Redshift Data Warehouse Modernization Playbook 2026

Amazon Redshift Data Warehouse Modernization Playbook (2026): Zero-ETL, Serverless, and Spectrum

Quick summary: For a retailer DW exit (~42 TB logical, 18 nightly Glue jobs), Aurora→Redshift zero-ETL retired 11 jobs and cut dashboard freshness from 6h → ~15 min — concurrency scaling on zero-ETL (March 2026) absorbed Monday open spikes.

Key Takeaways

In March 2026, Redshift announced concurrency scaling support for auto-copy and zero-ETL, so ingest peaks can add compute instead of stalling dashboards (What's New)
It is not Serverless vs provisioned tier choice, not the S3 Tables / Iceberg lake reference architecture, and not a Glue-only ETL tutorial
Benchmark silhouette (not a cited client) — Retail analytics, legacy DW ~42 TB logical, 18 nightly Glue jobs, dashboard freshness ~6 hours
After Aurora + RDS zero-ETL into Redshift Serverless and Spectrum for cold history: 11 Glue jobs retired, freshness ~15 minutes on operational facts
Monday open spikes absorbed via concurrency scaling on zero-ETL ingest (post–March 2026)

Zero-ETL integrations make operational data available in Amazon Redshift without you maintaining a classic ETL fleet for supported sources — including Aurora MySQL/PostgreSQL, RDS MySQL/PostgreSQL/Oracle, DynamoDB, and listed SaaS/apps (Redshift zero-ETL). In March 2026, Redshift announced concurrency scaling support for auto-copy and zero-ETL, so ingest peaks can add compute instead of stalling dashboards (What’s New).

This is the warehouse modernization playbook. It is not Serverless vs provisioned tier choice, not the S3 Tables / Iceberg lake reference architecture, and not a Glue-only ETL tutorial.

Artifacts: path matrix, TCO worksheet, architecture diagram (draw.io).

Benchmark silhouette (not a cited client) — Retail analytics, legacy DW ~42 TB logical, 18 nightly Glue jobs, dashboard freshness ~6 hours. After Aurora + RDS zero-ETL into Redshift Serverless and Spectrum for cold history: 11 Glue jobs retired, freshness ~15 minutes on operational facts. Monday open spikes absorbed via concurrency scaling on zero-ETL ingest (post–March 2026). Modeled monthly TCO drop ~$30k → ~$24k before legacy license exit (worksheet).

Modernization sequence

Assess — freshness SLOs, license cost, job inventory
Land — zero-ETL for supported sources; Glue only for true transforms
Size — Serverless default; RA3 if flat-high baseline (tier post)
Park cold — S3 / Iceberg + Spectrum; lake detail in modern data lake post
Cut over — BI first, then delete duplicate jobs

Opinionated take: Delete Glue jobs only after consumers read the zero-ETL schema for two billing cycles. Early deletion is how you recreate the pipeline under an incident bridge.

Reference flow

Aurora / RDS / DynamoDB ──► Zero-ETL ──► Redshift (Serverless or RA3)
On-prem / complex transforms ──► Glue ──► Redshift + S3
S3 cold / Iceberg ──► Spectrum / lake query ──► same BI tools

Target prerequisites include case sensitivity and IAM/resource policy setup (configure Redshift target).

Path matrix

Use modernization-path-matrix.md to choose Serverless vs RA3 vs zero-ETL vs Spectrum-heavy designs.

What broke — case sensitivity

What broke — Day 2 of zero-ETL. Integration stuck / tables mismatched because enable_case_sensitive_identifier was off on the Serverless workgroup. Detection: integration status + missing relations in the destination database. Fix: enable case sensitivity, recreate destination DB from integration, replay validation queries. Four hours lost — documented as a hard prerequisite in the runbook.

What to Do This Week

List every nightly job and tag replicate vs transform.
Enable zero-ETL on one Aurora or RDS source into a Serverless workgroup.
Turn on case sensitivity and confirm concurrency scaling eligibility in your Region.
Fill the TCO worksheet with your license line items.

What This Post Doesn’t Cover

Detailed Serverless RPU tuning (see tier-choice post)
Full lakehouse governance (SageMaker Catalog operating model)
Streaming-first architectures (Kinesis/MSK)
Oracle Exadata-specific exit tooling beyond DMA/partner programs

Frequently asked questions

When should we NOT use zero-ETL?

When the source is unsupported, you need heavy mid-flight transforms before analytics, or you only need a one-time historical dump. Use Glue/Spark or DMS-style pipelines for those cases.

Serverless or provisioned for the modern warehouse?

Default Serverless for spiky BI. Choose RA3 provisioned when baseline compute is flat and high 24×7. Details in the Serverless vs provisioned post — this playbook focuses on the modernization program.

What could go wrong with zero-ETL cutover?

Case-sensitivity (`enable_case_sensitive_identifier`) left off on the Redshift target, missing IAM/resource policies, and assuming Glue jobs can be deleted before consumers switch. Validate row counts and critical metrics before decommissioning ETL.

Where does Spectrum fit?

Cold/historical data on S3 (including Iceberg/lake patterns). Keep hot facts in Redshift local/managed storage. Do not Spectrum your entire hot path.

How is this different from the modern data lake S3 Tables post?

That post designs the lake/Iceberg plane. This playbook modernizes the warehouse consumer plane and operational replication into Redshift.

Did concurrency scaling change zero-ETL behavior in 2026?

Yes — March 2026 GA added concurrency scaling support for auto-copy and zero-ETL ingestion on Serverless and RA3, helping peak ingest without freezing BI reads.