TL;DR
Useful data lake examples show three things, which are what data went in, how it was stored and organized, and what it fed. Walmart streams business events and database changes into lakes that it began moving to an Apache Hudi lakehouse in 2023. Robinhood copies Postgres changes into Hudi tables on Amazon S3, which cut core-table freshness from 24 hours to under 15 minutes. In healthcare and manufacturing, a Microsoft reference design and two Kanerika client lakes land clinical, pharma and ERP data in governed layers. Uber, Airbnb and Netflix show the longer arc, as raw files gave way to open table formats such as Hudi and Iceberg. The shared pattern is to land raw data once, organize it in layers, add a table format and catalog, then serve BI and ML.
Key Takeaways
Walmart (2023) feeds Kafka event streams and database change data from supply chain, customer, inventory and search into lakes now moving to a Hudi-based lakehouse. Robinhood (2022) streams Postgres changes into Hudi tables on Amazon S3, so core tables refresh in under 15 minutes instead of 24 hours. Microsoft’s Fabric healthcare design lands FHIR records, DICOM imaging and CMS claims in OneLake, while Kanerika’s pharma data lake serves 15+ use cases. Kanerika unified SAP, Azure Synapse, SQL Server and Marketo data in OneLake for a packaging leader, with a 60% increase in data accessibility. Uber, Airbnb and Netflix show the long arc, from raw-once ingestion and small-file pain to open table formats like Hudi and Iceberg. The same blueprint maps to Amazon S3 with Glue and Lake Formation, to Azure Data Lake Storage, or to OneLake in Microsoft Fabric. Watch on YouTube
Microsoft Fabric for Manufacturing: SSMH Case Study & Business Impact
Kanerika walks through a real manufacturing deployment on Microsoft Fabric. It shows how plant and business data were brought into one governed store and what that store went on to feed.
Every Data Swamp Started as Somebody’s Data Lake
Nobody sets out to build a data swamp. A team points a few pipelines at cheap object storage, the first dashboards work, and other teams ask for access. Two years later the bucket holds thousands of folders, three copies of the customer table and a directory called final_v3 that nobody dares to delete.
That is the warning behind every example in this guide. The lakes that lasted were not always the biggest. They were the ones where someone decided what went in, how it was stored and what it had to feed, and then wrote it down.
Five of the eight data lake examples below come from engineering teams that published their own designs. The other three are a Microsoft reference architecture and two Kanerika client projects. Each one is dated, because a design from 2018 and a design from 2023 solve different problems. If your own lake already shows swamp symptoms, our guide to telling a data lake from a data swamp covers the warning signs in detail.
How to Read These Data Lake Examples
A data lake is a store for raw and refined data in open file formats, usually on cloud object storage, that several engines can read. The interesting part is what each organization did with it, so every example follows the same path.
Follow the Data From Source to Consumer
Each example starts with the sources, such as operational databases, event streams, files and SaaS apps. Then comes ingestion, which is batch, streaming or change data capture (CDC). After that, the data lands on object storage like Amazon S3, Azure Data Lake Storage or OneLake.
In Robinhood’s case, for example, Postgres is the source, Debezium handles ingestion and Hudi tables on Amazon S3 form the lake.
Inside the lake, data then moves through layers. Some teams call them raw, curated and serving, while others use the bronze, silver and gold names of the medallion architecture . Next, a table format and a catalog make files behave like tables. Finally, consumers read the result, whether that is BI, machine learning or an application.
Once you keep that path in mind, the examples become easy to compare. When you are ready to build one yourself, the step-by-step version lives in our data lake implementation guide .
What Counts as Documented Here
Every example comes from the organization’s own engineering blog, a Kanerika case study or a vendor’s published reference design. Because sources matter, third-party summaries and unverified vendor stories were left out. Paragraphs marked as a design note are Kanerika’s guidance for readers in that industry, so they are not claims about the company in the example.
Data Lake Examples at a Glance
The table lines up all eight examples by the three questions that matter most. Read across a row to see what went in and what came out, then jump to your industry.
