This article details how Yuedu Group designed and implemented an overseas big data platform, covering overall system architecture, offline data‑warehouse construction with dimensional modeling, real‑time streaming using Oceanus and ClickHouse, and future plans for cost reduction and data quality assurance.
This article shares Qunar's experience using Flink together with Apache Iceberg 0.11 to address real‑time data warehouse challenges, covering background pain points, Iceberg architecture, solutions for Kafka data loss and Hive latency, and optimization practices such as small‑file handling, sorting, and checkpoint management.
This article examines the challenges of using Kafka, Flink, and Hive for real‑time data warehousing, introduces Apache Iceberg 0.11 as a solution, details its architecture, query planning, Flink integration, code examples, optimization techniques, and summarizes the benefits for large‑scale data processing.
This article walks through a complete real‑time recommendation system built on Apache Flink, detailing its v2.0 architecture, modules for user behavior, interest, and product profiling, the recommendation algorithms (hot‑list, collaborative filtering, item similarity), and step‑by‑step Docker deployment of MySQL, Redis, HBase, and Kafka.
This article presents Qunar's practical experience with Flink and Iceberg 0.11, covering background challenges such as Kafka data loss and Hive metadata pressure, explaining Iceberg architecture, query planning, and detailed solutions including real‑time ingestion, small‑file handling, sorting, and code examples for seamless migration.
This article presents a comprehensive walkthrough of a Flink‑powered recommendation system, detailing its v2.0 architecture, module functions, recommendation algorithms (hotness, product similarity, collaborative filtering), front‑end and back‑end UI, and step‑by‑step Docker deployment of MySQL, Redis, HBase, and Kafka services.
Tencent Cloud Oceanus, a computing service powering internal apps like WeChat and external partners such as Bilibili, scales to over 30,000 cores handling 5 PB daily and 500,000 jobs, and tackles Flink SQL’s syntax, function and operational limits with table‑valued functions, incremental and enhanced tumble windows, and caching‑based retraction optimization that cuts downstream data volume up to 30× and improves join performance by about 20 %.
This article provides a detailed introduction to data middle platform concepts, covering data aggregation, ingestion tools, offline and real‑time development, data governance, service layers, monitoring, and deployment patterns, illustrating how enterprises build unified data ecosystems across various industries.
This article explains how JD constructs its OLAP platform from data ingestion to storage, querying, and management, describing the diverse data sources, real‑time and offline processing, scalability, consistency, fault tolerance, and future optimization plans, while addressing key technical challenges and solutions.
This article provides a comprehensive overview of data middle platforms, covering data aggregation, collection tools, offline and real‑time development, scheduling, baseline control, heterogeneous storage, data governance, service layers, monitoring, and the architectural differences between offline and real‑time data warehouses.
This article explains how to use Alibaba's Canal to capture MySQL binlog changes, forward them to Kafka, and process the incremental data in real time with Spark Streaming, including installation, configuration, client development, Spark code, testing, and troubleshooting dependency conflicts.
This article presents a comprehensive overview of data lake implementations, detailing Huawei's production‑scene platform, a real‑time financial data lake architecture using Kafka, Flink and Iceberg, and Soul's Delta Lake practice with Spark, Hive, and custom ETL tools, highlighting design choices, processing flows, and operational considerations.
Meituan built an SSD‑based application‑layer cache for Kafka that bypasses PageCache contention between real‑time and delayed jobs, classifies log segments across SSD and HDD, limits flush rates, and achieves up to 80% latency reduction while guaranteeing stable real‑time consumption.
This article explains how to build a real‑time data processing platform using Flink, covering the Lambda architecture, design approaches, SQL and custom‑Jar task definitions, UI drag‑and‑drop, cluster resource management on Yarn and Kubernetes, submission modes, scheduling, permission and metadata handling, logging, and monitoring with Prometheus and Grafana.
