Part of our Merchandise Engineering Series
In this series, we take a deep dive into Style Arcade's tech to explore what it takes to deliver rapid merchandise planning at scale - including the architectures, technologies, and techniques that we employ, the challenges we have faced along the way, as well as what drew us to this problem in the first place.
And it all starts with a question that many retail technology leaders have:
“Are buyers and planners insane?
We just made a huge technology investment to up-level how we work with data across the organisation - and yet the first thing they do is download to Excel!”
— You, secretly
If there's one thing that I want you to take away from this series, it's that actually, they aren't. Creating a tool that buyers and planners love to use requires more than just intelligence and great UX - it also requires specialised technology.
Because trust me - they don't want to be in Excel.
If you have ever wondered why merch reports run sooo slowly - no matter how successful the brand or how big the IT budget is - then you already know the problem Style Arcade set out to solve.
Today, Style Arcade is powered by our own analytics engine and tech stack that we call Unified Merchandise Analytics (UMA), which was purpose-built to solve the challenges buyers and planners face.
Now you could be forgiven for thinking this is just the output of a random buzzword generator, like "Blockchain AI" or "Agentic Microservices", but it's completely real and shares its DNA with technologies your business is already running on.
We don’t talk about it much (we’ve been too busy building!), but Unified Merchandise Analytics is the source of our speed, and it’s the reason that buyers and planners refuse to let our product go once they get their hands on it.
And we built it because we had no choice. To understand why, we have to go back to 2019: the year our tech stack was born, and the year Style Arcade almost went up in flames.
Ambitions of a planner
The Iconic is an early adopter of Style Arcade and also one of the largest online apparel retailers in the southern hemisphere. From very early on, we were working with a dataset of hundreds of thousands of products, enormous transaction volume, and hundreds of users all logging on at the same time for Monday trade.
Although getting basic metrics like stock, sales, and return rates to work out of a SQL (structured query language) database at this scale was straightforward, my co-founder and CEO, Michaela Wessels, had far greater ambitions. She envisioned:
- A new level of buying accuracy - and therefore profitable trade - powered by sophisticated metrics far too intense for Excel (especially sizing)
- Complete freedom for users to slice and dice these metrics however they wish, on demand
- A platform fast enough that users can answer any question IN the Monday trade meeting
- A platform scalable enough that we could give access to every user in every business
We designed and delivered #1 and #2. However, as our customer base grew, we quickly found that no amount of AWS (Amazon Web Services) spend could deliver #3 and #4. Our forecasted bill for the month was higher than our revenue, and users were still complaining. Not to mention, our app ran like a dog during peak periods despite the soaring costs.
The cause of this problem is best illustrated through an example of how buyers and planners think:
“December sales were soft in our flagship Sydney store, but we didn’t run out of stock of the products with poor sell-through, which also had poor size availability?” - Your head of buying and planning, trying to prevent your cash flow problems from recurring.
I love this question because it gets to the core of why merch is so hard:
- It’s par for the course in buying and planning, and not an unusual question at all
- It completely destroys database performance
Why merchandising questions kill databases
Any merchandise planner reading the above question already knows it’s much more complex than it seems, and answering this in Excel requires a lot of preparation.
I’ll be breaking this exact question down in great detail in the next article in the series, but there are 8 metrics across 5 sequential steps, and the databases powering typical analytics tech have no “one-shot” way to answer them.
If you answer this question without any pre-computation, there are 5 slowdowns in every single step:
- Wait for the step to get to the front of the queue
- Hope there is enough available memory to serve the request
- Re-open the files from the previous step from disk
- Download any additional data sets over the network
- Calculate the results in batches and save the results to file
That’s 5 different ways your question can slow down, repeated across 5 different steps, multiplied by 50 planners during Monday trade. No wonder it’s a grind.
And in order to answer it with pre-computation, you would need to decide in advance what metrics you were going to produce, and at what level of granularity. If you handle every possibility, then your batch jobs will take days to complete, and if you pick only a few, you won’t be able to serve the huge variety of questions buyers and planners ask on a daily basis.
You can’t win:
But wait…wouldn’t Excel be even worse?
Not at all!
In Excel, arbitrarily complex, heavily layered metrics run nearly instantly because there is zero latency, due to a crucial detail that most technology leaders miss: In Excel, all of your logic and data are co-located and run in a single context.
This means you load the file once, and iterate through 5 or 10 or even 50 layers of metrics and filters without any of the five slowdowns above. There’s no saving, loading, queueing, or waiting in between each layer. Truly grasping this would turn out to be the clue that not only saved Style Arcade, but propelled it far beyond where we thought possible.
Back to the story: a business in crisis
This was a crazy time for Michaela and I, because it really did look terminal. Out of desperation, we tried whatever we could to ease the load on our absolutely cooked database.
We started moving logic and data out of the database and into our serverless API layer, and saw results immediately. At first, the application ran more slowly, but we no longer had issues with peak traffic, and performance was stable.
Although it felt like one step forward, two steps backwards, we were convinced this was the right direction and followed the thread. Six months later, the database was no longer doing anything at all, and I had the deeply satisfying pleasure of hitting delete.
We were finally free, and what replaced it was something entirely new.
It wasn’t a better database.
It was the obvious thing we should have built all along, because merch isn’t a database problem; it’s a spreadsheet problem.
Enter: Unified Merchandise Analytics
UMA isn’t the answer to “how do we make a database powerful enough for merch”, it’s the answer to “how do we make a spreadsheet scale to billions of rows”:
- Like Excel, it places all of your retail data and metrics in one unified layer for zero-latency execution
- But unlike Excel, and definitely unlike other databases, you’ll have hundreds of CPU (central processing unit) cores and up to a terabyte of RAM dedicated to dividing and conquering your query, which means fast results on very large datasets, even when there are hundreds of other users
This is visualized below, where the yellow bricks = latency. Instead of a separate step for every layer and a dozen reasons to run slowly…:
.png)
…it’s just one layer, calculating all metrics in a single step:
.png)
Key Takeaways
Unified Merchandise Analytics is the merch spreadsheet - but much faster, larger, and more sophisticated than ever before:
- It’s fast because merch metrics are at the centre of the design, rather than an afterthought, and it’s fast at scale because we give every user dedicated resources through our serverless platform
- This speed also unlocks intelligence - we don’t force our users to choose between doing things accurately and doing them quickly
- It’s also so simple architecturally that our engineering team at Style Arcade spends less time managing infrastructure and more time delivering the features our users love
Unified Merchandise Analytics didn’t just solve the problems we had; it blew away the ceiling on what we could achieve.
