A rare silent data‑corruption bug in JDK 25’s G1GC caused Parquet and ORC files written by Spark and Flink to become unreadable, prompting a multi‑stage investigation that traced the issue to an optional evacuation flaw affecting JNI‑pinned objects, which was later back‑ported and fixed in the OpenJDK community.
Xiaohongshu, with over 3.5 billion monthly users and daily logs in the trillions, migrated 500 PB of data to Alibaba Cloud and iterated its data platform through four architecture generations—ClickHouse‑based ad‑hoc, Lambda, Lakehouse, and a unified incremental compute model—cutting resource, development, and storage costs to one‑third while delivering sub‑10‑second query latency at petabyte scale.
Xiaohongshu, with over 350 million monthly users and daily logs in the billions, migrated its data platform from AWS to Alibaba Cloud and iterated four times—from a ClickHouse‑based ad‑hoc layer to a Lambda architecture and finally a Lakehouse with incremental compute—cutting architecture complexity, resource cost and development effort each to about one‑third while delivering second‑level analytics on trillion‑scale data.
This article details Zhihu's end‑to‑end experience of migrating Spark SQL workloads to the open‑source Gluten framework, covering background performance benchmarks, the architecture of Gluten and Velox, consistency and performance challenges encountered during migration, the concrete fixes applied, and the resulting resource savings and future plans.
The article details Xiaohongshu's step‑by‑step migration from a simple ClickHouse‑based analytics stack to a Lambda‑style 2.0 architecture and finally to a Lakehouse‑based 3.0 design, highlighting concrete performance numbers, cost reductions, and the definition of a generic incremental‑compute model (SPOT) that underpins the evolution.
Xiaohongshu transformed its data platform from a simple ClickHouse‑based ad‑hoc analysis to a Lambda‑style architecture and finally to a lakehouse with generic incremental compute, cutting architecture complexity, resource and development costs by one‑third while delivering second‑level queries over trillions of rows.
This guide compiles the most frequent interview pitfalls for data warehouse roles, covering SQL join and aggregation errors, window function misuse, subquery versus CTE performance myths, dimensional modeling mistakes, SCD implementation traps, layered design issues, data quality handling, ETL traps, Hive and Spark performance questions, real‑time warehousing considerations, and effective interview strategies.
OPPO tackles explosive multimodal data growth by unifying metadata with Gravitino and boosting I/O performance using the open‑source Curvine cache, delivering a four‑layer data‑lake architecture that resolves data islands, metadata chaos, and bandwidth bottlenecks while achieving near‑commercial query speeds.
The article details Xiaohongshu's evolution from a simple ClickHouse‑based analytics layer to a Lambda‑enabled 2.0 stack and finally a Lakehouse‑based 3.0 architecture, showing how each iteration reduced infrastructure complexity, resource consumption and development effort by roughly one‑third while supporting trillions of daily events and AI‑driven use cases.
Xiaohongshu transformed its data platform from a simple ClickHouse‑based ad‑hoc analysis to a Lambda‑style architecture and finally to a lakehouse built on Iceberg, StarRocks, Flink and Spark, cutting architecture complexity, resource and development costs by two‑thirds while supporting trillions of daily events with sub‑second query latency.
The article details Xiaohongshu's four‑stage data‑platform evolution—from a simple ClickHouse ad‑hoc setup to a Lambda‑based 2.0 design and finally a lakehouse‑driven 3.0 architecture—highlighting the adoption of general incremental compute, cost‑reduction to one‑third, performance gains of up to ten‑fold, and the SPOT standards that guide the new system.
Xiaohongshu transformed its data platform from a simple ClickHouse‑based analytics stack to a unified lakehouse with generic incremental compute, cutting architecture complexity, resource cost, and development effort by roughly one‑third while supporting petabyte‑scale, sub‑second queries across its 350 million‑user app.
The article dissects a credit data system architecture, detailing six logical layers—from multi-source data collection and feature engineering (including graph features and feature stores) to model training, real‑time stream processing, decision engine integration, and privacy‑preserving computation—while explaining the trade‑offs, tools, and performance targets needed for accurate, low‑latency risk assessment.
