By replacing part of its Spark‑based offline pipeline with StarRocks, Xiaohongshu’s data‑warehouse team cut job execution from hours to minutes, reduced resource usage over 90 %, lowered back‑fill cost by 99 %, and accelerated daily data production by 1.5 hours.
This article details Xiaomi’s practical deployment of Trino, covering its architectural role, core and extended capabilities, performance comparisons, integration with Iceberg and Spark, operational enhancements, multi‑cluster and ad‑hoc query scenarios, future cloud‑storage plans, and a Q&A session.
Baidu’s Fusion Compute Engine, built on Spark with a one‑layer wide‑table model, combines data‑skipping, push‑down, code‑generation, vectorization and extensive tuning to cut ad‑hoc query latency to seconds, shrink storage by ~30 %, and accelerate ETL workloads while maintaining stability for massive data‑warehouse workloads.
This article analyzes Baidu's fusion compute engine for its data warehouse, detailing its architecture, optimization techniques such as data skipping, Parquet column indexing, ProjectLimit and CodeGen, and demonstrates how these innovations reduce query latency to seconds while cutting storage costs by about 30% on multi‑petabyte workloads.
This article details Bilibili's tag system from its 2021 inception through successive redesigns, describing the three‑layer architecture, data flow pipelines using Hive, Iceberg, Spark and ClickHouse, crowd selection DSL, online services with Redis, performance optimizations, and upcoming governance and quality initiatives.
Meituan’s Fulfillment Platform team built a high‑performance distributed causal‑forest framework—named Causal On Spark—that trains hundreds of trees on hundreds of millions of samples within minutes using MapReduce‑based histogram splitting, extensive memory optimizations, Parquet model serving, and novel distributed evaluation metrics, enabling scalable causal inference for pricing, subsidies, and marketing.
This article presents Xiaomi's sales data warehouse practice, detailing its evolution, positioning, dimensional modeling, layered architecture, Lambda design, Iceberg integration, capability building, security governance, and future directions toward data value and real‑time metrics.
The article describes how GF Securities designed and implemented a unified big‑data empowerment layer based on Apache Kyuubi to address data‑centric challenges, improve efficiency, ensure controllable governance, and support agile data scenarios across ingestion, processing, storage, and security.
This article explains the limitations of using JDBC for true streaming reads in Spark and Flink, demonstrates failed attempts with MySQL, shows workarounds that revert to batch processing, and recommends Flink CDC as the practical solution for incremental MySQL ingestion.
Apache Kyuubi 1.8 introduces a range of enhancements including multi‑tenant serverless SQL support on Spark and Flink, expanded batch and streaming capabilities, improved resource scheduling with database‑backed queues, stronger Kerberos/LDAP security, Flink YARN integration, and a new web UI for management.
Facing massive daily Spark shuffle volumes and unstable ETL performance, ZTO Express migrated from the community External Shuffle Service to Celeborn's Remote Shuffle Service, achieving higher disk I/O efficiency, better reliability, reduced network connections, and significant reductions in task failures and job latency.
This article introduces Apache Celeborn, explains the challenges of intermediate data in large‑scale compute engines, details its core architecture and design—including master, worker, lifecycle manager and shuffle client—covers its community history, version releases, performance comparisons with Spark ESS, real‑world deployment scenarios, and outlines future development plans.
The 2023 Yunqi Conference in Hangzhou showcased the latest advances in cloud computing and big‑data technologies, examined the evolution from big‑data 1.0 to 3.0, discussed the key difficulties of making big data cloud‑native, and presented a practical case study of MiHoYo’s cloud‑native transformation.
This article presents a comprehensive technical overview of how DXY's big data platform leverages Apache Kyuubi and Celeborn to unify Spark entry points, configure flexible task isolation, implement fine‑grained AuthZ, optimize small files and Z‑Order sorting, and accelerate large result set transmission with Arrow, while also discussing operational challenges and upcoming features.
