Data lake: Definition, architecture, use cases and platform guide.

Adobe for Business Team

09-09-2026

Every organisation generates more data than its existing systems were designed to hold and most of that data arrives in formats that traditional databases reject. A data lake solves this problem by accepting everything first and asking questions later, giving analysts and data scientists a single source of truth to work from.

This article answers the following questions:

• What is a data lake?
• How does a data lake work?
• What are the primary use cases for a data lake?
• Why do you need a data lake?
• When do you need a data lake?
• How does a data lake compare with a data warehouse and a data lakehouse?
• What is data lake architecture?
• How do you deploy a cloud data lake?
• What are the benefits and challenges of data lakes?
• How do you choose the right data lake platform?

What is a data lake?

A data lake is a centralised repository that stores structured, semi-structured and unstructured data in its original, raw format at any scale. Unlike a relational database, data is transformed or rejected at the time of ingestion. Clickstream event logs, IoT sensor readings, social media feeds, PDF documents and CRM exports can all coexist in a single lake without pre-processing.

Three roles interact with data lakes most often. Data engineers build the ingestion pipelines that feed the lake. Data scientists run exploratory models against raw data that has not been pre-aggregated or filtered. Analytics engineers prepare curated datasets from the lake for business intelligence (BI) tools and dashboards. Each role depends on the lake for different reasons, but all benefit from a single shared repository rather than scattered copies of the same source data.

The term data lake was coined to contrast with the concept of a data mart. A data mart provides a small, filtered pool of data optimised for a specific purpose, such as regional sales reporting. A data lake provides the whole body of data for any purpose you define later. This distinction matters because it changes depending on the design decisions made. In a data mart, you decide which questions to answer before storing the data. In a data lake, you store data first and decide which questions to answer when a business need arises.

Data lakes are the backbone of modern data platforms because they decouple storage from compute. You store everything once and apply different processing engines, whether SQL, machine learning (ML) or streaming, to the same underlying files without moving them. For example, a retailer can store two years of raw point-of-sale events in a data lake and run a demand forecasting model and a weekly revenue report against the same dataset, each using a different compute engine, without duplicating a single file.

How does a data lake work?

Data flows into a lake through an ingestion layer that pulls it from source systems, including databases, APIs, event streams and flat files and lands it in a raw storage zone without altering it. Even schema errors in the source system are preserved, which helps protect against data loss during upstream incidents. If a CRM vendor changes a field name overnight, the lake still captures the records and engineers can reconcile the change during the cleansing step rather than lose the data entirely.

Once landed, data moves through a series of processing zones. The raw zone contains untouched sourced data. The cleansing zone contains validated and deduplicated records. The curated zone stores business-ready datasets. This three-zone pattern prevents analysts from accidentally querying unclean data while preserving the original for reprocessing. A common reason for failure is skipping zone separation entirely, which can lead analysts to build dashboards on uncleaned data and discover errors only after reports have been shared with executives.

The compute layer is separate from the storage layer. SQL engines, ML frameworks and stream processing tools all read from the same files. For example, the same customer event log can simultaneously feed a fraud-detection model and a marketing attribution report without copying data. This separation also means that scaling compute resources for a demanding processing job does not require adding more storage, and vice versa.

A metadata catalogue sits above the storage layer and tracks what data exists, where it lives, who can access it and how current it is. Without a catalogue, a data lake can become a data swamp, where files accumulate with no way to discover, manage or govern them. Organisations that delay implementing a catalogue typically find that, within six months, data engineers spend more time searching for datasets than analysing them.

Data egress connects the data lake to downstream consumers. BI dashboards, AI and ML pipelines, customer experience platforms and operational databases all pull data from the curated zone rather than maintaining separate copies. This single-egress pattern helps eliminate the three versions of the total revenue problem that plagues organisations with siloed data shops.

What are the primary use cases for a data lake?

Data lakes support a wide range of analytical and operational workloads. The most common use cases include customer analytics, ML, security monitoring, IoT, compliance reporting and real-time decision-making.