Learn how to demonstrate real Spark optimization skills in data‑warehouse interviews by exploring two detailed case studies—small‑file merging in ODS‑to‑DWD ETL and shuffle‑skew mitigation in DWS aggregation—plus key interview questions and practical troubleshooting steps that separate theory from hands‑on expertise.
During a gray‑release of JDK 25 on Ctrip's massive Spark and Flink clusters, silent data corruption appeared in Parquet and ORC files, traced to a G1GC Optional Evacuation bug that moved JNI‑pinned objects, a root cause later back‑ported and fixed in JDK 25.0.3.
This article analyzes the challenges of scaling ClickHouse within Baidu’s MEG data platform and details a lake‑house solution that decouples storage and compute, integrates a meta‑service for transparent data access, optimizes query performance through caching, data roll‑up and layout tuning, and introduces a unified query gateway that gracefully falls back to Spark for complex workloads.
This article walks through three concrete methods for moving massive ClickHouse datasets—up to 100 billion rows—to Doris, detailing catalog integration, file export with stream load, and Spark‑based pipelines, while sharing real‑world performance results and pitfalls.
This article provides a comprehensive guide to Spark UI, explaining each primary and secondary tab, the key metrics they expose, and how to interpret them for performance bottleneck detection, followed by two detailed case studies and practical tuning recommendations for Spark workloads.
This article presents a comprehensive design for a layered storage and unified scheduling platform that archives MySQL historical data, reduces storage costs, ensures high‑performance queries, and enables efficient data analysis through tiered hot, warm, and cold storage using big‑data technologies.
A performance comparison of Amazon EMR Serverless Storage on a 3 TB TPC‑DS benchmark shows up to 55 % cost reduction and 25 % faster runtimes for shuffle‑intensive Spark jobs, while outlining usage limits and providing Python tools to analyze shuffle data from Spark event logs.
This article details how Alibaba Cloud EMR Serverless Spark combined with the Paimon lakehouse framework enables Taobao Flash Sale’s retail data team to achieve low‑latency, high‑throughput real‑time analytics, batch processing, and feature generation, outlining architecture evolution, performance gains, and practical Spark tuning techniques.
This article explains how to inject the LanceSpark plugin into Spark, covering the core LanceSparkSessionExtensions class, various ways to register extensions, the custom parser and planner strategy implementations, and the underlying Spark mechanisms such as injectParser, injectPlannerStrategy, and PredicateHelper.
Facing a ClickHouse cluster shutdown, the author explores three migration methods—using Doris’s ClickHouse catalog, exporting to files with Broker/Stream Load, and Spark—to transfer ~10 billion rows to Doris, evaluating each for simplicity, bugs, and performance, and sharing detailed steps, code snippets, and benchmark results.
This article compiles the most frequently asked interview questions for 2026 data‑warehouse development engineers, covering core concepts, layer architecture, SQL optimization, window functions, Hive vs Spark, data skew solutions, modeling metrics, slowly changing dimensions, scheduling tools, data quality monitoring, and real project experience.
This article explains GitHub's AI ecosystem—Models, Copilot, Agents, Spaces, Spark, Instructions, Skills, and the Model Context Protocol—clarifying each component, their relationships, and practical steps for developers to integrate them into their workflow.
This technical report details how Vivo’s big‑data platform adopted Celeborn as its remote shuffle service, evaluated alternatives, tuned hardware and software configurations, implemented performance and stability enhancements, and outlines future operational and community‑driven improvements for handling petabyte‑scale shuffle workloads.
This article chronicles a ten‑year evolution of a self‑built big data platform—detailing early Hadoop clusters, successive migrations to Spark, Hive, Hudi, and StarRocks, the operational challenges encountered, and the comprehensive shift to Alibaba Cloud EMR Serverless that delivered significant cost, performance, and stability gains while outlining future intelligent‑ecosystem plans.
