Platform comparison

Databricks vs Redshift

If you're all-in on AWS, the work is SQL reporting and dashboards, and your sources are Aurora, RDS, DynamoDB or SaaS apps that zero-ETL integrations can replicate, Redshift is the shorter path. Choose Databricks when data engineering, streaming, ML and AI agents matter as much as BI, or when you need the same platform outside AWS.

Databricks. Pipelines feeding models you serveRedshift. Zero-ETL feeds into an AWS warehouse

Checked: Against both sides' own documentation, October 10, 2026
Sources: 27 official pages, listed below
Written by: TechFabric engineers. We're a Databricks partner, and we say where the other one wins.

Talk through your case Talk to an engineer (30 min)

How each is built

Two different shapes, drawn from each one's own architecture docs.

Databricks

A data and AI platform that runs on AWS, Azure and Google Cloud, keeps tables in open Delta Lake or Apache Iceberg format in cloud object storage, and governs them with Unity Catalog.

Redshift

AWS's managed SQL data warehouse, run either as provisioned clusters on RA3 or the newer Graviton-based RG nodes, or as Redshift Serverless billed by capacity used.

Which one

When each one is the right call.

Choose

Databricks when

Python and Spark data engineering, ML and AI agents carry as much weight as dashboards, and you want them on one platform with one catalog.
You run on more than one cloud, or expect to, since Redshift runs only on AWS.
You want your main tables to be Iceberg or Delta in S3 by default, governed by one catalog that also runs outside AWS.
You want a managed Postgres for applications next to the analytics data; Lakebase is generally available on AWS.

Choose

Redshift when

Your operational data sits in Aurora, RDS, DynamoDB or apps such as Salesforce, SAP, ServiceNow or Zendesk, and you want it replicated into the warehouse without building pipelines.
Your team writes SQL, the workload is warehouse-shaped, and you'd like Redshift Serverless to scale against a price-performance target you set.
You stream from Kinesis or Amazon MSK and want the data landing in SQL materialized views with exactly-once ingestion.
You're committed to AWS services such as Lake Formation, SageMaker Unified Studio and Amazon Bedrock, and want the warehouse wired into them.

Side by side

10 differences, side by side.

Dimension	Databricks	Redshift
Architecture	A lakehouse: compute (clusters, serverless jobs and SQL warehouses) is separate from tables stored in cloud object storage, with Unity Catalog as the governance layer across them.	A massively parallel SQL warehouse. RA3 and RG nodes keep data in Redshift Managed Storage, billed separately from compute, and Redshift Serverless measures capacity in Redshift Processing Units (RPUs).^21,22,1,7
Storage and open table formats	Delta Lake by default. Unity Catalog managed Iceberg tables, foreign Iceberg tables and Iceberg v3 features became generally available on 21 May 2026, and external engines can read and write managed Iceberg tables through the Iceberg REST Catalog API.	Redshift Managed Storage by default. Redshift reads Iceberg v1 to v3 tables in the AWS Glue Data Catalog and can create and write Iceberg tables in Amazon S3 and S3 table buckets; Iceberg v3 needs Serverless or RG nodes.^15,16,3,4,5
Compute model	Choice of classic clusters in your cloud account, serverless compute, and SQL warehouses; serverless SQL warehouses typically start in 2 to 6 seconds.	Provisioned clusters on RA3 or RG nodes, or Serverless with a base capacity you set in RPUs (128 by default). RG nodes, generally available since 12 May 2026, run data lake queries on their own compute instead of Redshift Spectrum.^22,1,2,7
Data engineering and pipelines	Lakeflow covers ingestion connectors (Lakeflow Connect), Spark Declarative Pipelines for batch and streaming in SQL or Python, Lakeflow Jobs for orchestration, and Lakeflow Designer for visual data preparation.	Zero-ETL integrations replicate Aurora, RDS (MySQL, PostgreSQL, Oracle), DynamoDB, self-managed databases and apps such as Salesforce, SAP and ServiceNow into Redshift. Spark transformation runs in separate services such as AWS Glue or Amazon EMR.^20,6,10
Machine learning and AI	Training, tracking and serving stay on one platform. MLflow tracks experiments and manages the model lifecycle, Model Serving deploys custom models and LLMs as REST endpoints, and Agent Bricks builds governed agents; its Knowledge Assistant became generally available in January 2026. Models are governed in Unity Catalog with the tables.	Redshift ML trains models from SQL with CREATE MODEL, using SageMaker AI Autopilot (billed by SageMaker), and can call Amazon Bedrock foundation models for text tasks from SQL.^26,25,21,9
Governance and catalog	Unity Catalog governs tables, views, volumes, functions and models with grants, attribute-based policies, row and column filters, and automatic lineage. An open-source implementation also exists.	AWS Glue Data Catalog is the technical catalog and AWS Lake Formation provides fine-grained access, including on Iceberg tables Redshift queries and writes.^21,3,10
Streaming	Structured Streaming and Spark Declarative Pipelines handle streaming in the same code as batch. Real-time mode in Structured Streaming has been generally available since March 2026, and real-time mode in Spark Declarative Pipelines entered Public Preview in May 2026.	Streaming ingestion reads Kinesis Data Streams, Amazon MSK, Confluent Cloud or Apache Kafka topics into materialized views, with exactly-once processing; joins aren't supported directly on the streaming view.^20,27,15,8
Operational databases (OLTP)	Lakebase is a managed Postgres database inside the platform, generally available on AWS since January 2026, with autoscaling, scale to zero, branching and tables synced from Unity Catalog.	Redshift is not a transactional database. The AWS pattern is to run Aurora, RDS or DynamoDB and replicate into Redshift with zero-ETL.^17,18,19,6
Cloud availability	Runs on AWS, Azure and Google Cloud.	AWS only.^24,1
Lock-in and portability	Tables sit in your own S3 buckets as Delta or Iceberg, and any Iceberg client, inside AWS or not, can read and write managed Iceberg tables through Unity Catalog's REST endpoint. Lakeflow pipelines and jobs are Databricks-specific.	Native tables live in Redshift Managed Storage, which AWS now exposes to Iceberg engines through the Glue Iceberg REST catalog in the SageMaker lakehouse architecture.^16,10