Organization (year) Industry Storage and platform What went in What it fed
Walmart (2023) Retail Hadoop and Spark lakes moving to Apache Hudi Kafka business events, CDC from production databases, ML outputs Aggregation, reporting, analytics, ML Robinhood (2022) Financial services Amazon S3 with Apache Hudi tables Postgres changes via Debezium and Kafka Trino dashboards, product insights, reporting Microsoft reference design Healthcare OneLake bronze and silver lakehouses in Fabric FHIR records, DICOM imaging, CMS claims, SDOH data Clinical analytics, OMOP research models Kanerika pharma client Life sciences New Hadoop stack with improved pipelines Fragmented enterprise data sources 15+ use cases, self-service dashboards Kanerika packaging client Manufacturing OneLake in Microsoft Fabric with Purview SAP, Azure Synapse, SQL Server, Marketo Finance, HR and supply chain KPIs Uber (2018) Ride-hailing platform Hadoop (HDFS) with Parquet, then Hudi Raw data from online stores, ingested once Derived tables, analytics at 100+ PB scale Airbnb (2022) Travel marketplace Amazon S3, moving event data to Iceberg Event data, previously on HDFS and Hive Intraday and daily warehouse tables Netflix (2023) Streaming media Apache Iceberg tables Large datasets across the business A/B tests, studio production, model training
Two patterns jump out of the table. First, most of these lakes started with raw files and later added a table format. Second, almost all of them serve several consumers from one stored copy.
Two of the eight lakes, the healthcare design and the packaging client, run on Fabric’s OneLake. So if Fabric is on your shortlist, this Kanerika podcast asks why it is becoming the go-to data platform, with SSIS-to-Fabric moves in focus.
Retail Example: Walmart’s Data Lakes Move to a Lakehouse
Walmart’s engineering team described its lake in a May 2023 post titled Lakehouse at Fortune 1 Scale . It is one of the most detailed public retail data lake examples, because it explains what flows in and also why the design changed.
What Entered the Lake
High-volume streams of business events land first in a messaging layer backed mainly by Apache Kafka. Change data capture pulls updates from production databases, and machine learning and computer vision services add their own outputs. The domains Walmart names are supply chain, customer, inventory and search.
After that, the data feeds aggregation, reporting, analytics and ML work. So the same stored events serve reporting teams and data scientists across those domains.
Why the Design Changed
Batch-oriented lakes could not keep up once teams wanted near real-time data with warehouse-like behavior. Walmart wrote that it had begun to evolve its data lakes toward a lakehouse approach. The scale explains the stakes, since the company runs more than 600,000 cores of Hadoop and Spark across its internal cloud, Google and Azure.
To pick a table format, the team scored Apache Hudi, open-source Delta and Apache Iceberg with a weighted matrix. The criteria covered availability, compatibility, cost, performance, roadmap, support and total cost of ownership. In the end Hudi won, and migration of some critical workloads had started when the post went out.
Treat this as a 2023 decision, not a verdict on the formats. The author noted that the benchmark findings could go out of date, and all three formats have changed a lot since.
What Another Retailer Can Take From It
Design note. A table format choice is a multi-year commitment, so score it against your own engine mix and schema needs, not a vendor demo. Walmart’s lens included schema evolution for Protobuf, Avro and Thrift data, which matters when many teams publish events.
A typical retailer will also land point-of-sale data, e-commerce clickstream and inventory snapshots. Our post on retail data analytics shows what those sources tend to feed once they are in one place.
Kanerika Service
Data Architecture Services
Kanerika maps the sources, layers and table format your lake needs, then designs a target architecture on Fabric, Databricks or Snowflake that fits your engines.
Explore Data Architecture Financial Services Example: Robinhood’s CDC Data Lake on Amazon S3
Robinhood published its design in February 2022 under the title Fresher Data Lake on AWS S3 . The problem was simple to state. Core tables were copied once a day as snapshots, so analysts always worked with yesterday’s numbers.
The Stack, Layer by Layer
The stack is a clean example of change data capture feeding a lake. Each layer does one job, so the table below follows the data in order.