Apache Kyuubi, an enterprise‑grade multi‑tenant data gateway, replaces Livy and Flink SQL Gateway to support multiple engine versions, cross‑cluster elastic scheduling, high‑availability batch jobs, and traffic control, dramatically reducing deployment complexity, improving resource utilization, and accelerating release cycles for large‑scale Spark and Flink workloads.
Since its 2016 launch, Alibaba Cloud EMR has transformed from a basic open‑source Hadoop service into a high‑performance, AI‑enabled big‑data platform, delivering optimized I/O, vectorized processing, and integrated AI functions such as natural‑language SQL, StarRocks and Spark enhancements, while supporting diverse industry workloads.
Learn the six typical scenarios that cause data loss when writing to Paimon—ranging from checkpoint failures and misconfigured partial‑update mode to incorrect sequence fields, snapshot retention issues, concurrent bucket writes, and outdated Spark versions—and how to prevent each problem.
When big‑data workloads hit the CPU wall, BIGO’s adoption of the open‑source Gluten project delivers native‑engine execution for Spark and a roadmap for Flink, achieving up to 30% end‑to‑end speedup, 50% memory savings, and a scalable, cost‑effective data processing platform.
This article explains Spark Streaming's rate control mechanisms, covering static limits, the dynamic back‑pressure feature introduced in Spark 1.5, the PID‑based estimator, RPC communication, and how Guava's token‑bucket RateLimiter enforces the calculated thresholds to ensure stability and optimal throughput.
A comprehensive list of interview topics covering self‑introduction, career moves, data‑warehouse design, team building, architecture comparisons, fact‑table classification, common dimensions, performance tuning, and data‑governance for aspiring big‑data engineers.
This article compiles 18 essential interview topics for data warehouse engineer roles, covering self‑introduction, architecture layering, dimensional modeling, HDFS operations, Spark vs MapReduce, join implementation, SQL challenges, OLAP selection, real‑time quality assurance, and job transition considerations.
This article explores multiple techniques—including bitmap indexing, Bloom filters, external sorting, Spark, and layered bitmap structures—to efficiently remove duplicate QQ numbers from a dataset of up to one billion entries while keeping memory usage below a gigabyte and maintaining high accuracy.
This article explains how Baidu’s search data team redesigned its data warehouse with wide‑table modeling, Parquet columnar storage, and a Spark‑ClickHouse fusion engine, eliminating redundancy, cutting query latency from minutes to seconds, and enabling self‑service analytics for thousands of users.
Despite the hype around Flink and AI models, Spark 4.0’s release brings a lightweight Python client, Spark Connect GA, enhanced SQL optimization, vectorized execution, and AI integration, reaffirming its leading position in the big‑data ecosystem while hinting at future challenges and innovations.
This article details how a leading automotive parts supply‑chain platform migrated from a traditional Hadoop stack to Alibaba Cloud EMR Serverless Spark and DataWorks, achieving faster, more elastic, and cost‑effective data processing, enhanced AI integration, and significant operational improvements across multiple business scenarios.
This article reviews common interview questions for data development roles, covering Spark stage partitioning and optimization, criteria for evaluating data warehouses, Flink's handling of late data, and provides practical answers and resources to help candidates deliver standout responses.
Grab replaced its fragmented Grafana‑Superset stack with a StarRocks‑backed Iris platform, achieving over ten‑fold query speedups, 40% lower resource usage, and a unified real‑time and historical data store for Spark observability across its Southeast Asian super‑app ecosystem.
This article shares three high‑quality interview questions from a JD big‑data interview, covering practical Spark tuning, Flink's exactly‑once guarantees in production, and Paimon's asynchronous merge mechanism, and explains how to answer them with real‑world scenarios.
This article compiles a comprehensive set of Spark interview questions frequently asked by leading tech companies, providing detailed explanations of Spark’s performance mechanisms, architecture, RDD persistence, checkpointing, streaming, dependency types, HA setup, and practical coding examples to help data warehouse engineers prepare effectively.
The article explains how the vivo Internet Big Data team containerized offline Spark jobs and deployed them with the Spark Operator on a mixed online‑offline Kubernetes cluster, using elastic scheduling and resource‑over‑subscription to boost CPU utilization by 30‑40% and handle over 100,000 daily tasks.
