The market for managed cloud databases is expanding faster than almost any other layer of enterprise infrastructure. It was valued at $19.95 billion in 2024 and is projected to reach $49.78 billion by 2030, a compound annual growth rate of 16.7%. The pull is simple: companies want the performance and scale of distributed databases without the staffing, the patching schedules, and the 3 a.m. pages that come with running the clusters themselves. The harder question is whether a managed service can actually behave like the database it replaces.
Varsha Ganesh has spent more than 14 years building the systems that make that substitution possible. A Senior Software Development Engineer at Amazon Web Services and a Senior Member of the IEEE, she specializes in distributed databases and load balancing infrastructure, and she was one of the founding engineers behind Amazon Keyspaces, the company’s managed, serverless service for Apache Cassandra workloads. Her work sits at the point where the promise of a managed database meets the engineering reality of delivering it, where a customer’s existing application must continue working without requiring a single line of code to change.
The Cost of Owning the Cluster
For the teams running their own database clusters, availability is not an abstraction. More than 90% of mid-size and large enterprises now report that a single hour of downtime costs them upward of $300,000, and 41% put the figure above $1 million. Apache Cassandra earned its popularity precisely because it was built to stay available across many machines and regions, but that resilience carries a standing tax: capacity planning, version upgrades, repair cycles, and node failures that someone has to be awake to handle.
Keyspaces was built to remove that tax. Cassandra is one of the most widely used distributed data stores in the world, and the customers who lean on it hardest tend to operate large clusters with heavy maintenance overhead. The service Ganesh helped launch offers a serverless alternative that speaks the same interface and holds the same feature set, so a team can scale capacity in a day instead of provisioning hardware, and can do it without interrupting live traffic. The point was never to build a new database. It was to let people stop operating the one they already had.
“Most customers do not want a database. They want the thing their application needs the database to do,” says Varsha Ganesh. “Our job was to take away everything in between, the capacity math, the upgrades, the failure handling, and leave them with just the part they actually care about.”
Compatibility Is the Hard Part
Claiming compatibility is easy. Delivering it is where most managed services quietly fall short. A migration does not fail because an API is missing a method. It fails because a query that returned rows in one order on the old system returns them differently on the new one, or because an edge case in how data is filtered behaves a hair off from what the application expects. For a team betting a production workload on the move, those small divergences are the difference between a clean weekend cutover and a rollback.
Ganesh owned and architected the complete query processing layer for Keyspaces, the part of the system that translates Cassandra Query Language into operations against the underlying storage engine. Reaching genuine parity meant writing custom transformation logic to filter, reorder, and reshape the storage layer’s results so that every query pattern matched what Cassandra would have returned. In the largest regions today, the service runs at millions transactions per second, and every one of those requests travels through the critical path she designed at launch.
“Parity is not a feature you ship once. It is a contract,” Ganesh notes. “If a single query type behaves differently under load, you have not built a compatible system. You have built a similar one, and similar is what blocks a migration.”
What Parity Unlocks
The demand behind this work keeps climbing. Unstructured data already accounts for 93% of enterprise information assets, and flexible, schema-light databases like Cassandra are how a large share of it gets stored and queried. When that volume lives in a self-managed cluster, every capability the database lacks in its managed form becomes a reason a team cannot move. Feature gaps are not cosmetic. They decide whether a workload migrates at all.
After Keyspaces became generally available, several large customer migrations stalled on a single missing capability: time-to-live, the ability for rows and columns to expire automatically without application-side cleanup. Teams running session tokens, event logs, IoT telemetry, and time-series metrics treated TTL as a core design assumption rather than a convenience. Ganesh led the project to build it, coordinating 4 teams across the DynamoDB organization to rearchitect the underlying storage and compaction mechanisms. Delivering it brought Keyspaces to parity on one of Cassandra’s most-used features and, in the process, cleared the path for global replication, which depended on the same conflict-resolution machinery.
“TTL looks like a small feature from the outside. Inside, it touched the most sensitive parts of the storage engine,” Ganesh reflects. “But it was the thing standing between a whole class of customers and the migration they wanted to make. You fix the unglamorous blocker, and the door opens.”
Engineering Change Without Downtime
At the scale these systems operate, the hardest constraint is rarely the new feature itself. It is making the change to a system that cannot stop. A live database serving production traffic across regions does not get a maintenance window, and a fix that requires one is not a fix that customers can accept. Every alteration has to land underneath running workloads, invisibly, with the old behavior intact until the moment the new behavior is ready.
The TTL implementation was unusually intrusive by that measure. Every record written to Keyspaces had to support TTL metadata, including the records with no expiration set at all, which meant the change reached deep into a path already handling millions of transactions per second. Ganesh’s team deployed it without downtime and without disturbing the existing feature set, and the deletion system that removes expired records now runs at similarly high throughput in the busiest regions. The discipline that makes that possible, the careful sequencing and the backward compatibility held at every step, is the part of the job that never shows up in a release note.
“Nobody writes a press release about a deployment that changed nothing the customer could feel,” Ganesh observes. “But that silence is the whole point. The best infrastructure work is the work nobody notices.”
The Next Layer of Invisible Infrastructure
The pattern repeats one layer down. As more of the world’s data moves onto managed services, the systems that keep those services running, the control planes that create, scale, patch, and recover the infrastructure underneath, become the quiet determinant of whether the whole thing stays reliable. A control plane is the part customers never see and depend on completely, the brain coordinating every routine operation and every recovery in the background.
“Across databases and load balancers, the goal has not changed for me,” says Ganesh. “Take the operational weight off the customer and carry it inside the system, so the people building on top of us can think about their product instead of our plumbing. If we have done it right, they forget we are here. That is the entire job.”