Architecture^21,22,1,7

Databricks: A lakehouse: compute (clusters, serverless jobs and SQL warehouses) is separate from tables stored in cloud object storage, with Unity Catalog as the governance layer across them.

Redshift: A massively parallel SQL warehouse. RA3 and RG nodes keep data in Redshift Managed Storage, billed separately from compute, and Redshift Serverless measures capacity in Redshift Processing Units (RPUs).

Storage and open table formats^15,16,3,4,5

Databricks: Delta Lake by default. Unity Catalog managed Iceberg tables, foreign Iceberg tables and Iceberg v3 features became generally available on 21 May 2026, and external engines can read and write managed Iceberg tables through the Iceberg REST Catalog API.

Redshift: Redshift Managed Storage by default. Redshift reads Iceberg v1 to v3 tables in the AWS Glue Data Catalog and can create and write Iceberg tables in Amazon S3 and S3 table buckets; Iceberg v3 needs Serverless or RG nodes.

Compute model^22,1,2,7

Databricks: Choice of classic clusters in your cloud account, serverless compute, and SQL warehouses; serverless SQL warehouses typically start in 2 to 6 seconds.

Redshift: Provisioned clusters on RA3 or RG nodes, or Serverless with a base capacity you set in RPUs (128 by default). RG nodes, generally available since 12 May 2026, run data lake queries on their own compute instead of Redshift Spectrum.

Data engineering and pipelines^20,6,10

Databricks: Lakeflow covers ingestion connectors (Lakeflow Connect), Spark Declarative Pipelines for batch and streaming in SQL or Python, Lakeflow Jobs for orchestration, and Lakeflow Designer for visual data preparation.

Redshift: Zero-ETL integrations replicate Aurora, RDS (MySQL, PostgreSQL, Oracle), DynamoDB, self-managed databases and apps such as Salesforce, SAP and ServiceNow into Redshift. Spark transformation runs in separate services such as AWS Glue or Amazon EMR.

Machine learning and AI^26,25,21,9

Databricks: Training, tracking and serving stay on one platform. MLflow tracks experiments and manages the model lifecycle, Model Serving deploys custom models and LLMs as REST endpoints, and Agent Bricks builds governed agents; its Knowledge Assistant became generally available in January 2026. Models are governed in Unity Catalog with the tables.

Redshift: Redshift ML trains models from SQL with CREATE MODEL, using SageMaker AI Autopilot (billed by SageMaker), and can call Amazon Bedrock foundation models for text tasks from SQL.

Governance and catalog^21,3,10

Databricks: Unity Catalog governs tables, views, volumes, functions and models with grants, attribute-based policies, row and column filters, and automatic lineage. An open-source implementation also exists.

Redshift: AWS Glue Data Catalog is the technical catalog and AWS Lake Formation provides fine-grained access, including on Iceberg tables Redshift queries and writes.

Streaming^20,27,15,8

Databricks: Structured Streaming and Spark Declarative Pipelines handle streaming in the same code as batch. Real-time mode in Structured Streaming has been generally available since March 2026, and real-time mode in Spark Declarative Pipelines entered Public Preview in May 2026.

Redshift: Streaming ingestion reads Kinesis Data Streams, Amazon MSK, Confluent Cloud or Apache Kafka topics into materialized views, with exactly-once processing; joins aren't supported directly on the streaming view.

Operational databases (OLTP)^17,18,19,6

Databricks: Lakebase is a managed Postgres database inside the platform, generally available on AWS since January 2026, with autoscaling, scale to zero, branching and tables synced from Unity Catalog.