The Apache Hudi Asia technical salon held in Beijing on March 29 gathered over 230 on‑site participants and 16,000 online viewers, featuring expert talks from leading Chinese tech companies that showcased real‑world Hudi implementations, performance optimizations, and future roadmap for data‑lake technologies.
iQIYI transformed its real‑time data warehouse by replacing a costly Kafka‑based Lambda stack with a unified stream‑batch Iceberg lake, cutting storage expenses by 90%, halving compute costs, extending data retention, and delivering minute‑level freshness for 90% of use cases while preserving second‑level processing where needed.
This step‑by‑step guide explains how to use EMR Serverless Spark together with the StarRocks Spark Connector to create a workspace, upload the connector JAR, configure network connections, create databases and tables in StarRocks, and perform read/write operations via SQL sessions, Notebook sessions, or batch Spark jobs, complete with code examples and UI screenshots.
The article analyzes how the rapid rise of open‑source large‑model AI in 2025 is reshaping the data development profession, urging developers to transition from specialized data‑engineer roles to full‑stack AI data engineering skills such as distributed computing, lake‑house architectures, and model tuning.
The article introduces Blaze, Kuaishou's Rust‑powered native execution engine that vectorizes Spark SQL workloads, explains its architecture and operation, presents benchmark results showing up to 50% latency reduction, and details internal deployments, industry case studies, community collaborations, and the 2025 roadmap.
This article shares a step‑by‑step guide for migrating billions of rows safely and quickly, covering divide‑and‑conquer batching, dual‑write architectures, tool selection, shadow testing, and rollback plans, with concrete Java and Spark code examples and practical pitfalls to avoid.
This article analyzes the transition from a tightly coupled storage‑compute architecture to a decoupled model, detailing how Kubernetes, Kyuubi, Celeborn, Blaze, and Hue together solve resource inefficiencies, improve scalability, and boost query performance in modern big‑data environments.
The article compares Java, Scala, SQL, and Python for data‑engineering tasks, arguing that Python’s versatility, rich ecosystem, and ease of use make it the preferred language for both small‑scale and massive Spark workloads despite its performance trade‑offs.
This article explains the challenges of big‑data processing in the cloud era, introduces Spark’s native‑language SQL engine rewrites, discusses vectorization and code generation techniques, describes cloud‑native storage‑compute separation with Remote Shuffle services such as Apache Celeborn, and presents the production benefits of Alibaba Cloud’s EMR Serverless Spark.
Weifin, a fintech innovator, tackled massive data‑scale challenges by adopting Alibaba Cloud EMR Serverless Spark, building a unified Spark‑based platform that supports data collection, lake ingestion, distributed machine‑learning training, and intelligent risk‑control applications, while achieving performance gains, cost reduction, and scalable automation.
Chronon is an open‑source framework that centralizes feature definitions to guarantee training‑inference consistency, eliminates complex ETL pipelines, and supports real‑time and batch processing across diverse data sources, cutting feature‑development cycles from months to under a week, as demonstrated by Airbnb’s 40,000‑feature deployment.
This article shares practical techniques for writing clean, efficient SQL in large‑scale data environments, covering predicate pushdown, sub‑queries, deduplication strategies, bucket optimization, and automation with Python‑Spark integration to improve readability and execution speed.
This article details Zhihu's comprehensive cost‑reduction and efficiency‑boosting initiatives for its big‑data platform, covering FinOps‑driven financial operations, hybrid‑cloud architecture, cost allocation models, operational monitoring, and technical optimizations such as erasure coding, ZSTD compression, Spark auto‑tuning, and a remote shuffle service.
This presentation details Tencent's real‑time lakehouse architecture and the four key topics—lakehouse design, intelligent optimization services, scenario‑driven capabilities, and future outlook—covering components such as Spark, Flink, Iceberg, Auto‑Optimize Service, indexing, clustering, AutoEngine, and PyIceberg implementations.
