Distributed Databases: An Overview

On-Device AI: Benefits, Use Cases, and Challenges

On-device AI runs AI inference directly on local hardware rather than remote cloud servers. Models are trained in the cloud then compressed using techniques like quantization, pruning, and knowledge distillation before deployment to devices. Key benefits include reduced latency, improved privacy, offline functionality, and lower cloud costs. Challenges include limited compute/storage resources, model size constraints, complex OTA update distribution, and power consumption. Common use cases span mobile apps, IoT/smart home devices, retail personalization, and industrial monitoring. Hybrid edge-plus-cloud architectures offer a practical middle ground. Recommended frameworks include TensorFlow Lite, PyTorch Mobile, ONNX Runtime, Core ML, and Android NNAPI.

Your AI Agents Are Stuck in Pilot. It’s a Data Problem, Not a Model Problem.

Couchbase announces general availability of its AI Data Plane for self-managed Couchbase Enterprise deployments, extending beyond its managed Capella service. The platform addresses why enterprise AI agents stall in pilot by providing a unified data infrastructure layer with persistent agent memory, governed data access via an MCP Server, and an Agent Catalog for tool/prompt/trace management. The release also includes Apache Iceberg federation, a new Rust SDK, Bluetooth peer-to-peer sync for edge environments, React Native 1.1 support, and expanded infrastructure-as-code capabilities. The core argument is that AI agent production failures stem from fragmented data stacks rather than model limitations, and consolidating memory, caching, vector search, and operational data into one governed layer removes the bottlenecks that prevent pilots from reaching production.

Unlock Massive Scale for AI and Traditional Workloads: Meet the Couchbase Secondary Index Shard Dealer

Couchbase Server 8.0+ introduces the Shard Dealer, an automated system for managing secondary index placement across distributed clusters. It solves the 'Goldilocks problem' of shard count by dynamically scaling from 10–40 shards on small clusters to 600–2,000 on large ones. A multi-pass algorithm handles placement decisions, prioritizing reuse of under-capacity shards before creating new ones. Critically, the system enforces workload isolation by keeping traditional indexes and vector indexes in separate shards, preventing AI workloads from impacting standard query performance. Replica alignment is also maintained to speed up failover and recovery.

Convert XML to JSON in C# With Json.NET and Couchbase

A step-by-step guide to converting XML to JSON in C# using two approaches: Newtonsoft Json.NET (via SerializeXmlNode) and System.Text.Json (with a custom helper). Covers key differences between XML and JSON, common edge cases like attribute prefix handling, repeated element array coercion, namespace stripping, and large file streaming. Also demonstrates storing the resulting JSON documents in Couchbase Server or Capella using the .NET SDK 3.x UpsertAsync method, including parallel bulk insertion with Task.WhenAll.

Accelerating AI in Healthcare: Fix Data Infrastructure Before AI Fails Become a Board Priority

Healthcare AI initiatives are failing at high rates not because of model quality but due to fragmented, legacy data infrastructure. With 85% of failures attributed to poor data quality or infrastructure, the core issue is that most healthcare systems were built for batch ETL and relational databases, not real-time AI workloads. FHIR (Fast Healthcare Interoperability Resources) has become both a regulatory mandate and a technical foundation for modern healthcare platforms, but implementing it at scale is complex. Couchbase positions its distributed NoSQL platform as a solution by natively storing FHIR JSON resources, supporting full-text search on nested documents, enabling multi-region deployments, and consolidating operational, analytics, and AI vector workloads on a single platform while maintaining HIPAA and GDPR compliance. The post also highlights the CMS ACCESS Model as a new reimbursement framework that makes FHIR API infrastructure financially critical for healthcare organizations.

What Is DiskANN? Billion-Scale Vector Search Explained

DiskANN is a graph-based approximate nearest neighbor (ANN) algorithm from Microsoft Research (NeurIPS 2019) that enables billion-scale vector search by storing indexes on SSD rather than RAM. Built on the Vamana directed graph algorithm combined with product quantization (PQ), it keeps only compressed vectors in RAM for fast routing while reading full-precision vectors from SSD for final reranking. This achieves 95%+ recall with sub-5ms latency on 1B vectors using just 64GB RAM — 5-10x more vectors per machine than DRAM-only solutions like HNSW. The post covers how DiskANN works, compares it to HNSW and IVF, details tuning parameters (MaxDegree, SearchListSize, BeamWidthRatio), hardware sizing guidelines, and lists databases that support it including Couchbase 8.0, Azure Cosmos DB, Milvus, and pgvectorscale. FreshDiskANN extends the algorithm to support real-time updates without full index rebuilds.

Need a different partition key in Azure Cosmos DB? Pick the right approach

Khi partition key trong Azure Cosmos DB gây ra truy vấn chéo partition hoặc hot partition, bạn có bốn lựa chọn: sử dụng tính năng "Change partition key" trên portal (ít nỗ lực nhất), sao chép container qua CLI (có thể online/offline), hoặc di cư tùy chỉnh bằng bulk ingestion, Azure Data Factory/Spark cùng Change Feed (tùy chỉnh cao nhất). Nếu chỉ vấn đề nằm ở read pattern, Global Secondary Index (GSI) sẽ tạo container read-only, tự động đồng bộ với partition key mới mà không ảnh hưởng đến nguồn. Bài viết cũng phân tích điều kiện tiên quyết, đánh đổi (chi phí RU, TTL, tính nhất quán) và cung cấp bảng quyết định để chọn phương pháp phù hợp.

Lập trình viên nên đọc bài này để giải quyết hiệu quả vấn đề phân vùng không hiệu quả trong Cosmos DB, từ các giải pháp đơn giản đến phức tạp, giúp tối ưu hóa chi phí và hiệu suất cho ứng dụng của mình.