Data deskAerospike Cloud reaches general availabilityVoyager brings an agent-ready visual workspace to the databaseDatabase 8 adds distributed ACID transactions

Company profile / Enterprise infrastructure

The Database Built for the Worst Millisecond

Aerospike built its reputation in the least forgiving corner of software: decisions that must be correct before the next millisecond arrives. Now the database specialist is turning that discipline toward AI, graph workloads and a managed cloud business.

Every database wants to tell you it is fast. The more revealing question is what happens to the slowest request when a server fails, a cluster rebalances, a cricket match begins and ten million fans all reach for their phones. That ugly request lives at the far edge of a latency chart. For Aerospike, it is the whole plot.

The Mountain View company builds a distributed NoSQL database for systems that have very little time to think. An advertising platform may have milliseconds to decide which ad to bid on. A payments network needs to inspect a transaction for fraud while the customer is still staring at the spinner. A retailer has to fetch the right profile, inventory state or recommendation before attention moves elsewhere. Average speed is pleasant. Predictable speed under stress is the product.

Aerospike began in 2009 as Citrusleaf, founded by database engineers Srini Srinivasan and Brian Bulkowski. Their wager was partly about software and partly about hardware. Flash storage was getting faster and cheaper, but database designs still tended to treat disk as the slow room downstairs. Aerospike's Hybrid Memory Architecture put the primary index in DRAM while keeping data on SSDs. The intent was simple: retain memory-like responsiveness without paying to hold the entire dataset in expensive, volatile memory.

The seating chart stays in memory; most of the guests live on flash. A simplified view of Aerospike's signature architecture.

A database with a stopwatch

This architecture matters most at a scale where small inefficiencies multiply into buildings full of hardware. Aerospike says its shared-nothing design spreads data and work across nodes, automatically partitions records and supports replication across racks, data centers and regions. Operators can choose a mode biased toward availability or strong consistency. Database 8, introduced in 2025, added distributed ACID transactions for work that needs several records to change as one reliable unit.

The customer stories make the technical vocabulary tangible. Criteo, the advertising platform, reports handling roughly 200 million queries per second and 950 billion ad matches a day after consolidating a fleet of about 3,000 servers to 800. PhonePe says it processes more than 100 million payments each day, uses Aerospike for fraud and feature-store lookups, and cut the relevant server footprint eightfold. Airtel describes a real-time customer platform serving hundreds of millions of people and a roughly 90 percent reduction in total cost of ownership compared with its legacy setup.

200MApproximate Criteo queries per second
100M+PhonePe payments per day
90%Airtel-reported TCO reduction

These are vendor-published case studies, and every workload has its own physics. Yet the repeated unit is telling: not feature count, but machines retired, milliseconds controlled and incidents avoided. Aerospike is usually not the first database a small team reaches for to store a modest catalog. Its economics become interesting when traffic is high enough that per-request cloud pricing or all-memory infrastructure starts to feel like a tax on success.

“Our fraud detection system runs in real time, performing multiple checks in under 20 milliseconds out of Aerospike.”Burzin Engineer · Founder and Chief Reliability Officer, PhonePe

The product grew around the hard cases

The original key-value engine remains the center, but the perimeter has expanded. Aerospike Database now handles document data, vectors and transactions. Aerospike Graph combines the database storage layer with a scale-out service built on Apache TinkerPop and its Gremlin query language. Graph 3.0, released in July 2025, added developer conveniences and claimed tenfold faster ingestion with a storage footprint up to 50 percent smaller than the earlier version.

Connectors tie the platform to Kafka, Spark, Pulsar, Elasticsearch and Trino. Client libraries cover languages including Java, Python, Go, C, C#, Node.js, Rust, Ruby and PHP. Teams can run it themselves on bare metal, virtual machines or Kubernetes; hire Aerospike engineers to operate it in a cloud account; or use Aerospike Cloud, the self-service managed platform that reached general availability in January 2026.

That last option adjusts the business model as much as the product. Traditional Aerospike is expert infrastructure sold through enterprise subscriptions, support and hands-on services. Cloud packages the operations layer and prices around provisioned compute, storage, network transfer and management rather than charging for every transaction. For customers with a very busy database, the absence of a per-operation meter is not a small footnote.

Around the core sits a practical partnership network. AWS supplies marketplace distribution, infrastructure and integrations with services such as MSK and EMR; Aerospike has earned the cloud provider's Advertising and Marketing Technology Competency. A long technical collaboration with Intel explored persistent memory and high-density deployments. Apache projects supply familiar connective tissue through Kafka, Spark and TinkerPop. This is less a sealed product than a performance-conscious component meant to sit in an existing data estate.