This article presents Tencent's end‑to‑end real‑time lakehouse architecture, detailing its three‑layer design, the Auto Optimize Service modules such as compaction, indexing, clustering and engine acceleration, as well as scenario‑driven capabilities like multi‑stream joins, primary‑key tables, in‑place migration and PyIceberg support, and concludes with future optimization directions.
Bilibili replaced Spark’s unstable External Shuffle Service with a push‑based approach, then deployed Apache Celeborn’s remote shuffle on Kubernetes using HA masters, tiered workers, extensive monitoring, history‑based routing, chaos testing, and seamless Spark, Flink, and MapReduce integration, while planning self‑healing, elastic scaling, and priority‑aware I/O enhancements.
This article provides a comprehensive overview of Apache Paimon, covering its real‑time lake ingestion, unified stream‑batch processing, table types (primary‑key and append‑only), LSM‑tree storage, bucket mechanisms, merge‑engine options, compaction strategies, concurrency control, consumption methods, tag management, data cleanup, and system tables for big‑data workloads.
This article explains how Spark ≥ 3.3 introduces Storage Partitioned Join (SPJ) to avoid costly shuffle operations when joining partitioned V2 source tables such as Apache Iceberg, detailing the required conditions, configuration settings, practical code examples, and various join scenarios including mismatched partitions and data skew.
The article shares practical methods and code examples for making SQL both readable and high‑performing in large‑scale data platforms, covering predicate push‑down with subqueries, deduplication strategies, bucket utilization, and Python‑driven job parameter handling.
This presentation describes Tencent's real-time lakehouse architecture, including data lake compute, management, and storage layers, and details the intelligent optimization services—such as compaction, indexing, clustering, and auto-engine—designed to improve query performance, storage cost, and operational efficiency for large-scale data processing.
The article outlines Bilibili’s revamped recommendation data‑flow architecture that eliminates timing and calculation inconsistencies by snapshotting online features, unifying feature computation in a single C++ library accessed via JNI, and orchestrating label‑join and sample extraction through near‑line Kafka/Flink pipelines, with further performance gains and Iceberg‑based future extensions.
Uber’s data lake on Hadoop stores hundreds of petabytes in Parquet files and, by adopting ZSTD compression, column pruning, and column reordering, achieves up to 79% storage reduction and significant vCore savings, with detailed benchmarks guiding optimal compression levels and open‑source contributions.
This article explains how Spark ≥ 3.3’s Storage Partition Join (SPJ) can avoid costly shuffle operations by using Iceberg tables, outlines the required table properties and Spark configurations, demonstrates the effect with code examples and execution plans, and explores several realistic join scenarios.
The article details Bilibili's comprehensive redesign of its tag system—including background challenges, architectural layers, technical upgrades like Iceberg integration and shard‑based ClickHouse writes, crowd selection methods, online service guarantees, performance metrics, and future plans—showcasing a data‑driven solution that boosts stability, speed, and business coverage.
LinkedIn’s massive Spark workloads suffer from shuffle bottlenecks caused by tiny shuffle blocks, unreliable RPC connections, and data skew, so the authors design Magnet—a push‑merge shuffle service that merges blocks into large chunks, improves disk I/O, tolerates failures, and cuts end‑to‑end job time by nearly 30% regardless of hardware.
The article examines the DBOS concept—a database‑oriented operating system that places a distributed, transactional database at the core of OS services, tracing its roots from early database pioneers to modern cloud workloads and highlighting its potential advantages over traditional Linux‑Kubernetes stacks.
This article explains what constitutes a small file in HDFS, why they harm memory, compute and cluster load, outlines common sources such as data sources, streaming and dynamic partitioning, and provides detailed Hive and Spark solutions—including CombineHiveInputFormat, merge parameters, distribute by, and custom Spark extensions—to efficiently merge small files and improve job performance.
The 8th Tongcheng Travel Big Data Technology Salon in Suzhou featured four expert talks covering Tencent Cloud’s Meson Spark engine, near‑line computing for travel itineraries, a Flink‑based real‑time risk control system, and Apache Paimon’s latest lake‑warehouse innovations, followed by a data‑driven business perspective session.
