Adikteev delivers predictable real-time bidding at massive scale

Adikteev is a global mobile marketing platform specializing in performance-driven user acquisition and retargeting for mobile apps and games. Its real-time bidding platform evaluates advertising opportunities in milliseconds, using live user behavior data and machine learning models to determine whether to bid on an impression and how much to bid.

Adikteev’s real-time bidding platform must make decisions within strict latency limits while processing massive traffic volumes. Each ad impression request triggers multiple database lookups used to evaluate user behavior and determine a bid price.

Massive real-time scale

The system stores more than 1.3 billion user counters and processes up to 1.5 million database queries per second during peak traffic. Bid decisions must complete within an API timeout of roughly 10 milliseconds, requiring extremely fast and predictable data access.

Update-heavy workload

The platform continuously updates user activity counters using rolling time windows. As a result, roughly two-thirds of the dataset must be rewritten every day, creating extremely heavy write workloads alongside high read throughput.

Cassandra and ScyllaDB performance limits

Adikteev initially ran the system on Apache Cassandra and later migrated to ScyllaDB to improve performance. However, the update-heavy workload created compaction pressure that degraded read performance.

Adikteev needed a data platform capable of sustaining extremely high read and write rates while maintaining consistent latency under volatile traffic conditions.

After evaluating multiple technologies as alternatives to ScyllaDB, including Redis and RocksDB, Adikteev selected Aerospike to power the real-time data layer of its bidding platform.

Designed for predictable performance

Aerospike’s architecture keeps indexes in memory while storing data efficiently on disk. This design enables near in-memory performance while dramatically reducing RAM requirements and maintaining predictable latency even under extremely high throughput.

Production-ready before peak season

Adikteev ran ScyllaDB and Aerospike in parallel for roughly two months, executing the same queries against both systems and comparing results while gradually shifting production traffic. This extended production test confirmed that Aerospike could sustain the full workload with consistent sub-millisecond latency, allowing the team to complete the migration ahead of the Black Friday traffic surge.

"With ScyllaDB, we constantly had to think about compaction and how it would affect performance during peak traffic. With Aerospike, the system simply behaves predictably, even under heavy load."

Seiji Fouquet

AdikteevSRE

With Aerospike, Adikteev achieved predictable performance at massive scale while significantly improving reliability and infrastructure efficiency.

Sub-millisecond latency at scale

The platform now maintains p99 latency below one millisecond and p95 latency around 500 microseconds, even while serving extremely high request volumes.

Reliable performance under peak load

Database error rates improved to below 0.01%, ensuring no disruption to bidding operations.

1.5M queries per second across 1.3B keys

Aerospike supports more than 1.5 million database queries per second, enabling the bidding platform to evaluate advertising opportunities in real time at global scale.

Lower infrastructure cost with greater headroom

The new architecture reduced infrastructure cost by 39% compared with the previous ScyllaDB deployment, while dramatically shrinking the legacy cluster footprint by 81%. Even while operating both systems during the transition period, Adikteev achieved a net infrastructure savings of roughly 20%.