The article describes Alibaba Cloud's Data Lake Formation service, presenting a unified, real‑time, and low‑latency solution for ingesting heterogeneous data sources—including RDS, DTS, TableStore, and SLS—into an OSS‑backed data lake using templates, a Spark‑based ingestion engine, and modern file formats such as Delta Lake.
Xianyu’s premium‑product library combines interpretable, multi‑dimensional metric models built from structured product and user attributes with real‑time and offline pipelines to systematically tag high‑quality items, delivering services via HSF and a message bus, and has driven over 20% click‑through growth and nearly doubled conversion rates.
data pipelinefeature engineeringproduct recommendation
This article explains how to design a scalable log collection system for big‑data platforms by combining Flume for data ingestion, Kafka for buffering and high‑throughput transport, and downstream processing components, providing configuration examples and best‑practice recommendations.
The article describes Baidu App's real-time mobile super-resolution using a VDSR-based model with pruning and depthwise separable convolutions, optimized via application-layer and inference engine techniques to halve latency and memory, enabling on-device high‑def image/video enhancement, reducing server load, and supporting iOS/Android integration.
This article presents the concepts, architecture, and practical techniques of an online real‑time feature system used in intelligent risk‑control, covering feature definition, time‑window types, calculation functions, distributed processing, low‑latency storage, and operational challenges in high‑concurrency environments.
This article presents a modular architecture for real‑time ETL log analysis that combines Flume for log collection, Kafka as a buffering layer, Storm for stream processing, Drools for rule‑based data transformation, and Redis for fast storage, detailing installation, configuration, and code integration steps.
This talk presents the architecture, feature extraction, storage, real-time computation, monitoring, and optimization strategies of 58.com’s commercial DMP platform, detailing business requirements, system design across data, storage, compute, and service layers, and future plans for unified services and advanced analytics.
This article provides an extensive summary of data middle platform concepts, covering data aggregation, collection tools, offline and real‑time development, data governance, service layers, warehouse construction, and operational practices, illustrating how enterprises build and manage a unified data ecosystem.
This article presents an in‑depth overview of Weibo's large‑scale machine learning platform, detailing its multi‑layer architecture, development workflow, CTR model evolution, and how Apache Flink is employed for real‑time data processing, sample services, multi‑stream joins, multimedia feature generation, and future roadmap plans.
This article details the evolution of Ctrip's flight ticket data warehouse, describing its historical tech stack, current architecture—including Hive, Presto, ClickHouse, CrateDB, and Flink—data synchronization methods, layer design, quality monitoring, and a real‑time price‑monitoring use case.
This article details the design, installation, and modular integration of Flume, Kafka, Storm, Drools, and Redis to build a real‑time log analysis pipeline for ETL systems, discussing architecture, configuration, code examples, and practical considerations for scalability and fault tolerance.
This article presents a modular architecture for real‑time ETL log collection and analysis, detailing the installation and configuration of Flume, Kafka, Storm, Drools, and Redis, and explains how their integration improves fault tolerance, scalability, and processing speed.
This article explains how Alibaba’s XiaoMi team constructs a full‑cycle AI pipeline—covering real‑time and offline data processing, high‑dimensional visualization, model training, iterative feedback, and Spark‑based deployment—to accelerate intelligent product iteration while addressing common engineering pain points.
The article presents an AI‑driven follow‑subtitle system for video streaming that uses server‑side face detection and tracking to attach speech‑bubble bullets to characters, synchronizing trajectories with playback via a client SDK, while addressing cut‑scene handling, latency, and power constraints.
The presentation by Bilibili's real‑time platform lead details the design and implementation of a Flink‑based streaming data platform, explains how AI workloads are integrated, shares architectural decisions and operational insights, and provides the full slide deck for knowledge dissemination.
The article details Xiaomi's unified streaming data platform, its three‑generation evolution from Scribe/Kafka/Storm to Talos and Flink, the current architecture supporting billions of records daily, and future plans to unify offline and real‑time warehousing with Flink SQL.