The rebuilt tag system introduces a three‑layer architecture, standard pipelines, Iceberg‑backed storage and custom ClickHouse sharding, a DSL for crowd selection, and a stateless online service, achieving 99.9% success, sub‑5 ms latency, and supporting thousands of tags across dozens of business scenarios while planning real‑time processing and automated lifecycle management.
The article details Duodian DMALL’s migration from a traditional Hadoop stack to a cloud‑native Spark‑on‑Kubernetes architecture, explaining the motivations, design choices, component selections, operational challenges, and lessons learned through concrete examples and performance observations.
Magnus is Bilibili’s self‑developed intelligent service that continuously optimizes Iceberg tables by scheduling snapshot expiration, orphan‑file cleanup, manifest rewriting, and multi‑dimensional data optimizations—including small‑file merging, sorting, distribution, and index creation—while automatically recommending configurations from real‑time query logs, delivering over 99.9% task success and up to 30% scan‑data reduction.
Bilibili transformed its search indexing pipeline by replacing a manual, low‑throughput process with a distributed KV store (Taishan) and Spark‑based construction, achieving unified data ingestion, reduced resource consumption, faster full‑ and incremental builds, and a shift from daily to hourly indexing cycles.
This article explains document‑level, paragraph‑level, and sentence‑level deduplication for large‑scale LLM pre‑training, introduces the Simhash and Minhash algorithms with step‑by‑step Python examples, and shows how to implement efficient LSH‑based deduplication using Spark.
This article details Bilibili’s implementation of a large‑language‑model‑driven intelligent assistant for its massive big‑data platform, covering background, problem analysis, architectural design, knowledge‑base construction, precision and recall challenges, deployment across offline and real‑time Spark/Flink diagnostics, and future outlooks.
This guide analyzes common Spark Out‑Of‑Memory scenarios—such as massive data volumes, data skew, and improper resource allocation—and provides step‑by‑step configurations, memory‑management tweaks, partitioning strategies, and shuffle optimizations to prevent OOM failures in production.
This article presents JD Retail's data lake architecture overhaul, detailing the shortcomings of the Lambda model, the migration to Flink‑Hudi‑Spark pipelines, performance gains, storage savings, unified APIs, and upcoming improvements for resilience and automation.
This article explains the Spark SQL analysis layer, its core principles, how analysis rules such as ResolveRelations work, and the major pruning optimization introduced in Spark 3.2 that reduces unnecessary rule traversal, illustrated with concrete code examples and Q&A.
This article introduces Apache Hudi Streamer, a versatile Spark‑based data ingestion tool likened to a Swiss Army knife, detailing its core options—including table configuration, continuous mode, source classes, transformers, table services, catalog synchronization, and advanced features—while guiding users on practical pipeline setup.
The article shares practical study habits, detailed note‑taking, proactive questioning, effective communication, and a comprehensive set of interview questions covering Hive, Spark, Kafka, Flink, and other big‑data technologies, illustrated with real examples from a diligent student’s experience.
This article details Bilibili's transformation of its search offline indexing architecture—from a manual, low‑throughput MySQL‑centric process to a distributed, KV‑based, protobuf‑driven pipeline that leverages Taishan storage and Spark, cutting build cycles from days to hours while solving performance, consistency, and maintenance challenges.
Blaze is a Rust‑implemented, DataFusion‑based vectorized execution engine created by Kuaishou to accelerate Spark SQL queries, delivering up to 60% faster computation, 30% average compute‑power gains in production, and extensive architectural innovations such as native engine, protobuf protocol, JNI bridge, and Spark extension, while being open‑source and compatible with Spark 3.0‑3.5.
This article details how Ctrip’s data platform evolved from a single‑IDC design to a multi‑IDC, tiered storage and scheduling architecture, covering the challenges of rapid data growth, the migration to Spark 3 via Kyuubi, the introduction of Celeborn shuffle service, and the resulting performance and reliability gains.
