Zalando Documents Machine Learning Platform Architecture

Per Zalando's engineering blog, Zalando operates a multi-stage machine learning platform built to support experimentation and large-scale data processing. The engineering post describes a hosted experimentation environment called Datalab, which is a browser-accessible, JupyterHub-based workspace that includes R Studio, shell access, and preinstalled data-science libraries and connectors to internal data sources such as S3, BigQuery, and MicroStrategy. The same engineering blog reports that Zalando leverages Databricks for Apache Spark workloads when experiments require large-scale processing.

An AWS Machine Learning blog coauthored with Zalando documents how Zalando optimized large-scale inference and ML operations on Amazon SageMaker and explains the company's markdown-pricing solution, described as a forecast-then-optimize pipeline; that AWS post also states Zalando serves approximately 50 million active customers. Vendor writeups add supporting architecture detail: a Databricks blog describes a unified data foundation built using Databricks Unity Catalog and a federated access-control layer, while a ZenML summary outlines an ML lifecycle architecture that includes ZFlow and AWS Step Functions for orchestration.

Per the Zalando engineering blog, the platform separates exploratory work from large-scale processing: Datalab supports notebooks and ad-hoc analysis, while Databricks handles Spark-based scale and long-running jobs. The AWS coauthored post describes production inference on Amazon SageMaker, including optimizations for serving large numbers of SKUs and integrating discount-steering forecasts into downstream pricing optimization. The Databricks material documents a central data governance layer implemented with Databricks Unity Catalog and a federated access model on top of that catalog.

Companies building ML platforms at multi-million-customer scale commonly separate fast experimentation from reliable batch and serving infrastructure; this reduces friction for data scientists while limiting blast radius of experimental code. Industry-pattern observations: teams typically pair notebook-first environments with managed Spark or distributed compute and add a data-governance catalog to centralize schema, lineage, and access controls. Orchestration via workflow engines like AWS Step Functions and pipeline frameworks such as ZFlow mirrors broader MLOps practices for reproducibility and operational reliability.

Zalando's documentation illustrates a pragmatic, multi-vendor ML stack that combines developer-friendly experimentation, managed big-data compute, and cloud-native serving. The combination of a governed data catalog, Spark-on-Databricks processing, and managed inference on Amazon SageMaker reflects an architecture pattern that many large retailers and consumer platforms adopt to balance agility, governance, and cost. For practitioners, the concrete examples-hosted JupyterHub, federated catalog, forecast-then-optimize for pricing-provide reusable reference patterns for ML lifecycle design.

Observers should track additions to the data-governance layer (for example, extensions to Databricks Unity Catalog or federated access controls) and signals about how inference workloads are scaled on Amazon SageMaker. Industry observers will also watch adaptations of the forecast-then-optimize pattern-documented in the AWS post and cited research-for other retail tasks such as inventory or promotion planning.