Everyone wants to scale. Almost nobody wants to find their limit first

-

 I remember when Kubernetes landed and half the industry jumped ship overnight. Not because they'd hit a ceiling, but because it was the hype. Orchestration became the answer before anyone asked what the question was.

There are two ways to scale, and the order matters.

Vertical is handling more inside a single instance. Horizontal is adding instances behind a load balancer.

Most teams skip straight to horizontal. But if you don't know the ceiling of one instance, you're not scaling. You're guessing, and you're paying for that guess every month.

Here's what changes when you do the work: you learn that after rigorous optimization, one instance of size X handles 1,000 concurrent requests at $100/month. Now you have a base cost. Peak traffic becomes a number you can predict instead of a bill you explain after the fact. Your infra spend turns into a line item finance can actually plan around.

With that knowledge, if marketing wants to bring 3x more traffic to the app, they know exactly what that costs the company from a tech standpoint, because you know how much your system can handle.

Without that baseline, "just add more instances" becomes the default. It's easier, and in large orgs autoscaling quietly absorbs the decision. The cost doesn't disappear. It just stops being visible.

The AI era has its own version of this. Vibecoders miss it completely, because the goal is to finish, not to understand what they just shipped.

The irony is that AI is good at exactly this work. Connect an MCP to your monitoring tool and ask your model whether your current base scale is actually optimized. Then have it set the base and prepare the autoscaler around it.

Using AI doesn't mean you get to forget the basics. You still have to care. Care, and AI will help you scale better, not just faster.

Optimize vertically. Find the limit. Set your base. Then scale horizontally with intent.

Every cent you don't waste on unmeasured capacity is a cent that goes somewhere it matters.

Tech should be an enabler with a clear cost, not a cost center nobody questions.

Comments

Post a Comment

Popular posts from this blog

Simple FTP Server Solution With Python

-
I spent the last few days dealing with setting up an ftp server on by ubuntu vps and haven't figured out a solution. Then I tried asking my friends thru my facebook account. My friend Sakti  came up with a very simple and elegant solution using python. The solution uses pyftpdlib package. To install the package is really simple: $ easy_install pyftpdlib or $ pip install pyftpdlib after that create the folowing script and save it as a python file: from pyftpdlib.authorizers import DummyAuthorizer from pyftpdlib.handlers import FTPHandler from pyftpdlib.servers import FTPServer authorizer = DummyAuthorizer() authorizer.add_user("user", "password", "/home/user", perm="elradfmw") authorizer.add_anonymous("/home/nobody") handler = FTPHandler handler.authorizer = authorizer server = FTPServer(("0.0.0.0", 21), handler) server.serve_forever() T hen run the script and try accessing your server using ftp. Only 10 line...
Read more

Customizing Sanic's Logging output on Gunicorn

-
For those of you who didn't know, Sanic is a Python 3.6+ web server and web framework that’s written to go fast. It allows the usage of the async/await syntax added in Python 3.5, which makes your code non-blocking and speedy. Sanic's API is highly influenced by Flask and the API is really good. I worked on a personal project lately and found out that doing custom logging on Sanic is quite hard. You don't get a lot of information from Sanic's logging library and since it's based on the default Python 3 logging library , the information it provides is limited. My initial thought was since I will use Gunicorn, let's use Gunicorn's --access-logformat to get the information I needed. The functionality provided a lot of information and I can make use of it to get what I need. However, when you use Sanic Gunicorn Worker, which you need to use to make it work on Gunicorn, it will only use Sanic's logging output. I did a lot of googling and it seems like a lo...
Read more

5 Takeaways From My Past 5 Years With Coral, Prism, Midtrans and Gojek

-
Image
A little over 5 years ago, I decided to resign from Traveloka and joined hands with my friends to help build Coral. It was an exciting project. I have a big interest in the social commerce space and Coral seemed to be a very good take at it. We built it, launched it, learned from it and in the end decided that the business model will not be sustainable. Our experiment was evidently working. Merchant's conversion rate reached over 20% on our platform but we couldn't monetise it. Then we took the decision to pivot. We took the best part of Coral, the chat-to-buy capability, and breathe life into Prism. The product helped businesses dive head first into the social commerce space with their existing platforms. One thing that a lot of people don't know was that Coral and Prism was part of Midtrans from early on. It was acquired before launch. This enabled us to do the experiment without worrying about funding. Gojek acquired Midtrans during the closing days of 2017 and Prism w...
Read more