The article introduces Alibaba's Fast Attribution Engine (FAE), describing the technical challenges of user behavior and advertising attribution, its data model (AFile), system architecture, performance advantages over traditional OLAP solutions, and a range of application scenarios such as frequency analysis, crowd flow modeling, path, retention, funnel analysis, and visitor selection.
The article details the background, architecture, core features, scheduling mechanisms, Lisp‑based query DSL, and Alluxio integration of Vipshop's self‑developed Hera data service, illustrating how it unifies multi‑engine data access, improves SLA, and accelerates large‑scale crowd computing tasks.
This article explains the growing analytical demands in the big‑data era, the limitations of traditional OLAP, and how Kyligence’s distributed OLAP engine addresses data‑island issues, multi‑dimensional and many‑to‑many analysis, unified security, and performance optimization with MDX on Spark, delivering a seamless Excel‑like experience.
This article introduces the third edition of Stanford’s Data Mining Tutorial, highlighting its panoramic roadmap of data‑mining techniques for massive datasets, core features, comprehensive topic coverage, target audience, and supplementary resources while noting its popularity among students and professionals.
This article explains the evolution of cloud computing, the distinction between serial, parallel, and distributed computing models, and details the four classic computer architecture classifications (SISD, SIMD, MISD, MIMD) along with shared‑memory and distributed‑memory MIMD systems and their role in modern distributed system layers.
This article presents a detailed overview and ranking of numerous machine learning tools and libraries, distinguishing deep and shallow learning approaches, highlighting language support, GPU acceleration, and distributed computing capabilities, and provides insights into their academic and industrial usage.
This article systematically introduces NetEase Kyuubi, an open‑source high‑performance JDBC and SQL execution engine built on Apache Spark, covering its background, core architecture, service discovery, session and operation management, startup processes, and key source‑code implementations with detailed code examples.
Jet, an open‑source in‑memory distributed batch and stream processing engine, introduces dynamic ClassDefinition for SQL, JDBC sink batch size option, SQL aggregation memory optimizations, improved sliding‑window pickAny performance, numerous extension updates, critical bug fixes, and a major API change to DAG.toDotString for enhanced parallelism control.
This article details the design of a red packet calculation and distribution system for a Spring Festival promotion, addressing mixed-type packet computation, seamless transition to awarding, distributed task processing, stability guarantees, and successful implementation results.
This article explains SparkSQL's overall workflow, introduces the basic elements of joins, and details the physical execution processes for various join types—including sort‑merge, broadcast, and hash joins—along with their implementation conditions and optimization considerations.
This article traces the history of big‑data computing engines—from early MapReduce and Hadoop through Spark, Storm, Flink, and the newer Ray—explaining their technical advances, real‑world applications in AI and finance, and why graduates should consider a career in this rapidly evolving field.
This article explains the fundamentals of parallel computing in the big‑data era, compares parallelism and concurrency, outlines GPU and distributed‑computing solutions, and provides a detailed guide to Python’s multiprocessing module with code examples, performance tests, and practical tips.
This article provides a comprehensive overview of Hadoop, covering its origins, core components such as HDFS, MapReduce, and YARN, their architectures, data storage and processing mechanisms, fault‑tolerance features, scheduling strategies, and practical optimization techniques for large‑scale distributed computing.
The article explains how JD Cloud’s federated learning platform enables secure, privacy‑preserving collaborative AI across isolated data sources by using encrypted distributed training, flexible model architectures, and a range of algorithms, while highlighting its architecture, security mechanisms, deployment speed, and real‑world industry successes.
This article explains Apache Kylin’s core principles and technical architecture, then details how major Chinese companies such as Baidu Maps, Lianjia, and Didi have deployed Kylin for large‑scale OLAP, describing their system designs, performance results, and the challenges they encountered.
This article explains what data skew is, why it occurs in large‑scale Hadoop and Spark jobs, illustrates typical scenarios that cause it, and provides practical strategies and platform‑specific optimizations to detect, mitigate, and prevent skew in big‑data processing pipelines.
This article explains how Spark can be combined with MongoDB for large‑scale data processing, covering Spark fundamentals, comparisons with HDFS, practical integration patterns, performance benefits, real‑world case studies, and detailed code examples for deployment and analytics.
Meituan’s autonomous vehicle engine abstracts communication, scheduling, data handling, and tooling into three layers to ensure deterministic behavior across on‑vehicle and simulation environments, tackling consistency, scheduling, and resource‑utilization challenges by using unified computation models, distributed graph deployment, caching, and remote model serving, thereby accelerating autonomous delivery vehicle development.