This article describes the background, problems, and a series of architectural upgrades—including MQ replacement, thread‑pool isolation, Redis/TiKV redundancy, and Spark‑based compensation—to enhance the stability, scalability, and high‑availability of an advertising billing system.
This article details ByteDance's extensive migration of Spark Shuffle to a cloud‑native architecture, describing the massive data volumes, the underlying ESS and CSS services, the challenges of resource isolation, monitoring, throttling, spill‑splitting, and the performance gains achieved across stable and mixed‑resource clusters.
MiHoYo’s data platform team details their migration of Spark workloads to Alibaba Cloud’s ACK Kubernetes service, describing how the Spark‑on‑K8s + OSS‑HDFS architecture delivers elastic compute, up to 50% cost reduction, and true compute‑storage separation, while addressing operational challenges through custom operators, Celeborn, and robust monitoring.
This article offers a concise MVP learning plan for recent graduates aiming to secure a big‑data internship, covering essential computer fundamentals, core big‑data frameworks, project ideas, and algorithm/SQL practice, along with practical study tips and resource recommendations.
Facing rapid growth in offline data and compute demands, we migrated our big‑data platform to a cloud‑native architecture using Spark 3.2.3 on Kubernetes with OSS‑HDFS storage, achieving elastic scaling, cost reduction, and compute‑storage separation while detailing implementation, challenges, and operational insights.
This article presents a systematic performance comparison between Doris and ClickHouse, covering data ingestion speed, SQL syntax differences, hardware impact, and detailed query benchmarks across multiple scenarios, ultimately revealing that each system excels in different use cases.
This article provides a comprehensive technical overview of LakeSoul, an open‑source, cloud‑native lakehouse framework, covering its design philosophy, core features, architecture, performance benchmarks, real‑time ingestion, incremental computation, multi‑stream joining, security, community progress, and future roadmap.
This presentation introduces Iceberg's core capabilities, details Xiaomi's practical applications—including log ingestion, near‑real‑time warehousing, offline challenges, column‑level encryption, and Hive migration—and outlines future development directions such as materialized views and cloud migration, providing a comprehensive view of modern data‑lake engineering.
The article details the design and implementation of a content revenue settlement platform that aggregates traffic and ad data, uses a Spark‑plus‑PALO architecture for processing tens of millions of daily records, and employs a master‑worker model with idempotent tasks, temporary tables, and verification steps to ensure reliable monthly profit‑share calculations for authors, media, mini‑program owners, and users.
This article explains how NetEase Yanxuan upgraded its legacy Lambda architecture to an Iceberg‑based batch‑stream unified platform, detailing the original data pipeline, the challenges faced, the evaluation of Iceberg versus Hudi and DeltaLake, and the concrete engineering optimizations and governance measures implemented to achieve lower latency and higher query performance.
The Cloud Vision TV app implements a session‑id and placement‑id driven tracking pipeline that generates, collects, and processes lifecycle data across server and client layers, enabling fine‑grained engagement strategies, scene reconstruction via AC automata, and actionable BI dashboards to improve user retention and personalization.
The WeChat Experiment Platform migrated its 60,000 metric, 200,000 core, 30 PB plus data pipeline to an Iceberg based lakehouse, leveraging three layer metadata, fine grained partitioning, MERGE into writes, time travel snapshots and skew handling UDFs, which cut core time by 69%, saved ~100 PB storage, and reduced latency by up to 70%.
This guide explains the cloud‑native lakehouse concept, outlines its advantages and challenges, compares lake‑table projects such as Iceberg, and provides a step‑by‑step AWS deployment of Apache Iceberg and Amoro—including environment setup, AMS installation, catalog configuration, optimizer launch, data ingestion with Flink, and query verification with Spark.
This comprehensive tutorial details the process of deploying a complete Hadoop-based big data ecosystem on ARM architecture, covering the installation and configuration of essential components including Java, Zookeeper, Hadoop, MySQL, Hive, and Spark with practical code examples.
