On January 4, 2016, right before Netflix expanded into 130 new countries, its billing infrastructure went 100% AWS cloud-native — splitting one Oracle database into Cassandra for subscribers and MySQL for payments, because payment processing needed ACID transactions.

Netflix’s billing infrastructure tracked open and paid billing periods, account credit, payment status, charge requests, and the date each member had paid through — plus feeding financial systems for revenue and tax reporting. All of it ran against a single, giant Oracle-based system in Netflix’s own data center, with integrations reaching into member accounts, payment processing, customer service, and DVD shipping.

The scale was not small: billions of rows of constantly changing data, accumulated since Netflix’s founding in 1997, growing every minute in one large shared database.

Netflix disintegrated the single Oracle system into a services-based architecture and split the data across two purpose-built stores. Subscriber data moved to Cassandra, giving it the horizontal, multi-region scale the APIs needed. Payment processing integration needed ACID transactions, so all relevant data moved to MySQL instead — open-source, running on Netflix-managed EC2 instances, with a multi-region master using DRBD copy and read replicas in different regions.

On January 4, 2016 — just before Netflix expanded into 130 new countries — billing infrastructure became 100% AWS cloud-native.

The migration ran incrementally, not as a single cutover. Netflix built proxies and redirectors to keep applications in the data center working, unimpacted, until they were ready to move. New country launches gave the team a chance to test the new Cloud infrastructure on clean data first, before extending it to large, legacy countries like the US, which carried both streaming and DVD billing.

Migrating existing members’ data required real downtime: subscription data moved from Oracle to Cassandra country by country, with retry capability for failed processing and optimistic customer-state handling so members weren’t penalized while processing was halted. For the payments side, Netflix used Oracle GoldenGate to replicate tables — and their ongoing incremental changes — across the heterogeneous Oracle-to-MySQL boundary while production kept running, over a period of months. Weeks before the final switch, they ran their test database continuously against MySQL, fixing issues there before validating against Oracle and releasing to production.

Netflix’s own reflection is candid: “We underestimated testing automation needs. We did not have a good way to test end to end flows.” Being a SOX system added its own tax — all migration tooling needed to be SOX-compliant, with auditability of deployments and actions built in along the way, using Spinnaker, Atlas, and a comparator built to reconcile Cassandra data against Oracle by country.

Post-migration, Netflix reports being more efficient and lighter in software footprint, with applications able to scale horizontally to keep up with subscriber growth — and, in their words, this was “a major cross functional engineering effort” spanning core platform, security, database engineering, tooling, and big data platform teams.

When one dataset needs ACID guarantees and another needs multi-region scale, splitting them into two purpose-built stores can beat forcing both through the same database.