Apache Spark Core is a high‑performance, fault‑tolerant engine that abstracts distributed computation through SparkContext, DAG and Task schedulers, supports in‑memory and disk storage, runs on various cluster managers (YARN, Kubernetes, etc.), and unifies batch, streaming, ML and graph processing via its rich ecosystem.
This article introduces federated learning, edge computing, and cloud computing, explains each technology's principles and use cases, and then compares their similarities and differences, highlighting privacy‑preserving collaborative modeling, near‑source processing, and centralized resource provisioning.
By parallelizing Spark’s driver‑side commit, trash, and move phases—previously single‑threaded operations that caused costly copy‑on‑rename when writing massive files to object storage—the Tencent Cloud EMR case achieved over a tenfold (1,100 %) speedup, making object storage a viable alternative to HDFS.
This article explains how Spark SQL processes joins—including inner, outer, semi, and anti joins—by describing the overall query planning flow, the three physical join strategies (sort‑merge, broadcast, and hash), and the specific implementation details for each join type.
This article explains the various join types supported by Spark SQL, describes the overall Spark SQL execution flow, and details the physical implementation processes of inner, outer, semi, anti, broadcast, sort‑merge, and hash joins, helping developers grasp how joins are executed in a distributed environment.
DLFlow, an end‑to‑end framework from Didi’s user‑profile team, merges Spark and TensorFlow to automate feature preprocessing, large‑scale distributed training, and massive prediction for big‑data offline tasks, offering configuration‑driven pipelines, task scheduling, and easy deployment that dramatically speeds model development.
The article outlines Presto’s high‑performance, coordinator‑worker architecture and query flow, describes YouZan’s migration from mixed Hadoop deployment to dedicated low‑latency clusters, details challenges such as small‑file handling and regex backtracking with their fixes, and previews future enhancements like Alluxio integration, session property managers, and Ranger‑based multi‑tenant isolation.
The article introduces LPA-Detector, an open‑source project that redesigns the Label Propagation Algorithm using Spark GraphX to add node confidence weights and relationship influence, achieving significant improvements in execution efficiency and detection accuracy for massive graph data in risk‑control scenarios.
This article provides a comprehensive overview of Apache Flink, explaining how it unifies stream and batch processing on a single runtime, detailing its key features, APIs, libraries, architectural components, fault‑tolerance mechanisms, scheduling, iterative processing, and back‑pressure monitoring.
This article explains edge computing, its distributed architecture, key advantages such as higher security, lower latency and reduced bandwidth costs, explores major application scenarios like smart manufacturing and autonomous driving, and analyzes how 5G both drives and benefits from edge computing development.
Alibaba’s DAMO Academy and Zhejiang CDC have built an AI‑powered, fully automated whole‑genome sequencing platform that reduces COVID‑19 sample analysis from several hours to about half an hour, delivering rapid, accurate detection of viral mutations and supporting vaccine and drug research.
This article explains PySpark's multi‑process architecture, how the Python driver uses Py4J to call Java/Scala APIs, the implementation of RDD and DataFrame interfaces, executor‑side process communication and serialization with Arrow, and the design of Pandas UDFs, while also discussing current limitations and future directions.
The interview details how Ant Financial transitioned from offline to online machine learning by adopting the Ray distributed engine, describing their open architecture, fusion computing approach, technical advantages, encountered pitfalls, and plans to open‑source the system for broader AI and big‑data use.
This article explains the fundamentals of parallel and distributed computing, their definitions, core principles, advantages, required conditions, and key differences, highlighting how each approach tackles large‑scale tasks within high‑performance computing environments.
Edge computing moves data collection and analysis closer to the source, reducing latency, bandwidth costs, and privacy risks, and is increasingly essential for IoT, industrial automation, autonomous vehicles, smart cities, and other latency‑sensitive applications.
This article walks through the SparkEnv initialization process in Apache Spark, detailing how the driver and executor environments are created, the key components such as SecurityManager, RpcEnv, SerializerManager, BroadcastManager, MapOutputTracker, ShuffleManager, MemoryManager, BlockManager, MetricsSystem, and OutputCommitCoordinator are instantiated, and how the final SparkEnv instance is assembled and stored.
Tencent’s iData analysis center selected Spark as its new computing platform because, unlike ElasticSearch, TiDB, and other MPP solutions, Spark offers iterative processing, shuffle support, robust SQL and DAG scheduling, and flexible SMP‑style data exchange, enabling efficient OLAP on billions of game‑user records.