This article details MaFengWo’s practical experience using Kafka across three core scenarios—real‑time storage, analytical data source, and business data subscription—while describing a four‑stage evolution that includes version upgrades, resource isolation, security and monitoring enhancements, and a comprehensive subscription platform, followed by future improvement plans.
This article examines the design of a real-time data warehouse built on Flink, Kafka, and HBase, compares it with traditional offline warehouses, and discusses key challenges such as data accuracy, latency, and the complexity of maintaining real-time dimension tables.
This article presents a step‑by‑step guide for building a real‑time data pipeline using Kafka as a message buffer, Spark‑Streaming's Direct Approach for processing, and HBase for storage, including code examples, Maven configuration, local cluster setup, and troubleshooting tips.
This article traces the evolution of Youzu's data platform ingestion, comparing early HTTP/script methods with modern DTS and real‑time ETL solutions, evaluating middleware choices, detailing core system architectures, and outlining future improvements for reliable, scalable data access.
This article outlines the challenges of various big‑data scenarios such as financial risk control, recommendation systems, and social feeds, explains why Spark is chosen over alternatives, describes a one‑stop data platform architecture with Spark‑HBase integration, and shares best‑practice tips and case studies.
This article explains Ant Financial's online graph computing technology, covering its financial‑grade graph database, real‑time anti‑cashout use cases, high‑performance graph cache, flow‑graph fusion, dynamic DAG execution, and how these innovations support massive, low‑latency financial services.
This article explains how Spark Streaming's DStream abstraction is built from InputDStream through successive transform operators, details the internal ForEachDStream implementation, describes the job generation and scheduling workflow, and outlines how Beike's real‑time platform leverages these mechanisms for large‑scale streaming tasks.
This article explains how eBay’s Sherlock.IO monitoring platform processes billions of logs, events, and metrics daily using Flink Streaming jobs, detailing a metadata‑driven architecture, shared job strategies, Heartbeat‑based monitoring, job isolation, back‑pressure handling, and real‑world use cases such as Event Alerting, Eventzon, and Netmon.
This article explains how massive online data can be captured, structured, and analyzed in real time using a Lambda‑style architecture, then introduces a simplified Lambda‑Plus design built on Alibaba Cloud's Tablestore and Blink to meet both batch and streaming requirements while reducing operational complexity.
This article explains why Apache Kafka has become a cornerstone of modern big‑data pipelines by detailing its high‑throughput, fault‑tolerant publish‑subscribe architecture, real‑time processing capabilities, extensive language support, scalability mechanisms, and the wide range of use cases adopted by leading enterprises.
This article explores the design of a large‑scale, real‑time sentiment analysis system, detailing the data ingestion, processing, and storage requirements, comparing Lambda and Kappa architectures, and presenting an Alibaba Cloud solution that combines Tablestore and Blink for unified batch‑and‑stream processing.
To overcome severe performance degradation of Flink sliding windows over very long time ranges, Youzan engineers applied time‑slicing based on the greatest common divisor of window length and slide step, reducing state writes and timers, which yielded 3‑8× speedups in production.
This article introduces the design and implementation of a real‑time user behavior engineering platform at Qunar using Apache Flink, covering Flink's core characteristics, distributed runtime, DataStream programming model, fault‑tolerance, back‑pressure handling, event‑time processing, windowing, watermarks, and practical code examples for filtering, splitting, joining, and state management.
The article describes YouZan's business reconciliation platform for distributed systems, which detects and quantifies data inconsistencies by offering easy plug‑in integration, a four‑step orchestrated workflow, high‑throughput offline processing with Spark, second‑level real‑time event handling, a three‑layer architecture, and health monitoring for transaction chains.
The article chronicles Alibaba's ten‑year journey of building and scaling its big data platform—from early Oracle clusters and Hadoop‑based Cloud‑Ladder 1 to the self‑developed ODPS/MaxCompute, real‑time Blink engine, and the unified DataWorks ecosystem—highlighting key technical milestones, performance breakthroughs, and operational challenges that powered successive Double 11 shopping festivals.