This article presents a comprehensive technical overview of Alibaba Cloud AnalyticDB MySQL and its Serverless Spark integration, detailing architecture, core optimizations, security enhancements, and real‑world case studies that demonstrate how to achieve cost‑effective, high‑performance data lake analytics.
Since 2018 Youzan has migrated all online services to Kylin4, addressing long cube rebuilds, single‑point cache, CPU spikes, and throttling gaps by adding batch segment builds, low‑priority concurrency controls, Redis‑based query caching, parquet skew mitigation, range‑query acceleration, and class‑loader optimizations, which together doubled query‑per‑second capacity to 150, cut latency by up to 50 % and reduced CPU usage.
Guandata’s R&D leader outlines how their analytics platform leverages Delta Lake and Spark to deliver fast, ACID‑compliant BI and AI workloads, detailing architecture, key features like schema evolution and time travel, and practical performance tricks such as compaction, vacuuming, and multi‑engine integration.
To address the resource‑intensive challenges of joining billion‑row tables in data warehouses, this article examines common optimization approaches, analyzes Spark’s SortMergeJoin algorithm, and proposes a Bloom‑filter‑based solution that filters unchanged data early, dramatically improving performance and reducing cluster resource consumption.
This article summarizes a technical presentation on how Delta Lake is integrated into a BI+AI platform, covering the product background, data‑lake architecture, Delta Lake features such as ACID transactions, schema management, multi‑engine support, performance optimizations, and future development directions.
The article presents a detailed three‑month interview journey that led to a position at a top new‑energy automotive firm, outlining the questions and topics covered in five interview rounds—including Hive, Spark, Flink, Kafka, data modeling, and data governance—to help candidates prepare for big‑data roles.
This presentation details Iceberg's core capabilities—transactional writes, schema evolution, implicit partitioning, and row‑level updates—while showcasing Xiaomi's real‑world applications such as log ingestion redesign, near‑real‑time warehousing, offline optimizations, column‑level encryption, Hive migration strategies, and outlining upcoming enhancements like materialized views and cloud migration.
This article demonstrates how developers familiar with Java Stream APIs can quickly grasp fundamental Spark operations—including map, flatMap, groupBy, and reduce—by translating stream examples into Spark code, providing complete code snippets, explanations of transformations versus actions, and practical tips for handling exceptions in distributed processing.
This article explains the concept of data skew in offline big‑data jobs, describes its symptoms and root causes, and provides practical optimization techniques for Hive and Spark—including partitioning strategies, map‑join usage, adaptive query settings, and monitoring approaches—to prevent performance degradation and runtime failures.
This article details iQIYI's adoption of Iceberg for its data lake, covering the OLAP architecture, reasons for a data lake, Iceberg's table format advantages over Hive, platform construction, streaming ingestion, query and performance optimizations, real‑world business deployments, and future plans.
This article explains how Tencent Video leverages causal inference techniques—including uplift gain models, propensity‑score‑matching (PSM), and a distributed Spark‑based CausalML library—to identify incremental user effects, evaluate marketing interventions, and improve growth across advertising, internal flow, push notifications, and coupon strategies.
The article shares a fresh graduate's experience interviewing for offline data‑warehouse roles at companies like Ctrip, Meituan and Alibaba, outlines the common interview pattern, and lists detailed Spark, Hadoop, and data‑warehouse questions used by these firms.
This article presents a comprehensive case study of NetEase Cloud Music's data governance practice, covering data background, governance philosophy, detailed solutions across metadata, storage, compute, and model design, practical implementations, measurable cost savings, and future planning for sustainable data management.
This article explains how Apache Spark can be deployed using the traditional Hadoop YARN resource manager and the newer Kubernetes approach, detailing configuration steps, submission methods, and a comprehensive comparison of isolation, scalability, learning curve, logging, performance, and cost considerations.
