Skip to Content

Netflix RDS PostgreSQL to Aurora PostgreSQL Migration – Engineering Overview

22 February 2026 by
Suraj Barman

Context & History of Netflix’s Relational Database Modernization

In 2024 the Online Data Stores team at Netflix evaluated the relational database stack across the company. After reviewing functionality, performance, and total cost of ownership, the team standardized on Amazon Aurora PostgreSQL as the primary relational offering. The migration program, spanning 2024‑2025, aims to move existing workloads from Amazon RDS PostgreSQL to Aurora with minimal disruption, supporting a unified and scalable data platform for thousands of services.

Implementation & Best Practices for Self‑Service Aurora Migrations

Before diving into technical details, the migration workflow is organized into four phases: preparation, replication setup, quiescence and cut‑over, and post‑cut‑over validation. This roadmap ensures each step is repeatable and auditable, allowing service owners to execute migrations without deep database expertise.

Preparation Phase – Inventory and Baseline

The first step is to capture a complete view of database consumers. Netflix uses an eBPF‑based network attribution tool called flowlogs to record active TCP flows and enumerate every service that connects to the source RDS instance. This data feeds into a communication channel where stakeholders receive timelines and readiness checks. For a broader understanding of inventory techniques, see the Wikipedia article on asset management.

Replication Setup – Aurora Read Replica Creation

Automated backups must be enabled on the source RDS instance. A dedicated Aurora parameter group is cloned from the source configuration, preserving memory settings, connection limits, and planner behavior. The workflow then creates an Aurora read replica, which continuously streams write‑ahead logs (WAL) from the RDS source. This approach reduces downtime compared with snapshot‑only migrations.

Quiescence and Cut‑Over – Ensuring Consistency

During quiescence, write traffic to the RDS instance is stopped. To guarantee a clean state, the workflow detaches the RDS security groups and reboots the instance, forcing all sessions to close. Replication lag is monitored via the OldestReplicationSlotLag metric; a drop to zero followed by the expected 0→64 MB oscillation confirms that all meaningful WAL has been applied. For deeper insight into PostgreSQL replication, consult the official PostgreSQL documentation.

Post‑Cut‑Over Validation – Traffic Redirection and CDC Reset

After promotion, the Envoy‑based Data Access Layer updates its runtime configuration to point to the new Aurora writer endpoint. Because applications connect through the DAL, no code changes are required. Change Data Capture (CDC) slots are cleared before cut‑over and rebuilt on Aurora, triggering a full backfill of existing rows via REFRESH events. This guarantees that downstream consumers resume from a consistent baseline.

Key Takeaways

  • Self‑service workflow eliminates manual, error‑prone steps.
  • Read‑replica strategy achieves sub‑10‑minute write downtime.
  • Automated security‑group detachment provides deterministic quiescence.
  • CDC slot management prevents WAL retention issues.

For a practical guide on building reliable automation pipelines, refer to the internal article on monetizing an AI‑powered SaaS. Additionally, the discussion on grounded retrieval‑augmented generation illustrates how Netflix applies rigorous testing frameworks to data‑intensive services.