This article chronicles Alibaba's ten‑year journey of building and scaling its big‑data platform—from early Oracle clusters and Hadoop, through the launch of ODPS and MaxCompute, to global cloud expansion and cutting‑edge streaming innovations that now power billions of transactions each Double‑11.
This article details Vipshop's real‑time platform architecture, the migration from Storm and Spark to Flink, Flink's deployment on Kubernetes, and the latest Unified Data Management system that unifies data access across Kafka, Redis, Tair and HDFS.
This article compares Spark Streaming and Flink across runtime models, component roles, programming APIs, task scheduling, time semantics, dynamic Kafka partition detection, fault‑tolerance mechanisms, exactly‑once guarantees, and back‑pressure handling, providing code examples and practical insights for real‑time data processing.
The article describes the evolution of 360's Titan big‑data processing platform through three architectural stages, details its functional modules, explains the DITTO component framework, context and rule‑engine abstractions, and shares practical case studies and personal insights on building a flexible, self‑service data platform.
The Meitu Short Video Real-Time Classification Challenge (MTSVRC), co‑hosted by the PRCV conference, Meitu and the Chinese Academy of Sciences, releases the industry‑largest dataset of over 100,000 five‑to‑15‑second videos in 50 categories, judging both classification accuracy and real‑time speed, and offers cash prizes up to ¥100,000 plus presentation opportunities for top teams.
This article explains the internal workflow of Spark Streaming task submission, detailing how StreamingContext, DStream, receivers, and output operators are transformed into RDD jobs, and includes annotated Scala code examples that illustrate each step of the process.
This article examines Ctrip Ticket's real-time data processing ecosystem, comparing batch and streaming frameworks such as Hadoop, Spark, Storm, Flink, and Spark Streaming, detailing Kafka deployment and configuration, and describing how these technologies are applied in production for log analysis, seat‑occupancy detection, and anti‑crawling.
This article presents a rapid Q&A session with JD data scientists and architects, covering the benefits of algorithm contests for students, the unique advantages of the JDATA competition, scoring formulas, ways to improve results, strong feature extraction, real‑time modeling, algorithm selection, and the value of the competition’s special offer for future employment.
This guide details how to construct a real-time data processing platform on CentOS 7 using the Hadoop ecosystem—installing and configuring Zookeeper, Maven, Hadoop, Kafka, HBase, Spark, and Flume—followed by a Spark Streaming job that consumes Kafka messages and writes them into HBase.
This article explains the objectives, methods, architecture, and key algorithms of modern recommendation systems, covering popular, manual, related, and personalized approaches, the data pipeline, real‑time challenges, cold‑start handling, diversity, content quality, and exploration‑exploitation strategies.
This case study details how Beijing 恒远华信息技术有限公司 built a dynamic face‑capture and real‑time recognition solution on Huawei FusionInsight HD, leveraging deep‑learning algorithms, distributed storage, and stream processing to handle hundreds of millions of faces with high speed, efficiency, and security.
This article details Lianjia's journey of designing and implementing a low‑latency, stable real‑time computing platform using Spark Streaming on YARN, covering technical selection, architecture components, version compatibility challenges, exactly‑once semantics, graceful shutdown, Kafka tuning, and future enhancements.
Alibaba’s latest bundle‑mining system leverages weighted graph embedding and real‑time sampling to recommend complementary products, replacing traditional item‑to‑item similarity, boosting click‑through rates by up to 13% offline and 4% online during the Double‑11 promotion while handling billions of edges.
Analyzing 2017’s big data boom, the article explores how the 4V characteristics—volume, variety, velocity, and value—spurred innovations like distributed storage, NoSQL, real‑time stream processing, and AI integration, and predicts future hotspots such as SQL resurgence, cloud‑based platforms, and AI‑driven analytics.