This article introduces the AnalyticDB Spark architecture, explains the need for Spark vectorization, surveys industry vectorized solutions, details ADB Spark's own vectorized implementation with Gluten and Velox, and presents performance test results showing a 6.98‑fold speedup over open‑source Spark.
This article details Bilibili's transformation of its search offline indexing pipeline, moving from manual MySQL‑based processes to a high‑capacity, distributed KV store and Spark‑driven builds, addressing performance, maintenance, and scalability challenges while improving resource efficiency and iteration speed.
This article explains the complete Apache Hudi write pipeline, detailing each step from client creation to commit, and describes the various write operations such as Upsert, Insert, Bulk Insert, Delete, Delete Partition, and Insert‑Overwrite, providing a comprehensive overview for data‑lake practitioners.
This article shares detailed practical experiences from DingXiangYuan's big‑data platform on using Apache Kyuubi and Apache Celeborn, covering architecture, flexible configuration, AuthZ fine‑grained permissions, small‑file and Z‑Order optimizations, Arrow‑based large result transmission, and operational tips such as connection‑level issues and Netty cache handling.
This article introduces the fundamentals of distributed computing, compares major frameworks such as Hadoop, Spark, and Flink, and outlines their key components, performance characteristics, and typical application scenarios including big‑data analytics, cloud services, real‑time streaming, and scientific computing.
Blaze is Kuaishou's self‑developed native execution engine that leverages Rust, DataFusion, and SIMD vectorization to accelerate Spark workloads, offering a 30%+ compute boost, detailed architectural components, deep production‑grade optimizations, and a roadmap for broader adoption.
This article examines the critical role of shuffle in big‑data processing, compares Spark's native shuffle with the External Shuffle Service (ESS) and Remote Shuffle Service (RSS) solutions, introduces Uniffle's architecture and configuration, and shares practical deployment experiences and performance results within a 360 internal environment.
Baidu’s Mobile Ecology team transformed its Feed data warehouse through three progressive stages—hour‑level core tables, a real‑time wide table, and a unified day‑level multi‑version table—consolidating traffic, content, and user data into a single partitioned wide‑table architecture that resolves granularity inconsistencies, cuts processing cost, and delivers real‑time to daily latency for diverse analytics.
This article demonstrates how Java Stream operations such as map, flatMap, groupBy, and reduce can be directly applied to Spark, providing step‑by‑step code examples, explanations of transformation versus action operators, and practical tips for handling exceptions in distributed data processing.
This article explains how Apache Hudi integrates with Spark to read data, detailing the Spark‑SQL planning stages, the Spark‑Hudi read workflow, and the four main Hudi query types—snapshot, read‑optimized, time‑travel, and incremental—along with example SQL commands and code snippets.
This article presents a comprehensive technical overview of using ClickHouse as a native backend to accelerate Spark SQL execution, covering Spark performance bottlenecks, ClickHouse's CPU‑level optimizations, the design and implementation of the Spark‑Native integration, and detailed TPC‑DS benchmark results demonstrating up to 3.5× speedup.
The paper presents a Spark acceleration framework that replaces Java‑based task operators with a ClickHouse native library, converting plans via Protobuf and JNI, leveraging columnar storage, SIMD and JIT to achieve up to 3× speed‑up on TPC‑DS workloads while providing fallback mechanisms to ensure no performance loss.
This article presents a detailed technical overview of accelerating Spark's compute engine using ClickHouse as a native backend, covering Spark performance background, ClickHouse's advantages, the design and implementation of a Spark‑Native acceleration solution, and extensive performance evaluation results.
This article details Zhihu's migration of massive Spark and MapReduce shuffle workloads from the External Shuffle Service (ESS) to a push‑based Remote Shuffle Service (RSS) powered by Celeborn, covering background problems, evaluation of open‑source implementations, deployment architecture, encountered issues, solutions, performance gains, and future plans.
This article details Meituan's production‑grade adoption of Spark vectorized execution via the open‑source Gluten and Velox stack, explaining SIMD fundamentals, performance motivations, the end‑to‑end integration workflow, staged rollout, encountered challenges, and the resulting resource savings and speedups.
