The New York Times cut Oracle migration time from 18 months to 18 weeks

The NYT’s engineering team, together with AWS and the AWS partner Mactores, replaced a sprawling Oracle estate that spanned three schemas, 110 000 conversion items and roughly 6 million lines of code across Java, Groovy, Python, Go and shell scripts. By leveraging AWS Database Migration Service (DMS) Schema Conversion for bulk DDL changes and the Mactores Aedeon Agent—an AI‑native tool built on Amazon Bedrock’s Claude Opus 4.6—the project automated both schema transformation and the tedious remediation of embedded Oracle SQL. Continuous replication kept the source and target in sync, while a reverse‑replication path allowed a zero‑risk cut‑over that preserved full production throughput at go‑live. The effort eliminated about 200 CPU‑worth of Oracle licensing and positioned PLATO for multi‑region Aurora deployment.

The migration illustrates how generative‑AI services are moving from experimental to production‑grade utilities in enterprise data modernization. Traditional rule‑based converters struggle with heterogeneous codebases; Bedrock’s language‑agnostic model can ingest mixed‑grammar files and generate PostgreSQL‑compatible replacements, dramatically shrinking manual effort. AWS’s bundling of DMS, Schema Conversion, Step Functions and Bedrock creates a turnkey stack that rivals specialist migration consultancies, potentially reshaping the competitive dynamics between cloud providers and third‑party migration firms. As more media and subscription‑based businesses confront costly legacy licenses, the NYT case may accelerate adoption of AI‑augmented migration pipelines across the industry.

Looking ahead, the success hinges on the fidelity of AI‑generated code and the robustness of the rollback mechanism. Any subtle semantic drift in translated SQL could surface under peak traffic, risking data integrity or performance regressions. Organizations should therefore invest in extensive automated testing and staged rollouts, and monitor the evolving licensing model for Bedrock, which could affect total cost of ownership. Additionally, the repeatability of this approach will be tested as NYT scales PLATO across regions and integrates newer Aurora features, offering a real‑world barometer for AI‑driven migrations at enterprise scale.