This article walks through the design, deployment, and code implementation of a real‑time spam detection pipeline using Apache Storm, comparing it with Hadoop, detailing cluster setup, topology components, data flow, and how to package and run the solution on a distributed Storm cluster.
This article describes how Tongcheng Travel’s monitoring system evolved from a monolithic design to a distributed and big‑data‑based architecture, introducing real‑time processing with Storm, machine‑learning‑enhanced alerts, and a multivariate linear regression model that dramatically improves traffic anomaly detection accuracy.
Using Meituan‑Dianping’s hospitality online feature system as a case study, the article describes how feature production scheduling evolved from offline batch ETL to automated, metadata‑driven pipelines and sub‑second streaming, detailing the underlying architecture, incremental updates, storage abstraction, write‑shaving, atomicity, and recovery mechanisms.
This article explains why Ctrip creates user profiles, describes the product and technical architectures, and details the data collection, computation, storage, high‑availability querying, and monitoring components that power its personalized travel recommendations and services.
This article explores practical techniques for synchronizing heterogeneous data sources, performing batch and incremental analytics with Hadoop and Spark, designing low‑latency real‑time computation pipelines, implementing push notifications, and choosing appropriate storage solutions—from in‑memory caches to distributed databases—while addressing performance, reliability, and scalability challenges.
This tutorial explains Apache Storm’s core principles, architecture, and development workflow, covering its relationship with Hadoop, key concepts such as spouts, bolts, tuples, and topologies, and provides step‑by‑step code examples for a word‑count program and a call‑log analysis application.
This article describes how Ctrip's information security team redesigned its data platform from a simple RabbitMQ‑MySQL pipeline to a scalable, real‑time and offline big‑data architecture using Kafka, Storm, Hadoop, Spark, and a custom count server, dramatically improving processing capacity and supporting risk‑control operations.
This tutorial introduces Apache Storm’s fundamental principles and development workflow, providing a PDF guide and source code for two practical examples—real‑time word‑count and call‑record aggregation—while covering its definition, use cases, relationship with Hadoop, core concepts, cluster architecture, and step‑by‑step usage.
This technical guide demonstrates how to configure an iOS audio capture session using Swift, implement a custom multiplexer class to safely route microphone input to multiple processing destinations, and efficiently handle real-time audio data streams for language learning applications.
The article describes how Ctrip redesigned its real-time user behavior service using a Java‑Kafka‑Storm stack with Redis and MySQL, detailing the architecture, real‑time processing, availability, performance, scalability, and deployment strategies to handle billions of events daily.
This article details eBay's year‑long development of an enterprise‑grade, Kafka‑driven data transmission platform, covering its architecture, core services, monitoring and automation strategies, as well as performance tuning techniques that enable high throughput, low latency, and reliable cross‑data‑center replication.
Tencent Analytics (TA) is a free web analytics platform that processes terabytes of daily data in real time, using a custom architecture featuring JavaScript collection, event streaming, in‑memory computation, and NoSQL storage with Redis and LevelDB, offering site owners instant insights and high availability.
This article explains the architecture, real‑time computation framework, and storage solutions of Tencent Analytics, detailing how massive TB‑level web‑traffic data are collected via JavaScript, processed in memory‑centric streaming components, and stored using Redis and LevelDB to achieve second‑level updates.
This article outlines Alibaba Cloud's big‑data platform challenges and solutions during the 2016 Double 11 event, covering sub‑second real‑time processing, multi‑million‑records‑per‑second throughput, full‑day high availability, and massive offline workloads exceeding hundreds of petabytes.
This article presents a comprehensive overview of Ctrip's Aegis risk control system, detailing its modular architecture, rule engine, data service layer, Chloro analytics platform, and future directions, while highlighting the use of streaming, big‑data processing, and machine‑learning models for real‑time fraud detection.
This article explains how to apply machine learning to real‑time distributed processing, covering model basics, data collection, parameter learning, topology design, message flow control, and the Hurricane framework architecture.
