools and platforms. As I have stated in other blogs… #NeverStopLearning

Author: Robert F. Sonders
Technical Staff – Engineering Technologist
Multicloud Storage Software