November 2023 Summaries

Amazon's announcement of ElastiCache Serverless at re:Invent 2023 introduces a new level of simplicity to caching services, addressing several limitations of its non-serverless predecessor. The service supports native Memcache and Redis APIs, facilitating easier migrations from traditional setups, and offers multi-AZ deployment by default, eliminating the need for complex configurations for high availability. Although it removes the need for autoscaling groups and simplifies pricing with a pay-per-use model, the pricing structure starts at $90/month for up to 1GB and does not entirely achieve the ideal of scaling to zero. Despite these advancements, the service has areas for improvement, particularly in scaling speed during high demand spikes. The blog compares ElastiCache Serverless to Momento Cache, highlighting the latter's compliance with serverless principles by offering usage-based pricing without minimums and requiring no management or provisioning, positioning it as a competitive alternative.

Distributed locks are essential for maintaining order and fairness in high-demand distributed systems, such as online ticket sales, where multiple users may attempt to purchase the same item simultaneously. Traditional methods like Paxos, Raft, and Apache Zookeeper, while effective, can be complex to set up and manage. In response, Amazon's DynamoDB Lock Client offers a simplified solution, and Momento has developed a cost-effective alternative that integrates seamlessly with existing systems and reduces expenses by 50-90% depending on the workload. Momento's locking library emphasizes ease of integration and management, ensuring a smooth user experience by maintaining continuous locks during transactions through heartbeating, crucial for preventing issues like double bookings. While optimistic concurrency control might seem appealing, its limitations in high-traffic environments make pessimistic locking a more reliable choice, ensuring user trust and satisfaction. Momento's approach offers significant advantages such as cost-effectiveness, high performance, and reduced infrastructure management, making it an ideal choice for systems requiring robust distributed lock management.

A customer utilizing Node.js Lambda functions with Momento reported unsatisfactory cold start times, prompting an investigation to optimize performance. Initially, efforts to adjust memory configurations yielded no significant improvements. However, by employing the serverless-esbuild plugin, the team significantly reduced the Lambda package size from 1.5mb to approximately 260kb, resulting in a 90% reduction in cold start times to around 100ms. Inspired by this, Momento experimented with their production Lambda functions, achieving substantial performance gains by minimizing the bundle size through esbuild, minifying code, and externalizing source maps, ultimately reducing cold start times by 40%. Despite testing different architectures, x86 consistently outperformed arm_64 within the Lambda environment. The refined approach not only enhanced performance but also streamlined development processes, providing a template for future projects.

DynamoDB is praised for its reliability and ease of use, often becoming an integral part of teams' workflows without drawing much attention. This seamless integration, however, can lead to missed opportunities for optimization, particularly regarding cost and performance. The blog post highlights the potential for significant savings by understanding DynamoDB's billing modes: provisioned throughput and on-demand. While on-demand usage is typically 6.94 times more expensive than provisioned throughput, it can be more cost-effective for those unable to utilize 100% of their provisioned capacity, especially if they have irregular traffic patterns. The post also introduces the Infrequent Access (IA) storage class, which can offer a 60% discount on storage costs, making it a favorable option if the IOPS cost is less than 2.4 times the storage cost. By leveraging simple mathematical calculations and monitoring metrics like ConsumedWriteCapacityUnits and ProvisionedWriteCapacityUnits, users can optimize their DynamoDB usage and potentially reduce expenses.

On episode 2 of the Observe It podcast, Daniela Miao engages in a discussion with Austin Parker, Director of Open Source at Honeycomb and a member of the OpenTelemetry governance committee, about the challenges and advancements in the field of observability, particularly focusing on OpenTelemetry. Austin highlights the complexities in making observability tools more accessible and user-friendly for non-technical users and addresses the frequent changes and learning curve associated with OpenTelemetry. He stresses the importance of simplifying the process for application developers to write custom instrumentation and recognizes an organizational gap where developers may not prioritize metrics and traces. Austin envisions a future where developers can integrate observability into their workflows without needing to delve into the complexities of OpenTelemetry.

Rate-limiting is crucial for maintaining the health and quality of services by preventing resource overuse and potential system failures, particularly in distributed systems and multi-tenant applications. Stateless rate limiters, while simple, often fail in dynamic environments due to their inability to adapt to fluctuating demands, leading to inefficiencies. The blog post explores two approaches for creating a distributed rate limiter using Momento Cache. The first method, recommended for its accuracy and efficiency, uses Momento’s increment and updateTTL APIs to manage user transactions per minute, setting a time-to-live for requests. The second approach, inspired by traditional methods such as those used with Redis, relies on get and increment APIs but struggles with accuracy under high demand due to race conditions. Testing reveals that the first approach is superior in maintaining accuracy and managing latency in high contention scenarios, making it a robust solution for distributed rate-limiting.