The article outlines how a unified, big‑data‑based user profiling platform addresses traditional data silos, high costs, and limited functionality by standardizing tags, integrating Spark and RHadoop processing, and enabling a closed‑loop marketing workflow that improves accuracy and operational efficiency.
This article introduces Apache Hudi as a next‑generation streaming data‑lake platform, explains its core concepts, architecture, and table types, and showcases real‑world use cases at Tencent such as CDC ingestion, minute‑level real‑time warehousing, streaming analytics, multi‑stream joins, ad attribution, and stream‑to‑batch processing, while also outlining future directions.
This article presents Blaze, a Kuaishou‑built native execution middleware for SparkSQL that leverages Apache DataFusion to achieve vectorized operator execution, detailing its architecture, implementation, performance gains, current coverage, benchmark results, production rollout, and future development plans.
This article shares a recent graduate's successful job offer story, emphasizes preparing a detailed personal project summary, and provides extensive big‑data interview questions covering Hadoop, Spark, Flink, Kafka, Hive, ClickHouse, and related technologies to help candidates excel in interviews.
This article explains why tiny Hive files degrade storage and query efficiency, outlines how they are created, and presents practical Spark and Hive configuration techniques—including dynamic partitioning, AQE, Reduce tuning, and automated daily merge jobs—to effectively consolidate small files and improve overall data‑warehouse performance.
This article reviews Alibaba's Tongyi Qianwen large language model, shares personal impressions, provides a fish‑flavored pork recipe, conducts a SWOT analysis, and includes Scala Spark and Java code examples illustrating its capabilities and usage scenarios.
This article examines the causes and impacts of small file proliferation in Hive and Spark environments, and presents multiple mitigation techniques—including Spark 3 adaptive query execution, reducing reduce tasks, using DISTRIBUTE BY RAND(), post‑processing clean‑up, Hive and Spark configuration tweaks, and automated tooling—to improve performance and storage efficiency.
Shuttle2.0 extends OPPO's open‑source high‑availability Spark Remote Shuffle Service to support Flink, introduces a unified stream‑batch data model, pipelines shuffle with distributed sorting, and provides an Adaptive BroadcastJoin solution that dramatically improves performance and stability for large‑scale big‑data workloads.
The article reviews Alibaba’s Tongyi Qianwen large‑language model, shares a cooking recipe generated by the AI, presents a SWOT analysis, and provides code examples—including a Spark Scala script for same‑period month‑over‑month calculations and a Java bubble‑sort implementation.
This article presents Huawei's end‑to‑end solution for constructing a real‑time data lake on Hudi, covering requirement analysis, technology selection, architectural design, ingestion and processing challenges, practical optimizations, and future improvement directions.
This article provides a comprehensive guide on optimizing Spark and Flink workloads, covering parallelism settings, garbage‑collection tuning, out‑of‑memory mitigation, and full production‑grade configuration examples for both frameworks.
The article describes a data application platform that enables business users to configure custom data selection rules for opportunities, create scheduled tasks, perform large‑scale data comparison, handle task dispatch with Redis queues, and implement rate‑limiting using sliding windows to ensure reliable processing.
This article explains how to perform task‑level data cost governance by collecting storage and compute metrics from Hive tables, Spark jobs, and HDFS FsImage files, then estimating monthly expenses using replication factors and resource‑usage rates, while providing practical SQL and shell examples.
This article explains how to query, analyze, and visualize time‑series data stored in MongoDB using the aggregation framework, MongoDB Compass, read‑only views, the BI connector with SQL tools, as well as integrations with Spark and R for advanced analytics.
This article explains the concept of data skew, its symptoms such as slow tasks and OOM errors, and provides comprehensive mitigation techniques and configuration examples for Hive and Spark, including custom partitioning, map joins, adaptive execution, and key detection methods.