Hurricane frameworkTopology Designdistributed systems
This article outlines the design of a payment risk control system, detailing functional and non‑functional requirements, core components such as real‑time, near‑real‑time, and batch engines, rule and penalty centers, and explains the role of CEP and Drools in achieving flexible, high‑performance fraud detection.
The article explains the functional and non‑functional requirements, common pitfalls, and detailed architecture—including real‑time, near‑real‑time, and batch engines, rule and penalty centers, and CEP technology—of a payment risk control system aimed at detecting and mitigating fraud while maintaining performance and flexibility.
The article explains how Spark has become the leading open‑source big‑data platform, highlighting its superior speed, in‑memory processing, real‑time streaming, and built‑in machine‑learning library compared with Hadoop’s slower, disk‑based MapReduce approach and reliance on external storage and ML tools.
This article examines the design considerations of real-time distributed streaming systems, outlines their background and characteristics, compares the architectures of Spark Streaming and Storm, discusses primitives, message passing, high availability, storage models, and integration with production environments, providing practical insights for architects.
This article explains how Baidu Maps’ open platform handles massive daily location data through real‑time and offline pipelines, Hadoop‑based offline computing, stream processing, and query engines built on MySQL, Redis, and Apache Kylin, while outlining future big‑data enhancements.
This article details the architectural evolution of Kuaidi Dache from 2013‑2014, covering LBS bottlenecks, MongoDB scaling, long‑connection stability, distributed system refactoring, a wireless open platform, real‑time monitoring with Storm and HBase, and a data‑center built on sharding and synchronization.
Facing explosive traffic growth in 2014, Didi’s ride‑hailing platform tackled critical challenges by redesigning its LBS architecture, replacing unstable long‑connection services with an AIO‑based framework, partitioning databases, adopting Dubbo and RocketMQ for distributed processing, and building a real‑time monitoring and data center using Storm, HBase, and custom SQL‑to‑HBase translation.
This article details the rapid traffic growth challenges faced by Kuaidi Dache from 2013‑2014 and presents representative architectural bottlenecks and the engineering solutions—including LBS optimization, long‑connection redesign, distributed refactoring, a wireless open platform, real‑time monitoring, and data layer transformation—that enabled stable, scalable, high‑performance ride‑hailing services.
The article explains how Qiniu's data processing platform has evolved from a simple real‑time URL‑based model to a more complex architecture featuring separate caching, agent services, discover monitoring, and container‑based elastic scaling to handle massive unstructured data workloads.
This article explains why Kafka is essential for activity and operational data pipelines, outlines key use cases such as news feeds, relevance ranking, security, monitoring, and reporting, and details its deployment topology, design decisions, and message persistence strategies.
The article explains the need for low‑latency, high‑performance, distributed real‑time processing, outlines the challenges such systems must address, and introduces Storm as a Hadoop‑like framework for stream processing, detailing its architecture, fault‑tolerance mechanisms, transactional topology, and large‑scale deployment at Taobao.
Apache Storm, Spark Streaming, and Samza are three open‑source, low‑latency, scalable distributed systems for real‑time data processing; this article outlines their architectures, key concepts, differences in data handling, state management, delivery guarantees, and typical use‑cases to help you choose the right framework.
This article introduces Apache Storm, Spark Streaming, and Samza, explains their architectures, common features, key differences such as delivery guarantees and state management, and provides guidance on selecting the most suitable framework for various real‑time big‑data use cases.
This article chronicles Suning’s journey from early SAP‑based data warehouses to a modern, open‑source big data platform featuring real‑time collection, Hadoop‑Hive offline processing, Storm‑based streaming, and a visual development environment, highlighting how each layer addresses growing data volume, variety, and business demands.
This article introduces Apache Storm, Spark Streaming, and Apache Samza, outlines their architectures, highlights commonalities and differences such as delivery guarantees and state management, and offers guidance on selecting the most suitable framework for various real‑time big‑data use cases.