Layer What Robinhood used Job in the lake
Source Postgres on Amazon RDS Online transaction processing for the app Capture Debezium on Kafka Connect Reads database changes as they happen Transport Apache Kafka with Avro and Confluent Schema Registry Carries change events with a managed schema Table format Apache Hudi Writes lake tables incrementally Storage Amazon S3 Holds the data lake files Processing and query Apache Spark batch jobs, Trino, Hive Metastore, Airflow Batch pipelines, dashboards, schemas and orchestration
Robinhood also lists privacy primitives such as anonymization and access control as part of the ecosystem. That matters because, in a brokerage, account and trading data cannot sit in an open bucket.
What It Fed and What Changed
In the 2022 post, the lake fed business analytics dashboards, insights for product improvements, and business and ad hoc reporting. Before the change, freshness for critical tables was 24 hours, because daily snapshots were the bottleneck. With CDC and Hudi, freshness for those core tables dropped to under 15 minutes.
Fresher data came from changing the ingestion pattern, and the batch schedule mattered far less. Instead of re-copying whole tables, the pipeline only moves rows that changed.
Design Note for Banks
Banks and lenders usually add tighter access rules than a brokerage app’s analytics team needs. On AWS, AWS Lake Formation governs access to S3 data and Glue Data Catalog metadata at column, row and cell level. Those rules then apply across engines like Athena, Redshift Spectrum and EMR. So a fraud analyst and a marketing analyst can query the same table and see different slices of it.
Lineage is the other piece regulators ask about, because auditors want to trace a reported number back to its source rows. For regulated setups, see how banks approach data governance in banking . Then read how a data lakehouse for financial services handles these controls.
Healthcare and Life Sciences Examples: Clinical Data in OneLake and a Pharma Data Lake
Healthcare lakes carry the strictest data in this guide, so the boundary between raw and shareable data matters more than anywhere else. The two examples here are different in kind, since one is a vendor reference design and the other is a delivered client project.
A Reference Design for Clinical, Imaging and Claims Data
Microsoft documents healthcare data solutions in Microsoft Fabric as a set of ready-to-run pipelines. It is a vendor reference design rather than a customer deployment. Even so, it shows clearly what a clinical lake is expected to hold.
For example, the sources include FHIR-based clinical records, DICOM imaging data, CMS CCLF claims files and social determinants of health (SDOH) datasets. Exports from Azure Health Data Services also land in OneLake, and OMOP transformations prepare data for research. Unstructured clinical notes enrichment is available in preview.
The healthcare data foundations pipeline moves raw files into bronze and then silver lakehouses. Design note. Keep protected health information in restricted layers and de-identify it before wide use, because HIPAA obligations follow the data into every copy.
A Pharma Data Lake Serving 15+ Use Cases (Kanerika)
A multinational pharmaceutical company came to Kanerika with fragmented data that slowed its response to growing demand. Kanerika designed and deployed a data lake serving 15+ use cases. The build included a new, complete Hadoop stack with improved data pipes, plus new self-service dashboards.
The pharma data platform case study reports a 45% reduction in response times and a 40% increase in data accuracy. It also reports a 25% increase in data-driven decisions. The case study credits the gains to quicker access to accurate data from a single data lake. Our post on data analytics in pharma development covers where such a lake usually pays off first.
Case Study
45% Faster Response Times with Hadoop Data Platform
For a multinational pharmaceutical company, Kanerika designed and deployed a data lake serving 15+ use cases. The result was a 45% reduction in response times, a 40% increase in data accuracy and a 25% increase in data-driven decisions.
Read the Case Study → Manufacturing Example: One Governed Lake for a Global Packaging Leader
A multinational packaging company with a presence in over 100 countries had data spread across many systems and reports. Kanerika consolidated it in OneLake, the single data lake built into Microsoft Fabric . This example shows what governance looks like when it is designed into the lake rather than bolted on later.
What Went In
At first, the sources were spread across SAP, Azure Synapse, SQL Server and Marketo. The reporting estate was large, with more than 10,000 distinct measures across 400+ reports and 600 production dataflows. Many of those measures overlapped, so the same KPI could carry different numbers in different reports.
