ScyllaDB Feature Trails

This post details how Medium evolved its feature store to handle real-time recommendations at 1M OPS by migrating from a relational features data model to a list-based approach using ScyllaDB. This change was driven by the need for efficient storage and retrieval of time-series interaction data (e.g., stories read, recommendations shown) to maintain relevance for users. The new model leverages ScyllaDB's capabilities to store lists of relation IDs with timestamps, enabling faster querying and better performance for recommendation systems. The post highlights the challenges of scaling with relational models and the benefits of ScyllaDB for high-throughput, low-latency data access in such scenarios.

This post introduces an asymmetric io_uring backend for Seastar, designed to offload I/O operations from application cores to dedicated I/O cores. This aims to reduce contention on application cores, minimize context switching overhead, and mitigate cache pollution by separating computation and I/O handling. The new backend contrasts with the existing symmetric io_uring and linux-aio backends, focusing on improving performance and latency predictability for compute-heavy workloads.

This post introduces ScyllaDB Agent Skills, a set of best practices and usage patterns for AI agents interacting with ScyllaDB Cloud clusters. These skills are delivered as markdown files and aim to improve AI agent reliability and performance by providing domain-specific knowledge. Key areas covered include scylladb-cloud-setup for proper connection flows and driver selection, scylladb-data-modeling for designing efficient NoSQL schemas and queries, and scylladb-vector-search for optimizing vector index creation and queries. This builds upon the existing integration of ScyllaDB with AI agents by providing structured guidance for their development and deployment.

ScyllaDB successfully mitigated performance-degrading connection storms by optimizing caching, throttling, and password hashing to achieve a 1000x reduction in tail latency. This involved improvements across the connection establishment pipeline, including connection throttling, caching of connection-related data, optimization of password hashing, and the introduction of a dedicated service level for background traffic.

This post details the development of cqlsh-rs, a ground-up rewrite of the Python cqlsh in Rust. The primary motivations were to address the packaging nightmare associated with Python dependencies and to improve the reliability and performance of the COPY TO/FROM feature. The rewrite aims to provide a static binary with zero runtime dependencies, simplifying deployment and maintenance. It also focuses on resolving long-standing bugs and performance issues in the bulk data import/export functionality.