When Ichsan stayed in Tiket, I learnt several things about the cost optimisation. Sometimes, as an engineer, we always focus on making our service optimise, tune in the number of cores we spawn in production and ensure everything runs efficiently. He said, “That’s a good mindset if you want to focus on the cost optimisation, not everyone has the same mindset as you. However, there is another way to ensure we can run efficiently, especially if you have already worked with the enterprise system.” He mentioned utilising the CUD.
Disclaimer, before we run deeper, what I shared here is based on my experience in Tiket, as we are using GCP as our Cloud Platform. If you use AWS, Alibaba Cloud, Tencent, etc, I believe they offer the same thing as CUD. Please do your own research.
GCP offers 2 types of CUD:
- Resource-Based.
- Spend-Based.
Both have their plus and minus points. If you want to know the details of each CUD, you can directly read the docs here. I will not rewrite the details; I’ll share how we utilise and ensure we gain the maximum benefit for each CUD.
I will start with the resource-based. It’s a very specific CUD for committing to a specific machine type. To determine the machine type in GCP, please visit this link. They have a good simulator for you before you start using the machine. In Tiket, previously, we used multiple machine types and even used a custom machine. It’s not a good practice if you want to optimise the cost. You should stick to a few of the machine types (maximum 3 types) to make your life easier. For example, in Tiket, previously, we used N1 (an ancient type of machine), then we moved to T2D (a new generation machine, not the newest, but better than N1). Not only that, but we also use N2D. As my previous post, for the lower environment, we use the E2 machine. So in Tiket, we only maintain 3 types of machines, which are T2D, N2D and E2.
Next step, I checked the usage in the daily report for each type of machine. It’s a good thing, but based on a discussion with TAM (Technical Account Manager at Google), if you focus on cost optimisation, you should go to a granular level, which is to check the hourly usage.
What is the plus point of the resource-based CUD? You focus on a specific type of machine that you can control daily, and the discount will be fully on that machine. For example, if you apply CUD for the E2 machine, then you just need to focus on maintaining the commitment for E2 based on your daily usage. Let’s say your daily usage for E2 is for 10 cores and 40 GB RAM, then you just need to ensure it never reaches below that number because if it’s reached, then you still need to pay 10 cores 40 GB RAM per hour (that’s how the CUD is designed, you need to commit, like marriage LOL).
What is the minus point of the resource-based CUD? There will be a case when you have already committed to one type of machine and you find out, the machine you committed it feels too expensive. If you move to another cheaper type of machine, you can’t do that, as resource-based CUD is already tightly coupled to the machine type itself. In order to prevent that, Flexible Compute CUD will come to the picture.
Now we go with the spend-based. Actually, for spend-based, there are multiple SKUs; however, we will focus on the Flexible CUD for Compute. Most of the type machines are covered by this CUD; only one machine is not covered, which is T2D (at least until this post was published).
What is the plus point of the spend-based CUD? To resolve the minus point of CUD, except for T2D LOL. If you use Flexible Compute CUD, you can use whatever machine you like and get a discount on each type of machine. If you standardise to use multiple types of machines, it can help you to cover that. Based on experience in Tiket, Flex CUD is good for the machine in a lower environment, and when we shut down the machine, it can fill the discount to another type of machine. For example, E2 is covered by Flex CUD. Once we don’t use it in staging because it’s a weekend, and we turn it off in that environment, the discount will fill to the N2D machine in production. Great, right?
What is the minus point of the spend-based CUD? If you use this, it’s expected to use several machine types. Once again, if there are too many machine types, then there are too many machine types you need to maintain. Another thing is, Flex CUD offers less discount than resource-based CUD. The different pretty high, so we need to be meticulous if we want to pick this.
In Tiket, we apply both the CUD, because we use multiple types of machines, and at the same time, we gain the benefit for both CUD. The next question should be, how much do we need to commit? You should go with the percentile 95 of the daily usage for at least 3 months. For the disclaimer, it’s not an exact formula, but it’s based on what we learn during the cost optimisation. For example, if your daily usage is 100 cores and 400 GB RAM, then the number you need to commit is only 95 cores and 380 GB RAM. You don’t need to commit 100% to ensure you have the room for improvement. Another thing is, you need to apply the CUD gradually instead of directly committing with 95%. For example, you apply 50 cores and then in the next 2 weeks, you apply it again, 10 cores and next 2 weeks, apply again for another. This exercise is to prevent you from overcommitting the CUD, as once it’s committed, it can’t be cancelled after the grace period.
When you apply for CUD, there will be a certain period for you to choose, which is either one year or 3 years. Clearly, 3 years will have a higher percentage of discount. However, if you’re not sure, you can go with a one-year commitment first and review your business in the next year.
Back again to the finops, sometimes we’re focusing on our engineering improvement to gain performance, and after that can reduce the infrastructure cost. However, the development process is actually expensive, so if we can reduce the cost without any development, then the approach should be put as a priority.
For the enterprise company, GCP offers an Enterprise Discount Program (EDP) for companies that make for large commitment. It will apply to all SKUs in GCP. In order to do that, you can directly contact the GCP sales representative in your area. I don’t sell that GCP license lol. If you can go with EDP, you can add multiple discounts between EDP and CUD. Typically agreement will take up minimum of 3 years, so for the tips, if you want to apply CUD for 3 years, please ensure it is applied once you renew the contract with GCP.