Facing daily terabyte‑scale data ingestion and costly duplicate reads in its ODS‑to‑DWD pipeline, Bilibili introduced a Flink‑based real‑time incremental computation and multi‑level partition shuffling, dramatically reducing read amplification, cutting resource usage by ~20%, improving latency to minutes, and enhancing scalability.
Compass is an open‑source big‑data diagnostic platform developed by OPPO that provides non‑intrusive, real‑time monitoring and root‑cause analysis for offline and streaming tasks on schedulers such as DolphinScheduler and Airflow, covering workflow‑level failures, Spark engine anomalies, resource usage, and offering one‑click reports and extensible rule‑based diagnostics.
The article introduces MindAlpha, a high‑performance distributed machine‑learning platform built for large‑scale, sparse ad‑tech workloads, detailing its architecture, MLOps pipeline, Spark integration, sync/async training strategies, CPU/GPU choices, model‑splitting techniques, and future directions such as model pruning and AutoML.
The article introduces the Gluten vectorized engine, explains why Spark’s CPU bottleneck motivates integrating native vectorized back‑ends via Substrait, details its architecture, component design, current performance gains of up to three‑fold, and outlines ongoing development and future work.
This article walks through practical steps for governing compute resources in a data warehouse, covering problem identification, strategic thinking, Spark and Hive tuning, small‑file handling, DQC improvement, high‑consumption task optimization, scheduling adjustments, and measurable performance gains.
Apache Kyuubi 1.6.0 introduces major server‑side upgrades such as batch JAR task submission with RESTful APIs and a metadata store for HA, client‑side improvements including a unified JDBC driver and enhanced Beeline, plus mature Spark, Flink, Trino, and Hive engine plugins, while outlining the community’s roadmap.
The article provides a comprehensive walkthrough of Apache Kyuubi 1.6.0, detailing server‑side enhancements such as batch (JAR) task submission, metadata store and unified API/authentication, client‑side improvements to the built‑in JDBC driver and Beeline, as well as engine plugins for Spark, Flink, Trino and Hive, and concludes with the community’s roadmap and statistics.
This article presents Datacake's experience of integrating AI algorithms into big data governance, covering the bidirectional relationship between AI and big data, health‑score assessment of data tasks, intelligent Spark parameter tuning, SQL engine selection, and future application scenarios across the data lifecycle.
This article outlines Alibaba Cloud EMR's three‑stage evolution—compatibility, contribution, and beyond open source—detailing its early Hadoop adoption, Flink and Spark innovations, cloud‑native optimizations, and enterprise‑grade features such as Remote Shuffle Service, performance benchmarks, and integrated diagnostics.
We Analysis, the official data‑analysis platform for WeChat mini‑program providers, delivers a zero‑learning‑curve user‑profiling system that combines basic tag analysis and flexible, rule‑based segmentation, using an ETL pipeline to store pre‑computed data in TDSQL and online bitmap‑optimized queries in ClickHouse with RoaringBitmap, ensuring low‑latency, stable, and comprehensive analytics.
This article introduces the Iceberg table format, explains its core architecture and advantages such as transactionality, implicit partitioning and row‑level updates, details Xiaomi's practical deployments—including CDC pipelines, partition strategies, compaction services, and stream‑batch integration—and outlines future development directions.
With data volumes soaring, traditional Hadoop and Spark clusters become costly and cumbersome for small to medium workloads, prompting many to seek lighter alternatives; this article examines the technical, operational, and financial burdens of Hadoop/Spark and introduces the open‑source SPL engine as a fast, low‑cost, easy‑to‑use big‑data solution.
This comprehensive tutorial walks you through setting up a three‑node Hadoop 3.3.4 and Spark 3.3.1 environment on CentOS 7 virtual machines, covering system preparation, JDK and Scala installation, Zookeeper configuration, Hadoop and Spark deployment, and verification with practical command‑line examples.
