# Redshift DW modernization path matrix (2026)

| Scenario | Redshift Serverless | Redshift provisioned RA3 | Zero-ETL into Redshift | Glue/Spark + Spectrum/S3 | Stay on legacy DW + federation |
|----------|---------------------|--------------------------|------------------------|--------------------------|--------------------------------|
| Spiky BI / unknown concurrency | **2** | 1 | 1 | 1 | 0 |
| Steady 24×7 heavy ETL + predictable RPU | 1 | **2** | 1 | 1 | 0 |
| Near-real-time from Aurora/RDS/DynamoDB | 1 | 1 | **2** | 0 | 0 |
| Large historical cold data on S3/Iceberg | 1 | 1 | 0 | **2** | 0 |
| Oracle/SQL Server DW exit with fixed price risk | 1 | 1 | 0 | 1 | 0 — plan exit |
| Need datashares to multiple consumer accounts | **2** | **2** | 1 | 0 | 0 |

## Opinionated default

1. Land **operational** replicas via **zero-ETL** where source is supported (Aurora MySQL/PostgreSQL, RDS MySQL/PostgreSQL/Oracle, DynamoDB, and listed apps).
2. Put the **consumer warehouse** on **Redshift Serverless** unless you have flat, high baseline compute — then RA3.
3. Keep **cold/historical** in S3 (Iceberg/S3 Tables) and query via Spectrum / lakehouse patterns — do not lift every archive byte into local Redshift storage.
4. Use concurrency scaling for auto-copy and zero-ETL ingestion peaks (GA announced March 2026).

## When NOT to

- Do not rebuild a full Glue CDC pipeline for a source that already has zero-ETL.
- Do not choose provisioned "because we always did" without a 30-day Serverless baseline.
- Do not treat Spectrum as a substitute for modeling hot facts — hot path belongs in Redshift local/managed storage.