Redshift: Redshift is not a transactional database. The AWS pattern is to run Aurora, RDS or DynamoDB and replicate into Redshift with zero-ETL.

Cloud availability^24,1

Databricks: Runs on AWS, Azure and Google Cloud.

Redshift: AWS only.

Lock-in and portability^16,10

Databricks: Tables sit in your own S3 buckets as Delta or Iceberg, and any Iceberg client, inside AWS or not, can read and write managed Iceberg tables through Unity Catalog's REST endpoint. Lakeflow pipelines and jobs are Databricks-specific.

Redshift: Native tables live in Redshift Managed Storage, which AWS now exposes to Iceberg engines through the Glue Iceberg REST catalog in the SageMaker lakehouse architecture.

Cost

How each one bills you.

Databricks

Databricks bills in Databricks Units (DBUs), a normalized unit of processing, metered per second and priced per product and cloud on its price list. You pay as you go or commit to usage for discounts. Storage, networking and, for classic compute, the EC2 instances are billed by AWS.

Redshift

In US East (N. Virginia), Redshift Serverless costs $0. 375 per RPU-hour, billed per second with a 60-second minimum, and Serverless Reservations discount up to 24% (one year) or 45% (three years). Provisioned nodes bill per hour, for example $3.

26 for ra3. 4xlarge and about $3. 04 for rg. 4xlarge in that region. Managed storage is $0. 024 per GB-month there, and Spectrum lake queries from RA3 or DC2 clusters cost $5. 00 per TB scanned.

List prices change and committed-use discounts are negotiated, so compare quotes for your own workload rather than these units.^13,1

Using both

You don't always have to choose.

Databricks can query Redshift in place through Lakehouse Federation, pushing filters, joins, aggregates and sorts down to Redshift, though not window functions, and it can't reach Redshift's own external tables that way.

Iceberg tables registered in the AWS Glue Data Catalog are shared ground: Redshift reads and writes them, and Databricks reads them as foreign Iceberg tables, generally available since May 2026.

Sources ^14,4,15

Moving

Moving from Redshift to Databricks

01
Inventory the warehouse
Inventory schemas, stored procedures, Python UDFs (AWS ended support for them after 30 June 2026, so they need a new home anyway), zero-ETL integrations, streaming materialized views and BI connections.
02
Map schemas and permissions
Map Redshift databases and schemas to Unity Catalog catalogs and schemas, and IAM and Lake Formation permissions to Unity Catalog grants.
03
Federate during the move
Connect Redshift through Lakehouse Federation so teams can query it from Databricks during the move.
04
Unload and reload the data
Move bulk data with UNLOAD to Parquet in Amazon S3, then load into Delta or managed Iceberg tables; UNLOAD to Parquet drops the time zone from TIMESTAMPTZ values, so check those columns.
05
Convert SQL and feeds
Convert Redshift SQL with Databricks' agentic code converter (Beta, lists Redshift SQL as a source dialect), and rebuild zero-ETL and Kinesis or MSK feeds as Lakeflow Connect ingestion and Spark Declarative Pipelines.
06
Reconcile and retire
Run both in parallel, reconcile row counts and aggregates table by table, then repoint dashboards and pause or delete the cluster or workgroup.^{12,11,14,23,6}

Migration Readiness Sprint

Two weeks that inventory the warehouse and hand you a wave plan before anything moves

FAQ

Databricks vs Redshift: what people ask

Is Redshift cheaper than Databricks?

The meters differ, so test with your own queries. One Redshift charge depends on the node type. Data lake queries from RA3 clusters go through Redshift Spectrum at $5. 00 per TB scanned in US East (N.

Virginia), while RG nodes and Serverless run them on their own compute with no per-TB charge. Databricks bills DBUs per second plus AWS charges. Run a representative week on both and compare the bills.

Can Databricks read Redshift data without moving it?

Lakehouse Federation queries Redshift in place and pushes filters, joins, aggregates and sorts down to it. It can't query Redshift's own external tables. If you query the same Redshift data heavily every day, copy it.

Does Redshift support Apache Iceberg?

Yes, for reads and writes. Redshift queries Iceberg v1 to v3 tables in the AWS Glue Data Catalog and can create, insert into, update and merge Iceberg tables in Amazon S3. Iceberg v3 needs Redshift Serverless or RG nodes, and time travel queries on Iceberg aren't supported.

What are Redshift RG instances?

They are Graviton-based provisioned nodes, generally available since 12 May 2026. AWS says they're up to 2.2 times as fast as RA3 on data warehouse workloads and up to 2.4 times as fast on data lake workloads, at 30% lower price per vCPU. Lake queries on them run on cluster compute with no Spectrum per-TB charge.

We rely on Python UDFs in Redshift. Does that matter?

It does. AWS ended support for Python UDFs in Redshift after 30 June 2026 and recommends Lambda UDFs instead, so those functions need rewriting whichever platform you choose. Count them early in any plan.