This article examines the evolution, architectural differences, data and processing models, stateful handling, and programming APIs of Apache Spark and Apache Flink, highlighting their strengths, limitations, and the challenges of big‑data development and operations in the modern data‑driven era.
This article presents a curated collection of big‑data learning resources, including interview guides, in‑depth articles on Flink, Spark, Hive, ClickHouse, data governance, and personal growth, offering readers a one‑stop reference to boost their big‑data expertise and interview readiness.
This article presents a curated collection of the most comprehensive big‑data interview preparation resources, including expert guides, tutorials, and deep‑dive articles on Flink, Spark, Hive, ClickHouse, data governance, and related topics, accompanied by a call to engage with the content.
Distributed computing splits massive tasks across multiple servers, and this article explains the classic MapReduce batch engine and the modern Spark framework, covering their architectures, strengths, limitations, and evolution, while highlighting key features like fault tolerance, in‑memory processing, and real‑time streaming capabilities.
Facing massive, multi‑source traffic and the need for instant analytics, Kuaigou’s real‑time data warehouse evolved from Spark on‑premise to a cloud‑native stack using Alibaba Blink, Flink, and layered OLAP models, streamlining development, cutting costs, and enabling diverse real‑time applications.
This article explains how ByteDance designed and deployed a real‑time data warehouse on a data lake using Hudi, detailing three business scenarios, the challenges of latency, consistency and resource usage, and the engineering solutions—including upserts, compaction services, indexing, and future unified storage plans.
This article explains the shift of big data and machine‑learning workloads toward storage‑compute separation and cloud‑native architectures, outlines the technical challenges of running Spark on Kubernetes, and details the EMR on ACK solution with its architecture, performance gains, and real‑world adoption.
This article explains the design and implementation of WeChat's WeAnalysis image system, covering its basic tag and user‑group modules, multi‑source data ingestion, ETL processing, storage choices such as TDSQL and ClickHouse, bitmap handling, query performance, and service APIs for flexible, high‑performance user segmentation.
This article reviews the history of data platforms—from the first general‑purpose computers and early relational databases through traditional BI, agile BI, and big‑data technologies like Hadoop, Spark, and Flink, up to today’s cloud‑native modern data stack and its future outlook.
This article presents the background, requirements, architecture, key modules, and practical impact of OPPO's non‑intrusive big‑data diagnostic platform—named Compass—designed to quickly locate issues, provide optimization suggestions, and achieve cost‑saving and efficiency gains for large‑scale Spark and Hadoop workloads.
The article details how Tencent Advertising re‑architected its massive log pipeline by consolidating heterogeneous real‑time and offline logs into an Iceberg‑based data lake, introducing multi‑level partitioning, Spark and Flink ingestion, and numerous performance and cost optimizations for scalable big‑data analytics.
This article explains Spark SQL's CacheManager, how it stores cached query results using InMemoryRelation, the ways to trigger and release caches, the internal data structures like IndexedSeq and CachedData, and the role of canonicalization in determining cache reuse.
This article examines a recurring Spark batch task failure caused by OutOfMemory errors and data skew, details the investigation steps—including increasing executor memory, raising parallelism, and analyzing shuffle metrics—and proposes solutions such as data validation, filtering oversized keys, and memory adjustments.
This article presents a comprehensive overview of TikTok e-commerce's near‑real‑time data lake implementation, detailing data lake characteristics, architecture choices, practical use cases across analysis and operations, and for future challenges and plans.
This article investigates why Spark tasks that appear killed in the Web UI continue running on the driver, analyzes the role of daemon versus non‑daemon threads and SparkContext shutdown mechanisms, reproduces the issue with sample code, and provides practical solutions such as using daemon threads or checking SparkContext status.
Occlum v1.0, the open‑source trusted execution environment operating system released by Ant Group, delivers up to five‑fold performance improvements, supports over 150 Linux syscalls, introduces async I/O, dynamic memory management, and a Spark‑BigDL big‑data analysis solution, while outlining future GPU and TDX extensions.
