Put the logic back in the DB!

How to choose between DB-centric or horizontally scalable applications? This article explores the components of this choice, provides examples, and a practical checklist to answer this question.

Collecting data, grouping and moving those bits towards our final goal. In a way, this is what we do.
Photo by Siddhant Kumar / Unsplash

I advocate putting business logic into the database AS MUCH AS POSSIBLE.

At first, it may feel weird because, for the last two decades, they taught us the very opposite: CRUD and ORMs are The Way; you don't need to learn SQL; Indexes are the limit of your data knowledge. Focus on Java, focus on .NET. Scale.

"As Much As Possible"

In this article, we focus on the meaning of "as much as possible" and try to find tools that help understand the tradeoff between sheer performances, separation of concerns, and simple orthodoxy.

The Extremes and the Continuum

Let's warm up by considering three business requirements that could be solved in the Data Layer and weigh the pros and cons:

✅ Choose all the customers that placed at least three orders within the last week.

❌ Print the last ten product reviews with a sentiment score.

🧐 Report all the customers that placed at least three orders in the last week, and serve the data in JSON format, including the last three orders for each user.

The first is an easy pick to advocate for the DB as a business logic executor. Using SQL, we can solve this with JOIN and WHERE conditions. Without it, we would need to download the entire dataset on the Application Layer and run the filters, wasting memory and bandwidth.

Given enough data, pulling the entire dataset would become not feasible at all.

The second example is another extreme and advocates for WHAT NOT TO DO in a DB (of every kind). Even a simple sentiment score algorithm may involve thousands of lines of code if not using a Machine Learning model. This is the classic situation where pulling the dataset (ten product reviews) and running an API call toward an existing ML-based Sentiment Service would be efficient.

Let's now move to the third example.

Luke, follow the money!

The third example excites me because there is no simple way to decide how to solve it or WHERE to solve it. All in the DB? Partially in the Application Layer?

We can split the problem in two:

  1. find all the customers that placed at least three orders within the last week
  2. prepare a JSON document with such customer + last three orders

By now, it should be clear that Problem n.1 should be solved in the DB, and for this kind of relational problem, a relational database is possibly the best way to go.

You may have guessed that we will dive into the second part of the problem.

prepare a JSON document reporting
all the active customers
with their last three orders
From now on, I'll assume that the DB solution is based on PostgreSQL because of its ability to manipulate JSON format and its awesomness. Many system use PostgreSQL as data foundation anyway.

Problem n.2 is different and can be solved in many ways. For the sake of simplicity, we will constrain the discussion to these three possible implementations:

  1. use PostgreSQL and build aggregate JSON using SQL
  2. pull data from DB and build the JSON on the Application Layer
  3. pull data from DB and build the JSON in the Client App

We will dive into the technicalities of those solutions in another article. Today, let's explore how the three approaches impact the "who pays what?" question.

In Solution n.1, we (the company) pay for the entire computational cost in the DB, possibly reduce the bandwidth cost, and serve a ready-to-use JSON report.

In Solution n.2, we still pay for the entire computational cost, but we divide it between the Storage Layer (WHERE and JOIN) and the Application Layer (build the JSON). On the bright side, we reduce the workload on the DB for the sake of Horizontal Scalability. Still, we must pay the overhead of multiple machines, CPUs, and memory that runs the Application Layer.

Yes, Linux still needs CPU and memory to run properly.
And you pay for it.

Solution n. 3 taps into the Infinite Scalability of the Client App. Each new client brings its CPU and memory to the table. We may pay a little more bandwidth, but then we save on CPU and memory. Unfortunately, the Client App can not always be trusted (easy to hack), so we must be careful when choosing this approach and when we simply can not (due to requirements).

🎉 I choose Solution n.3 every time my requirements allow it!
I'm a big fan of the Infinite Scalability Opportunity. Also, I enjoy working with techs like modern Javascript and React.

Infinite Scalability is FREE,
as in Free Beer!

Where do I spend my dollars?

But using the Client App is not always possible, and more often than I like to admit, I find myself choosing between more costly options. (expensive for my company).

In the end, the difference between Solution n.1 and Solution n.2 lies in where we spend our money:

  • Do we pump up the DB?
  • Do we scale up horizontally?

Pumping the DB may lead us into the Vertical Limit; The other option carries the overhead of managing multiple machines and OSs resulting in an overwhelming maintenance nightmare.

Although Public Cloud Providers mitigate both problems, you still have to pay the overhead of horizontal scaling. One way or the other. Nothing comes for free. (never heard of entropy?)

Even with the uprising of great tools such as Kubernetes, we find ourselves dedicating more and more resources to "how to run the service" instead of "bringing functional value to our customers."

Jeff Bezos would be pissed.

Size Matters

Coming down to this economical choice, it is just a matter of sizing:

How big do you expect your business to be?
  • Are we talking about Facebook size?
    Go for Horizontal Scaling.
  • Are we talking about a small single-tenant management tool?
    Focus on the DB.

Don't make the mistake of pursuing Horizontal Scalability if your target business will never need it, and don't overtrust a single tool and its (promised) performances. (Yes, even if that tool is PostgreSQL, and I love it with all my heart).

Scalability is a business decision.

Knowledge Also Matters

Once you have figured out your domain constraints (e.g., no Client App) and business requirements (e.g., I can do that in the DB), you should still consider your team's experience and knowledge.

Creating a good DB schema, performant queries, and maintainable functions does not come for free. At the same time, taming all the challenges that come with Horizontal Scaling requires experience and dedication.

DevOp is a profession of its own!
(Cloud providers facilitates the maintenance and governance of a complex infrastructure, but the challenges there are far from being fully managed!)

You may decide to invest in your team skills development – and IMHO, you will gain some superpowers once they embrace the SQL way – or you can use what you have and pay a consultant to set up your Kube cluster.

Knowledge Development
is a business decision.

This is also a business decision, although I am biased toward growing the team's culture and knowledge 😉.

Scalability is a business decision

👉 The point here is that scalability is a business decision and brings some risks to the evaluation table!

It takes multiple actors to get out of the woods:

  • Marketing should provide estimates of the projected customer base growth.
  • Product should translate that input into data entity growth and design the worst-case (still realistic) scenario for data manipulation needs.
  • Engineers should run eager load tests simulating data at scale and provide metrics for the expected lifespan of a DB-centric approach. (aka, the POC)
  • DevSecOps should produce an operational cost chart for a monolithic approach and a service-based solution.
  • Architects should evaluate the costs of increasing the team's knowledge. Either to write more efficient software or to run a scalable system.
    And what would it be like to switch on the run.

the CEO must drive the decision.

She knows what is at stake and can use her professionals' inputs to craft a data-driven way forward.