Prediko

About the team

The founders of Prediko spotted a gap in the market, where numerous omnichannel commerce businesses still plan their inventory on excel spreadsheets. Unlike large retail companies, small retailers and e-commerce businesses do not have sufficient advanced predictive analytical capabilities for inventory.

The ability to predict and plan inventory is crucial to business success, as the correct prediction helps minimize waste, improve operational processes, and increase financial plan accuracy.

To address this problem for e-commerce businesses, the Prediko team resorted to their expertise in developing advanced enterprise operating software. Today, Prediko provides brands with an Inventory Operating System to forecast, plan, order, and finance their stock seamlessly.

Nicolas Sabatier, Co-founder and CTO of Prediko, shares with us how QuestDB works behind the scene for the Prediko Inventory OS.

Why QuestDB?

In the retail industry, Stock Keeping Unit (SKU) is used to track and manage stock levels internally. SKU is the unique combination of the product name and other product details such as color, size, location, etc. SKU management is the process to ensure that every piece of inventory meets the financial objectives of the business. Correct SKU management is key to the success of a business. The basis of Prediko is to make predictions of SKUs for the next twelve months for our users.

We started with machine-learning algorithms providing predictions based on historical data. However, it did not take long for us to realize that we needed a time series database for storage, because SKUs are updated daily and scale up exponentially as historical records grow. Our clients needed to fetch predictions for a product category in a specific period (weekly, monthly, quarterly). We needed a database from which clients could aggregate and fetch data promptly.

In addition, once we provided a baseline prediction, we wanted to allow our users to update the prediction and immediately see the changes in SKU management strategies. Data science is not a magic wand: small businesses might not have a large amount of historical data to rely on. Therefore, after creating the first benchmark, our users should be able to use their business expertise to adjust the predictions. We needed a database that could handle fast updates.

To sum up, we had two requirements for a large and growing amount of data:

• The ability to fetch data fast with SQL
• The ability to update the data easily

We were using Google BigQuery, which could not meet the query speed we required. We considered ClickHouse, TimescaleDB, DuckDB, Druid, and QuestDB. We narrowed down the list by looking for databases that were easy to get started. As a two-person team, a product that took too long to set up would be a waste of our precious time. We finally did a benchmark amongst TimeScaleDB, Druid, and QuestDB. Our queries aggregated various SKUs and fetched the latest version of prediction.

QuestDB stood out immediately from this benchmark: it completed test queries in just over one second, while similar queries took 4 seconds and 3 seconds for TimescaleDB and Apache Druid, respectively. Based on the benchmark results and the fact that we liked to try new exciting technologies, we decided to switch to QuestDB.

Architecture overview

A high-level overview of the Prediko architecture

We have one QuestDB instance on GCP. Both historical data and predictions are ingested and updated to QuestDB via REST API. We use Google's Cloud Run to communicate with the instance and fetch data.

All the SKUs are stored in QuestDB

Our clients have access to user-friendly dashboards to see the aggregated result for different "scenarios", which are recommended SKU plans based on predictions. Users can update the predictions and immediately fetch updated scenarios to change their inventory planning, thanks to QuestDB's excellent query speed.

The Prediko scenario dashboard

Behind the interactive dashboards, QuestDB's engine powers various aggregate functions and fetch data fast on the fly. Some of the aggregation functions look like the following:

SELECT
  coalesce(
    (forecast + manual_change_additive) * manual_change_multiplicative,
    0
  ) * price * (cast(is_active as int)) AS forecast,
  coalesce(
    (forecast + manual_change_additive) * manual_change_multiplicative,
    0
  ) * (cast(is_active as int)) AS units,
  last_year * price as last_year

The aggregated result is then presented using QuestDB's powerful `SAMPLE BY` SQL keyword, allowing an efficient way to summarize large datasets:

SAMPLE BY 1M ALIGN TO CALENDAR

Users can focus on the prediction of a specific product

Favorite SQL queries

The most powerful SQL query we use is undoubtedly the `UPDATE` keyword: it makes our task to update predictions very easy and provides superior query speed. Here is one example:

WITH prediction_update as (
  SELECT
    COALESCE(
      sum(forecast * manual_change_multiplicative) * 0.4519271611197119,
      0.0
    ) as bump,
    sku_id,
    warehouse_id
  FROM
    'read_48f5fda8-3f9a-425c-9584-045d8a3e5dc5_410fa30d-b95e-4463-81df-63e72042146c'
  WHERE
    date >= '2023-01-01'
    AND date < '2023-02-01'
    AND category_id IN ('4850b9e0-2019-46d9-a50b')
  UPDATE
    'read_48f5fda8-3f9a-425c-9584-045d8a3e5dc5_410fa30d' draft
  SET
    stock = CAST(stock + prediction_update.bump as double)
  FROM
    prediction_update
  WHERE
    draft.sku_id = prediction_update.sku_id
    and draft.warehouse_id = prediction_update.warehouse_id
    and prediction_update.bump != 0.0
    and draft.date >= '2023-01-01'
    AND category_id IN ('4850b9e0-2019-46d9-a50b-')

Conclusion

At Prediko, we are obsessed with offering the best sales predictions to e-merchants and to democratize operational excellence. Using QuestDB as our main "Prediction Store" means our users can digest, manipulate, and aggregate predictions at a great speed. Big kudos to the team!