September 2025 Summaries

P99 CONF 2025 is a virtual, interactive, and purely technical community event focused on database performance and distributed systems, featuring talks from industry leaders such as Pinterest, Clickhouse, Meta, and ScyllaDB. The conference covers a wide range of topics, including database internals, Rust and C++ integration, durable workflows, CPU scheduling, and vector search optimization. Highlights include discussions on rewriting SQLite in Rust, optimizing latency in TigerBeetle, and reimagining sorted sets with B+ trees. The event aims to provide insights into the engineering challenges and solutions in database technologies, with a keynote by Andy Pavlo promising additional valuable content.

In the final installment of a three-part series, ScyllaDB co-founder and CTO Avi Kivity discusses the architectural changes in ScyllaDB X Cloud aimed at improving efficiency for storage-bound workloads and enabling deployment on mixed-size clusters. These changes focus on optimizing storage utilization and compression techniques, particularly dictionary-based compression, which enhances data compression by matching repetitive byte sequences against samples. The discussion also highlights the challenges and strategies associated with managing clusters of varying sizes to maintain an optimal CPU-to-storage ratio, especially during fluctuating workload demands. Avi reflects on the evolution of ScyllaDB, acknowledging the complexity and unforeseen challenges in developing database architecture while emphasizing the importance of completing ongoing projects, such as the tablets initiative, before embarking on new developments.

The text discusses various caching strategies, focusing on their implications for latency and complexity, and provides a detailed examination of cache-aside, read-through, write-through, write-behind, client-side, and distributed caching strategies. Cache-aside caching is highlighted as a common approach where the application manages cache misses by fetching data from the backing store, while read-through caching involves the cache actively retrieving data from the backing store during a cache miss. Write-through caching ensures immediate synchronization between the cache and the backing store, albeit with higher latency, whereas write-behind caching defers database updates, allowing for lower write latency but at the cost of transactional guarantees. Client-side caching offers low latency by keeping cache within the application, but can increase memory usage and complicate transaction support. Distributed caching, involving multiple cache instances, addresses geographic latency and scalability but introduces challenges similar to those found in data replication and partitioning. The text emphasizes the trade-offs each strategy presents in terms of system performance and complexity.

The blog post features a conversation with ScyllaDB Co-Founder and CTO Avi Kivity, focusing on the technical intricacies and innovations behind ScyllaDB X Cloud's recent release. A key topic is the transition from vNodes to a tablets-based data distribution, which improves storage utilization and performance by allowing 90% storage usage and mitigating issues like hot partitions. The introduction of a system.tablets indirection table provides flexibility in data distribution, allowing different tables to be managed independently based on their size and activity levels. This flexibility aids in autoscaling, enabling the system to scale at 90% storage utilization, reducing costs and enhancing efficiency. The post also highlights how ScyllaDB distinguishes between foreground and background CPU tasks to optimize resource allocation and autoscaling, ensuring that scaling occurs only when necessary for real-time query performance.

Latency is a critical concern at ScyllaDB, prompting the exploration of caching as a method to reduce latency in applications and systems. Pekka Enberg's book "Latency" addresses the complexities of latency by revealing techniques often hidden in developer folklore and explores practical applications of these techniques, including caching. Caching is recommended for systems that do not require complex queries or transactions and face compute or storage constraints, as it can significantly reduce access time by storing temporary data copies. However, it involves trade-offs such as data freshness and storage needs. Effective caching is measured by the cache hit-to-miss ratio, with a high hit ratio indicating better performance. Cache eviction policies are essential for managing limited storage space, determining which data remains in the cache, and ensuring data consistency when source data changes. Caching strategies vary in how they handle cache misses, influencing latency and complexity. Despite its challenges, caching remains a valuable technique for latency reduction, particularly when system changes are not feasible.

ScyllaDB X Cloud represents a significant architectural evolution aimed at enhancing scalability and price performance, particularly for dynamic and storage-bound workloads, through the use of Raft and tablets-based data distribution. This shift, driven by the need to improve the scalability of large clusters and the inefficiencies of the previous static data distribution model, allows for faster and more flexible data management. The introduction of a coordinating authority via Raft centralizes topology management, while the tablets abstraction enables dynamic data rebalancing and efficient node scaling. This new architecture addresses the limitations of earlier approaches, such as the slow streaming mechanism and the inability to perform topology changes involving multiple nodes, thus reducing wasted resources and improving overall system efficiency.

Monster Scale Summit is a virtual technical conference designed to connect engineers, architects, and site reliability engineers (SREs) from around the world who work on performance-sensitive, data-intensive applications. Scheduled for March 11 and 12, 2026, the summit focuses on extreme scale engineering challenges, distinguishing itself from similar events like P99 CONF by emphasizing issues related to massive throughput, data, and global distribution. The event encourages speakers to share their expertise on topics such as distributed databases, real-time processing, and innovative system designs, with sessions lasting 18-20 minutes to emulate a TED Talk format. By offering a platform for first-hand experiences and insights, the conference aims to explore solutions for balancing latency, concurrency, and throughput in demanding environments.

ScyllaDB has evolved from being perceived as merely a faster alternative to Apache Cassandra into a database with its own distinctive features and identity, surpassing Cassandra in many aspects. Initially focused on optimizing performance through innovations like a shard-per-core asynchronous architecture and advanced schedulers, ScyllaDB now boasts features such as Materialized Views, Global and Local Secondary Indexes, and a superior Paxos implementation for lightweight transactions. Over successive generations, it has emphasized cloud integration, introducing its DynamoDB-compatible API, and enhancing data management with Incremental Compaction Strategy, Change Data Capture, and a raft of unique capabilities like workload prioritization, repair-based operations, and high storage utilization. ScyllaDB's recent advancements include strong consistency with Raft and Tablets for data distribution, allowing for rapid scalability and efficient resource use, all while maintaining low latencies. As it continues to innovate, ScyllaDB is poised to offer even more robust solutions, including vector search, strongly consistent tables, and dynamic provisioning, potentially making it a compelling choice for projects requiring highly predictable performance and scalability.