This article explains why YARN was created to overcome the limitations of MapReduce 1.x, describes its architecture—including ResourceManager, NodeManager, ApplicationMaster, Container, and Client—and outlines the step‑by‑step execution flow that enables multiple computation frameworks to run on Hadoop.
The article explains Hadoop's MapReduce framework as both a programming model and execution engine, detailing its map and reduce phases, the WordCount example code, job startup components, data shuffling, partitioning, and how large‑scale distributed computations are orchestrated across a cluster.
This article explains the concept, usage, precautions, and a practical example of broadcast variables in Apache Flink, illustrating how to initialize, broadcast, retrieve, and apply shared data across parallel operators with Java code snippets.
This article examines the historical evolution of Spark's shuffle implementations—from early Hash‑Based Shuffle to modern SortShuffleWriter, BypassMergeSortShuffleWriter, and UnsafeShuffleWriter—explaining their design choices, selection criteria, and the corresponding shuffle reader architecture in a production‑grade Spark 2.1.1 environment.
Euler, Alibaba's newly open‑sourced graph deep‑learning framework, combines distributed graph processing with neural network training to handle billions of nodes and edges, supports heterogeneous graphs, offers built‑in algorithms, and has already boosted advertising, fraud detection, and other industry applications.
AI infrastructureEuler frameworkGraph Neural Networks
Alibaba’s open‑source Mars engine extends NumPy‑style tensor operations to distributed environments, leveraging GPU acceleration, sparse matrices, and flexible scheduling to dramatically boost scientific and AI workloads beyond single‑machine limits.
This article explains how Spark's memory management models and configuration parameters can be tuned to handle massive billing data efficiently, covering StaticMemoryManager vs UnifiedMemoryManager, storage and shuffle memory fractions, common OOM and file‑not‑found issues, and practical performance‑optimisation tips.
This article provides a concise overview of numerous Apache open‑source projects—including Ignite, MapReduce, Pig, JAQL, Spark, Storm, Flink, Apex, REEF, Twill, and Beam—that enable distributed in‑memory storage, real‑time and batch processing, and advanced analytics for large‑scale data workloads.
This article provides an in‑depth technical overview of Alibaba's AnalyticDB, covering the challenges of massive real‑time analytics, the cloud‑native multi‑tenant architecture, data model, import/export capabilities, high‑performance SQL parser, the Xuanwu storage engine, Xihe compute engine, optimizer, GPU acceleration, and elastic scaling features.
The article reviews recent big-data trends—from Hadoop’s evolution and Spark’s in-memory advances to emerging storage like Ozone—while detailing data-warehouse models, query-optimizer techniques, and cloud-native architectures that integrate diverse data sources, enabling scalable, AI-ready analytics and modern data-lake capabilities.
An interviewee faces the challenge of determining whether a newly given integer exists within a set of 4 billion numbers, and the article explores efficient solutions—from naive disk‑I/O approaches to distributed processing and the memory‑saving bitmap technique—highlighting their performance trade‑offs and implementation details.
This article explains the concepts, principles, and key distinctions between parallel computing and distributed computing, describing their objectives, basic conditions, advantages, and typical use cases within high‑performance computing, and highlights how they differ from grid and cloud computing.
An interview question about checking whether a new integer belongs to a set of 4 billion numbers leads to a discussion of distributed loading across eight machines, bitmap representation using 500 MB of memory, and interval‑based external sorting, illustrating practical big‑data algorithm design.
The article summarizes senior AI expert Wu Jianjun’s presentation on machine‑learning and deep‑learning engineering at Ping An Life, detailing the company’s big‑data platform, data processing pipelines, model training frameworks, distributed computing strategies, and production model‑serving architecture for financial applications.
This article explains the storage challenges of big data, introduces RAID levels and their trade‑offs, describes the HDFS architecture with NameNode and DataNode replication, details the MapReduce programming model and execution flow, and shows how Hive translates SQL queries into MapReduce jobs.
This article introduces Hadoop’s open‑source big‑data framework, explains its core components HDFS and MapReduce, and outlines four key advantages—ease of deployment, robustness, scalability, and simplicity—while also covering HBase as the Hadoop‑based column‑oriented database.
This article reviews Alibaba's MaxCompute evolution from ODPS to a unified, multi‑cluster big‑data platform, detailing its architecture, development tools, large‑scale deployments, performance optimizations, typical workload scenarios, and why it is the preferred choice for enterprise data processing.
