July 2025 Summaries

Yieldmo, an online advertising platform, faced significant challenges with its initial reliance on Amazon DynamoDB due to escalating costs and the need for multicloud flexibility as the company expanded globally. The platform processes hundreds of billions of ad requests daily, requiring database lookups with single-digit millisecond latencies to maintain its ad-serving infrastructure. While DynamoDB offered simplicity and reliability, its costs proved unsustainable at scale, and geographic latency became an issue when attempting to deploy across multiple cloud providers. After considering alternatives, Yieldmo opted for ScyllaDB, which provided a DynamoDB-compatible API, cross-cloud support, and lower costs, enabling a smooth migration with minimal code changes. The transition involved moving multiple terabytes of data across regions and resulted in significant cost savings and modest latency improvements. A year after the migration, Yieldmo benefited from increased multicloud flexibility, reduced database costs by approximately half, and reliable performance, making ScyllaDB the primary database for its operations.

ScyllaDB Cloud has introduced a "Bring Your Own Account" (BYOA) model for Google Cloud, allowing users to run the fast and scalable NoSQL database within their own Google Cloud accounts. This setup ensures that users maintain full control over their data, which never leaves their private Google Cloud environment. The model is beneficial for industries with strict data privacy regulations, such as healthcare and finance, as it provides data sovereignty, compliance, and security features, including Bring Your Own Key (BYOK) encryption. Users can leverage existing Google Cloud commitments and discounts, as ScyllaDB Cloud costs are limited to license and support fees. The setup process involves using a wizard to configure necessary resources and permissions within a chosen Google Cloud project, with the option to select geographic regions, access types, and instance types according to anticipated traffic.

DynamoDB, while easy to adopt, presents significant challenges in cost management due to its complex pricing structure, which often catches teams off guard. The service charges based on Read and Write Capacity Units, with writes being notably more expensive, a factor exacerbated in on-demand mode where costs can escalate with write-heavy workloads. Provisioned capacity aims to offer cost savings but requires careful management of rates rather than total requests, leading to potential overprovisioning and idle capacity expenses. Reserved capacity offers further cost reductions but demands upfront commitments without partial payment options, complicating budget forecasts for teams. Additionally, data transfer costs are dictated by chunk sizes, with writes billed per 1KB and reads per 4KB, potentially leading to inefficiencies for variable-sized data. The necessity to overprovision to avoid throttling during peak loads further complicates cost management, as does the punitive rounding of item sizes, making DynamoDB's serverless promise feel paradoxically burdensome. Consequently, many teams remain unaware of their true costs until after significant financial impact, prompting the creation of calculators to better model real workloads and understand cost drivers.

ScyllaDB has partnered with Google Cloud to evaluate the performance of its database on Google Cloud's new Z3 small instances, which are designed for storage-optimized virtual machines to deliver low latency and high performance. The benchmarks demonstrated that ScyllaDB on Z3 instances achieved significant throughput improvements and reduced latency compared to the previous N2 instances, particularly under high load scenarios. The Z3 series, available in configurations ranging from 8 to 44 vCPUs, incorporates advanced CPU, memory, network, and SSD technologies, enabling ScyllaDB to cater to a broader range of workloads with enhanced performance. Testing involved various workloads such as read, write, and mixed operations, and the results highlighted substantial gains in throughput and latency across all scenarios. These findings underscore the potential of Z3 instances to offer greater flexibility and performance for ScyllaDB Cloud users, particularly in cost-sensitive environments where high storage IOPS and network bandwidth are required without high core counts. Future testing will continue to explore the Z3 instances' performance under different conditions to further refine ScyllaDB's deployment strategies.

The blog post explores the role of feature stores in real-time machine learning applications, highlighting their importance in managing features for both training and inference tasks. It emphasizes the distinction between offline stores, used for model training with large datasets, and online stores, crucial for real-time inference where low latency is essential. ScyllaDB is presented as a compelling choice for an online feature store due to its high-performance NoSQL capabilities, offering low latency, high throughput, and flexibility without vendor lock-in, making it suitable for scalable applications. The integration of ScyllaDB with the open-source feature store framework Feast is discussed, allowing users to leverage ScyllaDB's advantages while maintaining compatibility with existing data infrastructures. The post provides tutorials and resources for getting started with ScyllaDB and Feast to build real-time inference applications, addressing the need for fast and reliable feature serving in machine learning workflows.

ScyllaDB's Engineering Summit in Sofia, Bulgaria, brought together over 150 global participants, marking a significant event for the remote-first company. Held amidst the historical and cultural backdrop of Sofia, the summit featured a blend of insightful talks, collaborative sessions, and a hackathon that emphasized creativity, teamwork, and problem-solving. Attendees, including both engineering and Customer Experience teams, engaged in discussions about the company's past achievements and future innovations, while also working on real-world challenges through a hackathon that fostered collaboration across different teams. The event was rounded off with social activities, including a lively party and explorations of Sofia's cultural landmarks and the picturesque Rila Mountains, offering participants a chance to connect beyond work. The summit not only reinforced ScyllaDB's commitment to innovation and teamwork but also left attendees inspired and eager for future gatherings.

ScyllaDB employs a customized version of the Rust-based benchmarking tool, "latte," to simulate real-world production workloads that are specific to its high-performance distributed database needs. Originally designed for latency measurements in healthy database clusters, ScyllaDB's fork of latte includes unique features such as preferred datacenter and rack support, sine-wave rate limiting for workload prioritization, and real-time HDR histogram data streaming to avoid RAM leaks. The tool's flexibility allows the ScyllaDB testing team to define any schema and workload functions, using rune scripts as business logic. This enhanced version aims to address specific challenges like coordinated omission in latency measurements and saturated database impacts, enabling the testing of complex customer scenarios with controlled disruptions. By adopting latte, ScyllaDB can quickly reproduce customer workloads in test environments, which is crucial for troubleshooting issues like latency spikes and throughput drops.