• Customer 360 profiling. Organisations ingest behavioural data, such as web clicks, app events and purchase history, along with CRM records and support tickets, into a single data lake to build unified customer profiles. For example, a retail bank might combine branch-visit data, mobile app session data and call-centre transcripts into a single data lake so that a downstream customer data platform (CDP) can assemble a single view of each customer. This is a foundational use case for any business that delivers personalisation at scale. Adobe Experience Platform reflects this pattern — its governed data lake stores raw and processed customer events as the system of record, while a separate real-time customer profile feature merges that data into a unified view of each customer for activation across channels.
• ML and AI model training. Data scientists need access to large volumes of historical, unfiltered data to train accurate models. A data lake serves as the primary training data repository for recommendation engines, churn-prediction models and demand forecasting systems. In these use cases, data variety and volume directly contribute to model performance. A summarised or aggregated dataset from a data warehouse is often insufficient for training robust ML models.
• Log analytics and security monitoring. Application logs, server events and network traffic generate large volumes of data and schema variability that traditional databases struggle to handle efficiently. Storing this data in a data lake and running SQL or streaming queries against it allows real-time anomaly detection and post-incident forensic analysis. For instance, a security operations team at a SaaS company might ingest billions of authentication events per day and query them for brute-force attack patterns within minutes.
• IoT and sensor data processing. Manufacturing, logistics and utilities companies route device telemetry directly into a data lake because the data arrives continuously and in proprietary binary formats. Analysts process the data later, either in batches or via stream processing, to detect equipment failures before they happen. A wind farm operator, for example, might store turbine vibration data in a data lake and run predictive maintenance models weekly.
• Regulatory archiving and audit. Financial services and healthcare organisations use data lakes as a cost-efficient, immutable archive layer. Retaining raw transaction or patient-record data in object storage costs a fraction of what it would cost to store it in a database, while still keeping the data queryable for compliance audits.
• Real-time analytics and data streaming. Modern data lakes support streaming data ingestion, meaning that event data from a website or mobile app can be queried within seconds of an event occurring. This capability allows use cases such as real-time offer decisioning and live A/B test measurement, where waiting for an overnight batch load would make the insights stale.

Why do you need a data lake?

Traditional databases reject data that doesn't match a predefined schema. When a new data source arrives, whether from a new mobile app, a third-party partner feed or a sensor network, a data lake can ingest it immediately while schema design happens in parallel. This cuts the time to insight from weeks to hours. For example, a marketing team launching a new mobile app can start capturing raw event data on day one rather than waiting for the data engineering team to model a schema.

A single data lake eliminates the proliferation of siloed copies. Without one, each team typically builds its own extract, copies it to a local server and transforms it differently. The result is three versions of total revenue that never align. A data lake with a shared catalogue creates a single authoritative source that every team can query.

Cost efficiency is another major driver. Storing raw data in a cloud data lake's object storage, such as Amazon S3, Azure Data Lake Storage or Google Cloud Storage, costs significantly less per terabyte than storing equivalent data in a relational database. Organisations that migrate large volumes of historical data from a warehouse to a data lake often find that storage costs drop by 80% or more for cold-data tiers.

Data lakes also future-proof the analytics stack. Because data is stored in open file formats such as Parquet, Optimised Row Columnar (ORC) or Delta Lake, organisations are not locked into a single vendor's processing engine. They can switch from one SQL engine to another without re-ingesting their data.

Finally, AI and ML at scale require a data lake. AI models need access to the complete, raw history of a dataset, not a summarised or aggregated snapshot. A data lake is therefore structurally necessary for any organisation building predictive or generative AI capabilities at enterprise scale.

When do you need a data lake?

• When the data volume outgrows the database. Once a transactional database is being queried for analytics, it slows down for both workloads. A data lake offloads analytical queries and protects operational systems from query contention. A midsize ecommerce company running reporting queries against its production order database is a textbook candidate.
• When data variety exceeds the schema. The moment a business needs to combine structured records (orders, accounts and more) with unstructured content (emails, PDFs and images) or semi-structured feeds (JSON events, XML files and more), a schema-enforced database becomes a bottleneck. A data lake handles all three natively.
• When teams need exploratory access. If data scientists request raw data exports from the database administrator (DBA) team on a weekly basis, a data lake with governed self-service access removes that bottleneck entirely. Analysts can query the lake directly using appropriate access controls and the DBA team can focus on operational database health.
• When batch analytics latency is too high. If the business needs to act on data within minutes rather than relying on overnight batch windows, combining a data lake with a streaming ingestion layer, data pipeline and stream processing engine can reduce latency to near real time without replacing the batch layer.
• When regulatory retention requires low-cost archiving. Industries with multi-year retention mandates, such as financial services, healthcare and legal services, often use a data lake as a tiered archive. Hot data stays in a warehouse for fast querying, while cold data is stored in the data lake at a fraction of the cost.

How does a data lake compare with a data warehouse and a data lakehouse?

A data warehouse shops pre-modelled, structured data optimised for fast BI queries. It enforces a schema-on-write approach, meaning data must be cleaned and transformed before it is loaded. In contrast, a data lake follows a schema-on-read approach, allowing raw data to be stored first and structured only when it is queried.

The practical trade-off is that a warehouse is faster for known, recurring queries (such as monthly revenue by region), but it is less suited to storing unstructured data or supporting ad hoc exploration of raw events. A data lake may be slower for structured queries, but it can accommodate any data type and support ML workloads that a traditional data warehouse cannot.