How It Was Organized and Governed
First, Kanerika unified the enterprise sources in OneLake for centralized, governed analytics. An automated metadata repository was built in a Fabric Lakehouse, and some legacy dataflows moved to Dataflow Gen2 and PySpark with CI/CD. Then naming conventions, role-based access control and audit trails were set up with Microsoft Purview .
What It Fed
As a result, finance, HR and supply chain teams now work from unified, consistent KPIs. The packaging case study reports a 60% increase in data accessibility, a 30% reduction in ETL processing time and a 45% improvement in decision-making.
Design Note: Adding Sensor Data
Many manufacturers also want to add plant data. Kanerika’s data engineering practice describes this as piping sensor, ERP and quality data into a unified layer. That layer then supports OEE tracking, defect analysis and production planning. That is the same pattern as the packaging lake, with a streaming source added at the front.
Keep one limit in mind, though. A data lake is analytical storage, not a control system, so machine control should never depend on a lake query. For a closer look at plant use cases, read about the data lakehouse for manufacturing .
Case Study
60% More Data Accessibility with Microsoft Fabric
Kanerika unified SAP, Azure Synapse, SQL Server and Marketo data in OneLake for a global packaging leader. The result was a 60% increase in data accessibility, a 30% reduction in ETL processing time and a 45% improvement in decision-making.
Read the Case Study → Tech Platform Examples: Uber, Airbnb and Netflix
The three tech companies below ran lakes at a scale most enterprises will never reach. Their write-ups still help, because they hit the same problems early and documented the fixes.
Uber: Raw Data In Once, Then the Small-Files Problem
Uber’s October 2018 post, Uber’s Big Data Platform , is candid about growing pains. Its first-generation warehouse was effectively being used as a data lake. So the second generation introduced a Hadoop data lake where all raw data from online stores was ingested only once, with no transformation during ingestion.
At first, that design scaled to tens of terabytes of new data per day, stored as Parquet. It also created two problems. Small files put pressure on the HDFS NameNodes, and new data was only available every 24 hours.
Uber’s answer was Hudi, short for Hadoop Upserts anD Incremental, built as an open-source Spark library over HDFS and Parquet. With it, ETL jobs finished in under 30 minutes, and end-to-end latency for all derived tables fell below one hour. By then the platform held over 100 PB.
Airbnb: From HDFS to S3, Then Iceberg
Airbnb’s September 2022 post on upgrading its data warehouse infrastructure picks up after an earlier move from HDFS to S3. After that move, the new limits were Hive metastore load, slow Hive and S3 interaction, schema evolution and fixed partitioning.
So the team moved event data ingestion to Apache Iceberg with Spark 3. Compared with the previous Tez and Hive stack, it reported more than 50% compute resource savings and a 40% reduction in job elapsed time.
Netflix: Iceberg Tables Behind A/B Tests and Studio Production
Netflix’s November 2023 post on incremental processing with Maestro and Apache Iceberg lists what its data feeds. That covers A/B tests, studio production, training algorithms, content acquisition, security breach detection and payments.
In that 2023 post, large datasets live in Iceberg tables. Netflix’s incremental processing captures table changes without copying data, which helps with freshness, accuracy and backfills. In other words, it is the same instinct Robinhood followed, applied at much larger scale.
Checklist
Data Engineering Checklist for Enterprise Teams
A practical list of pipeline, testing, orchestration and monitoring checks to run before your lake takes on its next source or consumer.
Get the Checklist → The Same Data Lake Design on AWS, Azure and Microsoft Fabric
Although the examples above run on different clouds, the blueprint barely changes. Storage holds open files, a catalog turns them into tables and an access layer decides who sees what. The table maps that blueprint to the three platforms most enterprise teams compare.
