A weekly roundup of the architecture insights and releases we’ve been reading.
Infrastructure
🔹 FSZ: Breaking the Prediction-Throughput Trade-off in GPU Lossy Compression
A new approach to lossy GPU compression increases throughput while preserving prediction quality. The work shows how compression can become a tool for optimizing compute infrastructure, rather than simply a way to reduce memory usage. Read the release (EN)
🔹 Robust KV Cache Management for LLM Serving under Output Token Length Uncertainty
KV cache management becomes particularly challenging when the length of an LLM response is unknown in advance. The proposed mechanism accounts for this uncertainty when planning resources, helping improve serving efficiency and reduce latency. Read the release (EN)
🔹 Hardware-Transparent I/O Governance in Disaggregated Heterogeneous Storage
A new I/O governance model for disaggregated heterogeneous storage separates resource allocation policies from the underlying hardware. This approach simplifies infrastructure management and enables more efficient use of different storage resources. Read the release (EN)
Architecture
🔹 DoorDash Uses Envoy and Valkey for a 1.5M RPS Proxy Cache with 99.99999% Availability
DoorDash built a proxy cache based on Envoy and Valkey, designed to handle up to 1.5 million requests per second. The architecture demonstrates how combining proxy and cache layers can deliver both extreme throughput and 99.99999% availability. Read the release (EN)
🔹 Multi-Cluster Databases on Kubernetes: Architecture and Deployment
Multi-cluster databases on Kubernetes are emerging as an architectural pattern for systems that require both scalability and high availability. The article explores key approaches to designing and deploying these configurations. Read the release (EN)
Developer Tools
🔹 How to Scale GitOps in the Enterprise: From Single Cluster to Fleet Management
Moving from a single Kubernetes cluster to managing an entire fleet requires a new level of GitOps practices. The article explores how to scale automation, configuration management, and deployment workflows across large infrastructure environments. Read the release (EN)
🔹 Wait-Free Locks Should Not Fear Later Arrivals
The paper examines wait-free synchronization in concurrent systems and the challenge posed by threads arriving after an operation has already started. The proposed approach aims to improve the predictability and efficiency of concurrent algorithms without relying on blocking. Read the release (EN)
AI and Machine Learning
🔹 Ascend to Science: Exploration of AI Chips for Scientific Computing
AI chips are increasingly moving beyond traditional machine learning workloads. The paper explores their use in scientific computing and examines which architectural features could make specialized accelerators effective tools for HPC workloads. Read the release (EN)
🔹 OpenEvoShield: Dual Non-Stationary Continual Defense for Open-World Multi-Agent System Attacks
OpenEvoShield proposes an adaptive approach to defending multi-agent systems against attacks in dynamic environments. The architecture accounts for the fact that threat patterns can change over time, requiring the defense mechanism itself to continuously adapt. Read the release (EN)