A data lakehouse is a newer hybrid architecture that combines the open storage capabilities of a data lake with ACID transaction support and performance optimisations of a data warehouse. Technologies such as Delta Lake and Apache Iceberg add table-level metadata, versioning and time-travel capabilities on top of object storage, providing warehouse-style reliability and governance at data-lake scale.

Key differences: Data lake vs. Data warehouse vs. Data lakehouse

Choosing between the three depends on the dominant workload. If 80% of queries are standard BI reports, a data warehouse is typically the most appropriate choice. If the organisation runs ML pipelines or needs to store raw event streams, a data lake or lakehouse is often necessary. Today, most enterprise data platforms use all three technologies within a tiered architecture.

What is data lake architecture?

Data lake architecture describes the layered set of components that govern how data enters, moves through, and exits the lake. The five core layers are ingestion, storage zones, processing, catalog and governance and consumption. Each layer has a distinct responsibility. Conflating these layers, for example, by transforming data during ingestion, can create technical debt that is expensive to unwind.

• Ingestion. Connectors and data pipelines pull data from source systems on a scheduled basis or in real time via data streaming. Batch connectors handle database exports and file transfers, while streaming connectors handle event queues (such as Kafka and Kinesis) and IoT telemetry. The choice between batch and streaming data ingestion depends on how quickly downstream consumers need the data. A nightly product catalogue sync is well-suited to batch processing, whereas a fraud alert triggered by a credit card transaction requires streaming.
• Storage zones. The raw zone stores immutable source data. Data should never be overwritten or deleted, unless required by retention policies. The cleansing zone contains validated, de-duplicated records, while the curated zone holds business-ready datasets that are often aggregated. This separation ensures that if a bug is discovered in the cleansing logic, the data can be reprocessed from the raw zone without information loss.
• Processing and computing. Distributed processing engines (such as Apache Spark, Presto, Trino and Flink) read data from storage zones and write the results to downstream zones or external systems. Processing can be batch-based (scheduled overnight jobs), micro-batch (every few minutes) or streaming (continuous). Stream processing is increasingly common in use cases such as real-time personalisation and fraud detection. The separation of compute from storage allows processing capacity to scale independently of storage costs.
• Cataloguing and governance. A data catalogue registers every dataset with its schema, lineage, owner, classification and access policies. Without this layer, data engineers often spend more time searching for data than analysing it and compliance audits become manual, error-prone processes. Data quality rules and lineage tracking are applied at this layer, not in the storage layer.
• Consumption. BI tools, AI and ML platforms, APIs and operational databases connect to the curated zone. A well-designed consumption layer exposes clean, versioned datasets via a semantic layer or data API, so downstream consumers are isolated from changes in the underlying storage structure.

How do you deploy a cloud data lake?

Cloud data lakes are built on object storage services that decouple storage from compute. The three dominant cloud providers each offer a managed object store that serves as the foundation of a data lake: Amazon S3 (AWS), Azure Data Lake Storage Gen2 (Microsoft Azure) and Google Cloud Storage (Google Cloud). All three support open file formats, fine-grained access control and tiered storage pricing.

• Lift-and-shift. Migrate existing on-premises Hadoop Distributed File System (HDFS) clusters to cloud object storage. This approach preserves existing processing logic but does not take full advantage of cloud-native elastic compute. It is best suited for organisations that have made significant investments in Hadoop and need to reduce infrastructure overhead quickly.
• Cloud-native build. Design the data lake from scratch using managed cloud services for ingestion (data ingestion tools such as AWS Glue, Azure Data Factory or Google Cloud Dataflow), storage (object store), processing (managed Spark or Presto) and cataloguing (AWS Glue Data Catalogue, Microsoft Purview or Google Dataplex). This approach reduces operational overhead but typically requires existing pipelines to be re-architected.
• Lakehouse on cloud. Layer a table format (such as Delta Lake, Apache Iceberg or Apache Hudi) over cloud object storage to add ACID transactions, schema evolution and time-travel queries. This approach is used by platforms such as Databricks and is increasingly becoming the default for new enterprise data lake builds. Organisations adopting this pattern often pair it with a streaming data platform to support both real-time and batch workloads on the same storage layer.

Key governance decisions should be made at deployment, not retrofitted. Define retention policies before data accumulates, because retroactive clean-up is expensive. Enforce column-level encryption for personally identifiable information (PII) at the storage layer. Implement tag-based access controls so new datasets inherit the right policies automatically.

Cost management in a cloud data lake requires active monitoring. Storage costs are low but scale linearly with data volume. Without lifecycle policies that move cold data to archival tiers, organisations are likely to see their lake costs growing faster than their data value. Compute costs are the larger risk: unoptimised queries on a large lake can generate significant processing spend in a single run.

What are the benefits and challenges of data lakes?