Building block AWS Azure Microsoft Fabric
Storage Amazon S3 Azure Data Lake Storage, which is Blob Storage with the hierarchical namespace turned on OneLake, built on Azure Data Lake Storage, one per tenant Catalog and access AWS Glue Data Catalog with Lake Formation permissions at column, row and cell level Directory-level structure and finer-grained security on the namespace Workspaces and items inside Fabric, with Purview for governance Open table format S3 Tables, fully managed Apache Iceberg tables since December 2024 Set by the engine you run on top Tables stored as Delta Parquet or Iceberg Engines and access Athena, Redshift Spectrum, EMR and Glue Hadoop-compatible access for Spark and other engines ADLS Gen2 API compatibility, plus shortcuts to S3, Google Cloud Storage and ADLS
Platform Notes for Each Cloud
On AWS, Amazon S3 Tables , announced on 3 December 2024, is one recent piece. AWS says they deliver up to 3x faster query throughput and up to 10x higher transactions per second than self-managed tables. Those are AWS’s own figures, so test them on your workload before you plan around them.
On Azure, Azure Data Lake Storage gives Hadoop-compatible access and folder-level security, which is why many older Azure lakes sit there.
In Fabric, Microsoft describes OneLake as a unified data lake for the whole organization. In fact, every Fabric tenant gets one automatically. Through OneLake shortcuts , it can also point at data in Amazon S3, Google Cloud Storage or ADLS without copying it. Our guides to the Microsoft Fabric data lake and OneLake shortcuts go further into both.
What These Data Lake Examples Have in Common
Put the eight designs side by side and five lessons repeat. None of them is new on its own, but seeing them in real systems makes them easier to act on.
Land raw data once. Uber ingested raw data from online stores only once, with no transformation on the way in. Every later fix depended on having that untouched copy.Freshness comes from incremental loads. Robinhood and Netflix both moved changed rows rather than re-copying whole tables, and that is where the speed came from.A table format is a long-term decision. Walmart scored three formats with a weighted matrix, and Airbnb rebuilt event ingestion around Iceberg. Neither choice was quick to undo.Small files and metadata load will arrive. Uber hit NameNode pressure and Airbnb hit Hive metastore limits, so plan for compaction from the start.Govern at the catalog, not the folder. Lake Formation permissions and Purview controls in the packaging case apply to tables and columns, which folder rules alone cannot do.The fifth lesson is the one that keeps a lake from becoming a swamp. A shared enterprise data catalog lets people find a trusted table before they build a new copy of it.
Most of These Lakes Became Lakehouses
There is also a pattern in where these lakes ended up. Walmart, Robinhood, Uber, Airbnb and Netflix all added a table format on top of their files. That step is what turns a data lake into a lakehouse. If you want the lines drawn clearly, read data lake vs data warehouse , then data lake vs lakehouse and what a data lakehouse is . When the real choice is between a lakehouse and a warehouse, our comparison of the data lakehouse vs data warehouse picks up from there.
Which Data Lake Example Matches Your Situation?
Start from the problem you have, and then copy the example that solved the closest one. The table pairs common needs with the matching design and a sensible first step.
Your situation Closest example Where to start
Product data is a day old when analysts need it in minutes Robinhood Run CDC on the three to five tables people query most Many teams report different numbers for the same KPI Kanerika packaging client Unify sources in one governed lake with a shared metadata layer Clinical or patient data has to stay restricted Microsoft healthcare reference design Keep a locked raw layer and de-identify before the curated layer Event volume is huge and growing fast Uber and Walmart Choose a table format early and plan compaction Teams wait on central IT for every new dataset Kanerika pharma client Build one lake for many use cases, with self-service dashboards An older Hadoop lake is costly and slow Airbnb Move storage to object storage, then adopt an open table format
For teams on an aging cluster, our guide to Hadoop to Databricks migration walks through the move. To see where the lake sits among ingestion, transformation and BI tools, read our overview of the modern data stack . Once curated tables exist, shaping them for reporting is its own discipline, which our post on dbt data modeling covers.
How Kanerika Builds Data Lakes That Teams Use
Kanerika’s data engineering services come in three engagement models, and a data lake project can start with any of them. The wording below is how the service page describes each one.
