DeepSeek-R1: Reasoning via Reinforcement Learning

Spanner's Globally-Distributed Database: Query Execution

This paper details the evolution of Google's Spanner, a globally-distributed database system, from a key-value store to a fully-fledged SQL system. Key improvements discussed include distributed query execution, handling of transient failures via query restarts, efficient range extraction for data retrieval, and the adoption of a common SQL dialect. The authors also explain the transition from a Bigtable-like storage format to a more efficient blockwise-columnar store (Ressi). Finally, the paper highlights lessons learned during Spanner's large-scale deployment and outlines remaining challenges.

Change Data Capture (CDC): Three Implementation Methods

The article explores Change Data Capture (CDC), a method for tracking database changes, highlighting its advantages over traditional daily snapshots. It details three CDC implementation approaches: using database triggers (e.g., in PostgreSQL), capturing API requests and using a message broker (e.g., Kafka), and leveraging change streams within a data warehouse (e.g., Snowflake). The article compares these methods, weighing their pros and cons in terms of performance, scalability, and ease of implementation. A subsequent discussion critiques the presented methods, suggesting alternative, more robust solutions based on logical replication tools like Debezium.

DeepSeek-R1: Reasoning via Reinforcement Learning

This research paper introduces DeepSeek-R1, a large language model enhanced for reasoning capabilities using reinforcement learning (RL). Two versions are presented: DeepSeek-R1-Zero, trained purely via RL without supervised fine-tuning, and DeepSeek-R1, which incorporates additional multi-stage training and cold-start data for improved readability and performance. DeepSeek-R1 achieves results comparable to OpenAI's o1-1217 on various reasoning benchmarks. The study also explores distilling DeepSeek-R1's reasoning capabilities into smaller, more efficient models, achieving state-of-the-art results. Finally, the paper discusses unsuccessful attempts using process reward models and Monte Carlo Tree Search, providing valuable insights for future research. https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf

Jira Cloud Performance Enhancement with Protobuf

This Atlassian blog post details the migration of Jira Cloud's Issue Service from JSON to Protocol Buffers (Protobuf) to enhance performance. The switch involved a phased approach to minimise downtime, creating new endpoints and logic to handle both formats concurrently before a complete transition. The results showcased significant improvements: 75% less Memcached CPU usage, 80% smaller data size, and a substantially faster response time. Challenges encountered included Protobuf's handling of null values and incompatibility with Spring's default error controller, which required workarounds. Ultimately, the migration yielded substantial performance gains and reduced infrastructure needs. https://www.atlassian.com/blog/atlassian-engineering/using-protobuf-to-make-jira-cloud-faster

Hyaline: Fast and Transparent Lock-Free Memory Reclamation

This research paper introduces Hyaline, a novel family of memory reclamation schemes for lock-free data structures in unmanaged C/C++ code. Hyaline leverages reference counting, but only during reclamation, minimising overhead during object access and balancing workload across threads. The paper details Hyaline's design, including a scalable multi-list version and robust extensions to handle stalled threads. Extensive testing across multiple architectures demonstrates Hyaline's superior performance and memory efficiency compared to existing schemes like epoch-based reclamation and hazard pointers, particularly in read-dominated and oversubscribed scenarios. The paper concludes by proving Hyaline's correctness and lock-freedom properties.