Automatic hot/cold data tiering keeps hot data in memory while persisting the full dataset to a high-performance storage pool, delivering millisecond-level latency for high-frequency interactions. 

A multi-tier storage architecture spanning memory, SSDs, and object storage scales smoothly to hundreds of terabytes of capacity while significantly reducing storage and O&M costs.

Built on a distributed architecture with decoupled storage and compute, the system can scale compute nodes in minutes and storage in seconds without interrupting services. 

GeminiDB combines an advanced symmetric architecture, in which all nodes support reads and writes, with a multithreaded parallel processing engine to enable high-speed ingestion of massive feature datasets. It delivers up to 5× the data ingestion performance of open-source Redis and supports real-time reads at million-level QPS with millisecond-level latency.

To meet the storage demands of massive volumes of historical data used for tagging and feature engineering, GeminiDB combines decoupled storage and compute with multi-tier storage. The full dataset is persistently stored with high reliability, while only hot data remains in memory, significantly reducing storage costs compared with traditional fully in-memory solutions.

GeminiDB provides strongly consistent, real-time persistence with three-replica data redundancy. In the event of a hardware failure, the cluster can automatically fail over and take over compute workloads within seconds, helping ensure continuous operation of upstream ad bidding services.

Native support for Hash, List, Sorted Set, and other core data structures provides data primitives needed for complex game logic, significantly accelerating development and iteration efficiency.

Built to handle the sudden concurrency spikes typical of gaming workloads, GeminiDB supports seamless storage scale-up in seconds and smooth scale-out of compute nodes in minutes, helping maintain responsive player experiences and uninterrupted core services during peak hours.

GeminiDB uses decoupled storage and compute combined with a high-performance persistent storage pool to provide hundreds of terabytes of storage. Intelligent hot/cold data tiering keeps hot data in memory while moving cold data to the storage pool, significantly reducing storage costs compared with traditional fully in-memory solutions and providing cost-effective storage for massive volumes of player asset and equipment data.

Purpose-built for time-series workloads, GeminiDB delivers 4.5× the small-batch write performance and 3.3× the large-batch write performance of open-source InfluxDB, easily handling high-concurrency ingestion spikes from massive fleets of IoT devices.

Cold data is automatically archived to lower-cost storage media, reducing storage costs by up to 90%. Hot data and cold data are accessed in the same way, requiring no application code changes or adaptation. The hot/cold data age threshold can also be flexibly defined and dynamically adjusted based on business requirements.

Built-in efficient data conversion and transfer mechanisms automatically convert online time-series data to standard Parquet files and export them to object storage, where they can serve as a core input source for big data processing engines. This enables deeper mining and complex analysis of massive datasets.