Pipeline Audit and Design. Map existing pipelines and gaps, define the target architecture and tooling, and deliver a build-ready technical blueprint.Build and Modernize. Rebuild legacy ETL on a cloud-native stack, implement lakehouse and streaming, and hand off tested, production-ready pipelines.Ongoing Support. Extend your team with dedicated engineers, monitor pipeline health and data quality, and ship new workstreams as priorities shift.The same page describes the lakehouse work as structured layers from raw to refined. It also lists quality checks run on every pipeline and lineage tracked across data movements. Those are the habits that kept the examples in this guide out of swamp territory.
Kanerika Service
Data Engineering Services
Kanerika designs, builds and runs data pipelines and lakehouses with structured layers from raw to refined. Every pipeline gets quality checks, and lineage is tracked across data movements.
Explore Data Engineering The two Kanerika lakes above show the results. For the pharma client, a data lake serving 15+ use cases cut response times by 45%. For the packaging client, one governed OneLake raised data accessibility by 60% and cut ETL processing time by 30%.
Kanerika is a Microsoft Solutions Partner. It also holds ISO 27001, ISO 27701 and SOC 2 Type II certifications, so security and audit controls are part of every build. Teams planning the target design first can also start with data architecture services , and regulated teams often pair the lake with data governance services .
Data Lake Examples Worth Copying
The best data lake examples answer the same three questions. What went in, how was it stored and organized, and what did it feed? Walmart and Robinhood show event and change data feeding analytics. Meanwhile, the healthcare and pharma cases show restricted layers feeding many users. Finally, the packaging lake shows governance feeding trusted KPIs.
In short, pick the example closest to your problem and copy its decisions rather than its size. If you want a second opinion on your own design, talk to Kanerika’s data engineers .
Frequently Asked Questions
What is a real-world example of a data lake?
Robinhood is a clear one. It streams changes from its Postgres databases through Debezium and Kafka into Apache Hudi tables on Amazon S3. Analysts query those tables with Trino, and freshness for core tables dropped from 24 hours to under 15 minutes. Walmart, Uber, Airbnb and Netflix have published similar designs on their engineering blogs.
What are data lake examples on AWS?
Robinhood runs its data lake on Amazon S3 with Apache Hudi tables, Spark batch jobs, Hive Metastore and Trino. A typical AWS design pairs S3 storage with the Glue Data Catalog and Lake Formation for access control. Since December 2024, S3 Tables also provide fully managed Apache Iceberg tables inside S3 for analytics workloads.
Is Amazon S3 a data lake?
Amazon S3 is the storage layer that many data lakes use, but a bucket alone is not a complete lake. A working lake also needs ingestion pipelines, a catalog such as AWS Glue, access rules such as Lake Formation and usually a table format. Engines like Athena, Spark or Trino then read the data for analytics.
Are Azure Data Lake Storage and OneLake the same thing?
No. Azure Data Lake Storage is a set of capabilities on Azure Blob Storage, enabled by the hierarchical namespace, that you provision per storage account. OneLake is the single data lake that comes with every Microsoft Fabric tenant. OneLake is built on Azure Data Lake Storage and supports the same ADLS Gen2 APIs.
What data goes into a data lake?
Almost any data an organization produces can land in a data lake. The examples in this guide hold database change records, business event streams, ERP and CRM extracts, marketing data, clinical FHIR records, DICOM images and claims files. The common rule is to land the raw copy once and refine it in later layers for each consumer.
How do banks and fintechs use data lakes?
Fintechs like Robinhood use change data capture to keep analytics tables fresh within minutes for dashboards and reporting. Banks add stricter controls on top, such as column, row and cell permissions in AWS Lake Formation, plus lineage so auditors can trace reported numbers back to source data. Fraud, risk and customer analytics are common uses.
What is the difference between a data lake example and a lakehouse example?
A data lake example stores raw and refined files on object storage for several engines. A lakehouse example adds an open table format such as Iceberg, Delta or Hudi, which brings transactions, schema control and faster queries to those files. Many lakes in this guide, including Walmart, Uber and Airbnb, later evolved into lakehouses.
Why do data lakes turn into data swamps?
A data lake turns into a swamp when data lands without owners, documentation or a catalog. Teams cannot find trusted tables, so they make new copies, and quality drops with every copy. The examples in this guide avoid it by landing raw data once, organizing it in layers and governing access at the catalog level.