This article uses a human‑computer analogy and a playing‑card counting example to explain the fundamentals of distributed computing, why single machines cannot handle massive data, and how the MapReduce model’s four steps—split, transform, shuffle, and merge—solve big‑data problems.
In this interview, Baidu’s senior big‑data architect Ma Ruyue discusses his career transition from Hadoop to online databases, the design philosophy behind Baidu’s Palo ROLAP system, the future of HTAP, and his views on the evolving big‑data ecosystem including Spark, AI, and containerization.
This article explains Hadoop's core components, the MapReduce programming model, the detailed shuffle and merge processes, and how YARN replaces the classic JobTracker/TaskTracker architecture to improve scalability and resource utilization in large‑scale data processing clusters.
The article explains how Ethereum aims to replace centralized cloud services with a decentralized network of nodes, giving users control over their data and creating a global, democratic computing platform, while also noting the challenges of security, usefulness, and scalability.
The paper describes a high‑performance user‑behavior analysis system that processes hundreds of billions of daily logs for Meituan‑Dianping, using an inverted‑index structure with bitmap UUID sets and timestamp sequences, combined with Spark, Spring and Alluxio optimizations to cut query times from hours to under five seconds.
This article presents a first‑principles framework that categorizes blockchain infrastructure components—storage, computation, and communication—by mapping them to concrete projects such as Ethereum, IPFS, BigchainDB, and others, illustrating how these modules interoperate to build efficient decentralized applications.
SparkSql optimizes inner‑join queries by pushing predicates to the scan phase, allowing filters connected with AND to be applied before the join without changing results, while OR‑connected filters can be unsafe except when they involve the join key or partitioned tables which use partition pruning.
The article explains how to combine Distributed TensorFlow with Kubernetes—using GlusterFS storage, Deployments for parameter servers, Jobs for workers, service discovery, monitoring, and a Jinja2‑generated YAML template—to create isolated, scalable training clusters with Jupyter and TensorBoard access.
This article details how Suning built a Hadoop‑based big data platform and leveraged Apache Spark to process terabytes of product price and inventory data, describing the system architecture, four key technical practices, performance results, and future data‑lake directions.
This tutorial walks you through setting up a three‑node Apache Storm cluster on CentOS 6.9, covering hostname configuration, firewall disabling, Zookeeper preparation, Storm installation, startup of Nimbus, UI and supervisors, and finally submitting a sample topology.
The article explains MapReduce’s divide‑and‑conquer model and Hadoop’s execution pipeline—including map, partition, spill, merge, shuffle, and reduce phases—illustrated with a WordCount example that shows how mappers emit word‑1 pairs and reducers aggregate counts to produce final frequencies on HDFS.
This article introduces Hadoop as a widely used big‑data framework, explains its core components HDFS and MapReduce, describes the cluster node roles, presents typical command‑line usage and a sample MapReduce workflow, and offers guidance for further learning.
This tutorial quickly introduces the MapReduce model, explains its core principles and execution flow, and guides you through seven practical examples—from basic WordCount to custom serialization, partitioning, joins, and friend‑recommendation—while providing test data and an optional ready‑made Hadoop environment for hands‑on practice.
This article walks through Spark’s origins, its core RDD concept, how it improves on Hadoop’s MapReduce, the role of in‑memory processing, functional programming support, and the emergence of Spark SQL with DataFrames and the Catalyst optimizer.
This article provides a comprehensive overview of Hadoop as the leading open‑source platform for big‑data processing, detailing its core components HDFS and MapReduce, the evolution to Hadoop 2.0/YARN, and the extensive ecosystem of tools and commercial solutions that enable scalable storage, analysis, and machine‑learning on massive data sets.
TensorFlow is a dataflow‑based programming model for large‑scale machine learning that uses directed acyclic graphs to represent computations, supports single‑device, multi‑device, and distributed execution with sophisticated node placement, communication, fault‑tolerance, and optimization techniques, and provides tools such as TensorBoard for visualization.
This article explains Hadoop as an Apache‑managed open‑source platform for storing massive data on distributed clusters and running robust, efficient analytics via its two core components—HDFS for storage and the Java‑based MapReduce framework for processing—highlighting modularity, high availability, and common tooling.
Apache Beam is an open‑source, unified programming model for distributed data processing that lets developers write pipelines once and run them on multiple execution engines such as Spark, Flink, or Dataflow, simplifying code reuse and easing migration between frameworks.