This article analyzes a midnight HDFS DataNode failure caused by excessive GC and OOM due to Spark batch jobs, examines how an unexpected surge in block count overloaded default memory settings, and presents concrete remediation steps and optimization recommendations to stabilize the cluster.
This guide explains common business scenarios that require custom SQL parsing, walks through setting up Antlr4 in IntelliJ IDEA, configuring Maven dependencies, generating parser code, and provides Java examples for extracting table names from Spark SQL statements, including handling of prediction modes and execution results.
NVIDIA’s security team adopted the formally verified SPARK language, replacing C in safety‑critical components, and after a successful proof‑of‑concept demonstrated improved security, verification efficiency, and unchanged performance, leading to widespread internal adoption across many products.
ByteDance’s cloud‑native computing team, led by Li Yakun, details how they transformed a Hadoop‑centric big‑data stack into a Kubernetes‑driven platform—customizing storage, middleware, and scheduling—to support petabyte‑scale workloads, achieve over 40% resource utilization, and sustain rapid product growth.
This article introduces Iceberg’s table‑format fundamentals, details Xiaomi’s large‑scale deployment of Iceberg for CDC and log ingestion, explores their streaming‑batch integration experiments, outlines future roadmap items, and provides a comprehensive Q&A covering practical challenges and solutions.
A high‑CPU anomaly on a Spark‑enabled machine was traced through application checks, network TIME_WAIT analysis, and Airflow inspection, leading to kernel tweaks and an Airflow configuration change that finally restored normal CPU usage.
The article details how Meituan’s Turing machine‑learning platform cut offline resource use by 80% and task time by 63% through memory‑level techniques such as column pruning, adaptive caching, and a deep dive into Spark’s mapPartitions operator, including source‑code analysis, GC behavior, and a low‑memory batch‑iterator best practice.
This article walks through a Spark job’s garbage‑collection tuning workflow, presents step‑by‑step experiments with different JVM options and collectors, compares performance under tight and normal memory conditions, and offers practical recommendations for choosing the optimal GC strategy in big‑data workloads.
This article explains the importance of layering in data warehouse modeling, outlines the four ETL steps, describes common pitfalls, presents a typical technical stack, and details each warehouse layer (ODS, DWD, DWS, ADS) along with best‑practice naming conventions and implementation tips for big‑data environments.
This article examines the challenges of traditional Spark clusters and explains how integrating Spark with cloud‑native platforms—through Kubernetes deployment modes, EMR on ACK practices, Remote Shuffle Service, and serverless Spark on DLF—provides elastic scaling, lower operational costs, and advanced features such as executor rolling and custom scheduler support.
This article presents an in‑depth overview of Apache Iceberg as used at Tencent, covering its table format architecture, Spark read/write mechanisms, production challenges and optimizations such as schema evolution, file filtering, upsert strategies, and the surrounding data‑governance services.
This article examines the challenges of using Apache Hudi for real‑time data lake writes, analyzes the COW and MOR write models, and presents a custom master‑worker architecture with index optimization and repartitioning that reduces write resource consumption by over 85% while boosting throughput up to 300‑fold.
Bilibili adopted the open‑source Kyuubi proxy to replace its unstable STS layer, enabling multi‑tenant, multi‑engine (Spark, Presto, Flink) SQL/Scala processing with Hive Thrift compatibility, fine‑grained queue isolation, UI monitoring, stability safeguards, and Kubernetes/YARN deployment, while planning further cloud‑native extensions.
Hulu’s data platform team describes how they migrated large‑scale Spark workloads from Yarn to native Spark on Kubernetes, leveraging AWS services such as EKS, S3, and custom operators to achieve dynamic scaling, unified monitoring, cost‑effective resource management, and improved stability for search, recommendation, and advertising pipelines.