The core benefits of a data lake are well established:

• Schema flexibility. A data lake can accept virtually any data format without requiring upfront modelling, allowing new data sources to be onboarded in hours rather than weeks.
• Cost-efficient storage. Object storage is significantly cheaper per terabyte than traditional database storage, making multi-year retention economically viable.
• Support for AI and ML. Data scientists can access complete, raw historical data without requesting extracts from a DBA team.
• Single source of truth. One governed repository eliminates redundant copies and the reconciliation efforts associated with maintaining them.
• Independent scalability. Storage and compute resources can scale independently, so spikes in ingestion volume do not force compute upgrades.

The challenges are equally significant. The risk of creating a data swamp is the most cited concern. Without enforced metadata, lineage tracking and access governance, a data lake can become an unmanageable accumulation of files. Organisations that skip the catalogue layer during deployment consistently find that data engineers spend more time locating data than using it.

Data quality is not automatic. A data lake accepts all incoming data, including corrupt records, duplicate events and schema drift from source systems. Data quality must be enforced within the cleansing zone through explicit validation rules. It does not emerge from storage alone.

Security and compliance complexity also increase with a data lake. Storing PII, financial records and healthcare data in a single lake with operational data creates a broad attack surface. Therefore, column-level encryption, attribute-based access control and audit logging must be implemented as foundational architectural decisions rather than afterthoughts.

From an organisational perspective, a data lake requires cross-functional ownership. Different teams often manage storage, ingestion, governance and consumption. Without a designated data platform team and a shared data contract standard, the lake can fragment into team-specific zones, recreating the silos it was meant to eliminate.

Finally, performance trade-offs exist. Ad hoc queries that are run against large, unpartitioned datasets in a data lake are slower than equivalent queries executed in a data warehouse. Organisations that need both exploration flexibility and fast BI response times typically adopt a data lakehouse architecture or maintain a separate data warehouse fed by the curated zone of the data lake.

How do you choose the right data lake platform?

The right data lake platform depends on your organisation's current infrastructure, dominant workload and governance requirements. Four common organisational conditions map to different platform types:

1. Cloud-native with no legacy Hadoop infrastructure and a preference for managed services. Choose a cloud-native data lakehouse platform such as managed Databricks, Snowflake with external tables or Google BigLake.
2. Significant existing investment in AWS, Azure or Google Cloud. Extend the native data lake services of your existing cloud provider before evaluating third-party platforms. The integration overhead of adding a new vendor often outweighs the marginal feature advantages.
3. Dominant workload focused on customer data activation and real-time personalisation rather than analytics. Evaluate a CDP that pairs a data lake with a dedicated customer profile store. For example, Adobe Experience Platform uses a governed data lake as the system of record for ingested data. It feeds a separate Real-Time Customer Profile, which unifies customer data and powers real-time segmentation and journey activation. The data is ingested once and shared across analytics and activation use cases rather than copied into a separate marketing stack.
4. Enterprise environment with multi-cloud or hybrid requirements. Evaluate platforms that support open table formats, such as Delta Lake and Iceberg, to reduce the risk of vendor lock-in at the storage layer.

Every buyer should apply a consistent set of evaluation criteria:

• Open format support. Can your data be accessed using a different processing engine if you switch vendors?
• Governance maturity. Does the platform include a built-in catalogue, lineage tracking and column-level security or are these capabilities available only as add-ons?
• Streaming and batch parity. Can the platform handle both real-time data streaming and overnight batch processing on the same infrastructure?
• Total cost of ownership. Consider egress charges, compute costs for inactive clusters and the engineering effort required to maintain pipelines, not just storage price per terabyte.
• Ecosystem integrations. How many of your existing source systems and downstream tools connect natively and how many require customised connectors?

For enterprise data lake deployments at scale, the platform decision is inseparable from the data governance model. Organisations that choose a platform with strong governance primitives (such as attribute-based access, audit logs or data contracts) consistently report lower remediation costs when compliance requirements change. Evaluating an enterprise data lake solution should include a proof of concept that tests governance under realistic data volumes and team access patterns, not just data ingestion speed.

Before committing to any data lake platform, assess three factors: the number of data sources to connect within the first 90 days, whether the data science team needs notebook access to raw data or works exclusively with curated datasets and the compliance classification of the data being ingested. These three factors will help to determine whether you need a general-purpose data lake, a lakehouse or a domain-specific platform such as a CDP.

How Adobe can help your business.

If unifying customer data for real-time personalisation and activation is your priority, explore how Adobe Experience Platform combines a governed data lake for storage and analytics with a real-time customer profile built to unify customer data for segmentation and journey orchestration at enterprise scale.

Learn more about what Adobe Experience Platform can do for you.

https://business.adobe.com/fragments/resources/cards/thank-you-collections/rtcdp