This article chronicles Hadoop’s ten‑year evolution from its early HDFS and MapReduce roots to a mature big‑data platform, detailing its historical milestones, architectural layers, ecosystem components, industry adoption, and future trends in storage, processing, security, and cloud integration.
Data warehouses centralize and transform enterprise data for multidimensional analysis, and modern demands have spawned four types—traditional, real‑time, associative discovery, and data marts—each with distinct technical requirements, while Hadoop‑based solutions like Transwarp Data Hub address challenges of scale, variety, latency, and security.
Meituan’s delivery team leverages massive real‑time data mining, distributed parallel optimization and a simulation platform to intelligently assign orders, boost rider efficiency, cut costs, and continuously refine algorithms, while integrating offline‑online learning and upstream‑downstream coordination to enhance overall logistics performance.
AI in deliverydistributed computinglogistics optimization
The article reviews Hadoop’s origins from Google’s pioneering papers, explains its architecture and ecosystem, evaluates its strengths such as scalability and benchmarks, discusses current limitations like single‑point failures and complex programming, and outlines upcoming improvements including HDFS Federation and next‑generation MapReduce.
The article details the author’s experience upgrading a production Spark cluster from version 1.4.1 to 1.6.1, exposing memory‑spill, unified memory, BlockManager deadlock, Yarn‑kill, UI quirks, and Spark‑SQL compatibility issues, and proposes concrete code‑level fixes for each problem.
This article explains the end‑to‑end workflow of Spark on YARN, covering client initialization, ApplicationMaster actions, driver and executor roles, RDD fundamentals, SparkSQL processing, and practical code examples for building and tuning distributed Spark jobs.
This article introduces Apache Spark’s core architecture, explains how Spark runs on YARN, details driver and executor roles, describes RDD concepts and dependencies, and outlines SparkSQL’s schema‑based query processing, providing code examples for HiveContext and JDBC integration.
This article describes how Meituan migrated from Hive‑SQL and MapReduce to Spark on YARN, built an interactive Zeppelin‑based development platform, created reusable ETL templates, constructed a Spark‑driven feature and data‑mining platform, and applied Spark to interactive user‑behavior analysis and large‑scale SEM services, highlighting performance gains and operational benefits.
The article chronicles Doug Cutting’s path from his Stanford studies and early Xerox work through the creation of Lucene, Nutch, and Hadoop, highlighting how open‑source innovations and Google’s technologies propelled Hadoop to become a cornerstone of modern big‑data processing and its future outlook.
This article explains SparkSQL's query processing pipeline—from parsing and logical planning through optimization and physical execution—highlighting why it often outperforms Hive on MapReduce by reducing I/O, minimizing shuffle stages, and reusing JVMs.
This article explains the fundamentals of distributed computing, covering sharding algorithms, message‑queue based task distribution, an overview of Hadoop and its MapReduce model, and the characteristics of offline batch processing for large‑scale data workloads.
This article explores the 4V characteristics of big data, real‑world data growth examples, historical analogies, Google’s GFS‑MapReduce‑BigTable model, Hadoop’s architecture and HDFS processes, HBase components, NoSQL alternatives, and practical big‑data applications at Tencent and beyond.
This article explains how bitmap data structures, combined with compression and in‑memory techniques, enable fast, flexible, and scalable multi‑dimensional analytics for large‑scale data platforms, addressing historical marketing inefficiencies and outlining future directions such as memory‑mapped files and distributed bitmap computation.
The author recounts delivering a MapReduce talk, then uses a vivid pizza sauce preparation story to illustrate how mapping chops ingredients and reducing blends them, effectively explaining distributed data processing concepts to a non‑technical audience.
This article chronicles Baidu’s eight‑year journey from early Hadoop adoption to advanced MPI, DAG engines, and real‑time streaming platforms, detailing architectural milestones, performance optimizations, and practical lessons for large‑scale offline and online data processing.
The article explains Apache Spark’s memory‑based distributed computing model, its advantages over Hadoop’s MapReduce, key features, fault tolerance, deployment modes, ecosystem components, and the scenarios where Spark is most effective for large‑scale data analytics.
The interview with Fourinone founder Peng Yuan explores the framework's evolution from a parallel computing project to a 220 KB distributed system with its own NoSQL database engine CoolHash, compares it to Hadoop, and discusses its open‑source release, technical design choices, and real‑world deployments in finance and enterprise environments.