The expertise is correspondingly specific. Aerospike engineers spend their time on data distribution, storage media, replication, consistency and the behavior of systems during failure. The company’s public voice has an operator’s bias: benchmarks are discussed alongside cluster size; performance claims arrive with server counts; customer praise often sounds like relief. Dream11 executive Amit Garde’s line that the database “doesn’t keep us awake at night” may be the least glamorous and most persuasive endorsement in infrastructure software.

Customer-reported server consolidation · selected workloads
Criteo before
3,000
Criteo after
800
PayPal before
360
PayPal after
20
A hardware diet with unusually specific before-and-after photos. Figures describe customer-cited workloads, not entire company infrastructures.

AI needs a memory that can keep up

Aerospike now describes itself as “the real-time database for AI.” The phrase could dissolve into conference-booth fog, but the underlying idea is concrete. Predictive systems need fresh features at inference time. Generative applications need user context and vector retrieval. Agents create intermediate state, pause, resume and call other systems. A model can be clever while the surrounding data path remains slow, stale or ruinously expensive.

Aerospike's bet is that key-value records, documents, graph relationships, vectors and transaction state can share a scalable operational core. Vector search can locate a neighborhood of similar items; a graph can then follow the relationships among them. That combination is relevant to fraud rings, identity resolution, recommendations and retrieval systems. It also places Aerospike in a busy market where cloud providers, specialist vector databases and established database companies all want to become the memory layer for AI.

Cache alternatives

Redis and Memcached trade in speed and familiarity. Aerospike argues for durability and a smaller all-memory bill at large scale.

Cloud key-value

DynamoDB, Bigtable and Cosmos DB offer deep cloud integration. Aerospike counters with deployment choice and no per-transaction cloud fee.

Distributed NoSQL

Cassandra, ScyllaDB and MongoDB meet broad operational needs. Aerospike concentrates on bounded latency and hardware efficiency.

Graph specialists

Neo4j and Neptune have graph mindshare. Aerospike approaches the category from a high-throughput storage engine.

The practical question for a buyer is not whether Aerospike wins a generic benchmark. It is whether the application's awkward shape matches the company's expertise. Test the 99th-percentile response, not only the average. Remove a node. Rebalance the cluster. Mix reads and writes. Price the replicas, network and operator time. A system that looks excessive at one terabyte can look economical at a petabyte; the reverse can be true at smaller scale.

A specialist learns to invite people in

Specialization creates a sales challenge. A product known for expert tuning and enormous deployments can intimidate developers who merely want to try it before lunch. The Community Edition, Docker images, public client libraries and extensive documentation have long formed the on-ramp. In 2026, Aerospike made that entrance more visual.

Voyager is a free desktop workspace for Mac, Windows and Linux. It browses records, builds filters, troubleshoots queries and includes safeguards for potentially expensive operations. It also ships with an MCP server so coding agents in tools such as Claude or Cursor can work with a cluster. Alongside new Java and Python SDKs, Voyager turns the first encounter from a command-line exercise into something closer to an inspectable workbench.

This is the company’s balancing act. Aerospike must become easier to buy and easier to use without losing the characteristics that made it valuable when the load becomes unpleasant. It must sell AI infrastructure without pretending that a database alone makes an AI product. And it must broaden from one sharp key-value use case into a platform without becoming a vague catalog of modes.

That tension also shapes its culture. The company describes a mission to “unleash the power of the world's real-time data,” but its more distinctive habits are observational: publish architecture, expose client code, test failure modes and let customers talk in operational measurements. Aerospike has kept an open-source-friendly route through Community Edition while reserving enterprise features, support and managed operation for paying customers. The resulting culture feels closer to a systems lab with a sales team than a consumer-tech brand chasing daily attention.

The Aerospike promise is not that traffic spikes disappear. It is that the database stops making them dramatic.

The name comes from a rocket nozzle designed to remain efficient across changing altitudes. It is an unusually apt metaphor for enterprise infrastructure: nobody buys the nozzle for its own sake. They buy it because the rest of the vehicle has somewhere urgent to go. Aerospike’s customers are chasing a payment, a bid, a fraud signal, a recommendation or the next useful piece of context. The database has one quiet assignment - answer before the moment is gone.

Distributed databasesReal-time dataEnterprise AIDeveloper toolsCloudNoSQL