Predictable P99s. True elasticity. And unmatched price performance. ScyllaDB delivers predictable performance at any scale, handling millions of ops/sec, billions of vector embeddings, and petabytes of storage.
ScyllaDB is a shard-per-core, lock-free, hardware-optimized database designed for predictable low-latency and massive throughput. With autotuning, data rebalancing, elastic scaling and workload prioritization, ScyllaDB helps you hit SLAs without the usual database drama.
Millions
Billions
< 10ms
Performance bottlenecks inevitably surface as you scale data-intensive applications. Adding nodes and caches might buy time, but it doesn’t fix the underlying problem: you need a more efficient database.
Sustain massive concurrent traffic with smaller, more efficient clusters.
Maintain predictable single-digit millisecond P99s, even under heavy load.
Reduce infrastructure & networking costs, avoid overprovisioning.
Scale linearly with growth and elastically with traffic fluctuations and spikes.
Power fast vector search & feature retrieval in massive scale inference pipelines.
Switch without rewriting your apps;CQL & DynamoDB APIs are supported.
Survive node, rack, or even full-region failures with active-active replication.
Avoid manual tuning – CPU, memory, and I/O balance automatically.
Escape cloud vendor lock-in; run across clouds, regions, or on-prem.
Where extreme performance meets massive scale.
Best for write‑heavy, query‑driven data models where partition and clustering keys are well-defined.
<1 ms
Reliable P99 latency
+1M ops/sec
Millions of ops/sec
Best when vector search is a core application function and data freshness matters for accurate AI context.
250 QPS
Single digit P99 latency
+1B Vectors
High dimensionality
ScyllaDB’s close-to-the-metal architecture delivers millions of ops/sec per node with single-digit millisecond P99 latency. This means fewer nodes, less admin, and lower costs.
ScyllaDB runs one shard (thread) per CPU core with isolated memory and async I/O. This shared nothing architecture eliminates locking, providing linear scalability as you add cores/ or nodes.
Userspace schedulers adjust I/O and CPU dispatch to maximize hardware utilization. This prevents background tasks from starving latency-sensitive queries.
ScyllaDB caches hot data directly in memory on each CPU core. It self-tunes based on access patterns and memory pressure to keep reads fast as workloads change.
Tablets partition data into small chunks that dynamically rebalance across the cluster. This provides elastic scaling – helping you handle peaks with lower infrastructure costs.
ScyllaDB lets you control how your workloads compete for system resources. This ensures latency-sensitive queries are fast, even with other heavy workloads running on the same cluster.
Tripadvisor uses ScyllaDB on AWS to power real-time ML personalization. At peak, they handle ~500K ops/sec with P99 latencies of 1-3 ms. Their feature store serves up to 5 million static features/sec and 0.5 million user features/sec.
Freshworks uses ScyllaDB to power its AI-driven data platform (e.g., feature stores, model training caching layers, workflow automation, and customer data services). After migrating to ScyllaDB, they hit single-digit P99.99 latency and reduced storage 50%.
ShareChat uses ScyllaDB as the backbone of its ML feature store that recommends fresh social media content in 15+ different languages. They scaled from 1M to 1B features per second with 10X cost reduction and P99 latencies from 10-20ms.
ZEE5 uses ScyllaDB to power “Continue Watching” and recommendations across 190 countries. The platform maintains single-digit millisecond P99 latency while processing 1M+ concurrent requests and 1TB of daily state changes.
ScyllaDB was designed for efficiency, with the goal of delivering predictable low tail latency at scale. ScyllaDB’s close to the hardware design leverages a shard-per-core architecture and autotuning capabilities to maximize hardware utilization. This efficiency also translates to lower cost: less hardware is required to run similar workloads on ScyllaDB than with Cassandra or DynamoDB. Also, ScyllaDB runs anywhere; it’s available as a Database as a Service and can be deployed on any public cloud, or even on-premises. See NoSQL database comparisons and benchmarks.
ScyllaDB’s design includes a built-in efficient cache layer. This internal cache ensures predictable low latency for workloads at scale. For this reason, many teams have replaced caching layers (e.g., Redis, ElasticCache, DAX) with ScyllaDB. In DynamoDB cases, ScyllaDB can replace the underlying database (through its support of the DynamoDB-compatible API) as well as replace the cache associated with it. ScyllaDB’s design helps reduce overall costs by making efficient use of infrastructure, as well as running anywhere. Why teams are replacing their external cache with ScyllaDB.
ScyllaDB is designed for low latency at scale, including flexible scaling to meet growing needs. Workloads that have thousands to millions of operations per second, as well as multiple terabytes or petabytes of data, will get the greatest benefits from ScyllaDB. ScyllaDB is designed for applications that work with semi-structured or structured data and query that data with known/predictable patterns. High cardinality with evenly distributed access patterns is also helpful. Is ScyllaDB a good fit.
ScyllaDB’s performance-focused design relies on its shard-per-core architecture, enabling efficient CPU utilization with a shared-nothing approach. Seastar, the framework ScyllaDB is built upon, allows for maximizing concurrency and reducing the latency of operations. ScyllaDB also bypasses OS-level memory management, performing direct I/O operations and leveraging its internal cache. Since ScyllaDB is written in C++, it reduces the complexity usually associated with tuning JVM parameters and avoids Java’s garbage collection pauses. What Makes ScyllaDB So Fast?
ScyllaDB is a NoSQL database designed for high performance, high throughput, and predictable low latency. Some of the tradeoffs are the lack of traditional relational database functionalities, such as arbitrary joins and ad-hoc querying. It does provide linearizable, single-partition transactions via lightweight transactions (Paxos) today—and Raft-based strong consistency for metadata (and soon for user tables)—but it does not yet support full ACID transactions spanning multiple partitions or tables. ScyllaDB also requires careful data modeling to ensure data is distributed according to its access patterns. On the other hand, its performance and scalability make it ideal for workloads that require predictable performance at scale. Read about database performance tradeoffs.
ScyllaDB distributes data using a shard-per-core architecture. It attributes partitions to CPU virtual cores for efficient parallelism and contention reduction. By leveraging consistent hashing for data distribution, it ensures the load is evenly distributed across the cluster and simplifies scalability. It replicates data across nodes distributed across multiple availability zones, which provides fault tolerance and high availability on the cloud. Additionally, ScyllaDB tablets allow rapid scaling of clusters in response to traffic spikes and increasing demand. ScyllaDB Architecture Overview.
ScyllaDB is source available. You can access the full ScyllaDB Enterprise feature set for free, up to 10 TB of total storage and 50 vCPUs across all clusters. This includes community support only.
