This article explains JD’s self‑developed data‑skew mitigation solution for Spark, detailing the problem of uneven key distribution, the limitations of the open‑source AQE implementation, and JD’s OptimizeSkewedJoinV2 algorithm that dramatically reduces stage latency in large‑scale join workloads.
This article introduces Kyuubi, the first NetEase project contributed to the Apache Foundation, describing its core features, multi‑tenant architecture, Spark‑based execution engine, cloud‑native capabilities, and real‑world use cases within NetEase’s data‑warehouse, ad‑hoc, and internal systems, along with performance gains and community resources.
This article provides an in‑depth guide to Spark performance optimization, covering the ten development principles, static and unified memory models, resource parameter tuning, data skew detection and mitigation techniques, as well as shuffle‑related configuration adjustments, supplemented with practical code examples and diagrams.
SeaTunnel, the China‑originated data‑integration platform built on Spark and Flink, has been accepted into the Apache Incubator, and this article introduces its history, architecture, plugin ecosystem, deployment requirements, and numerous enterprise deployments across batch and streaming big‑data scenarios.
This article presents a detailed interview-style walkthrough covering Hadoop cluster setup, HDFS components, MapReduce workflow, YARN advantages, Spark fundamentals, Kafka replication, Hive table types, and related big‑data concepts, providing concise explanations and practical insights for data engineers.
This article introduces the Project Matrix initiative that re‑examines and restructures Spark‑ML linear models, detailing the background of Spark‑ML usage at JD, the performance‑focused optimizations such as blockification and virtual centering, and outlines upcoming work to further improve scalability and accuracy.
This article presents a comprehensive list of essential Spark configuration settings—including executor memory, off‑heap memory, memory fractions, shuffle options, and adaptive query execution parameters—each accompanied by a concise description to help users fine‑tune Spark performance.
This article explains the two most important Spark 3.0 features—Adaptive Query Execution and Dynamic Partition Pruning—detailing how AQE dynamically optimizes join strategies, partition coalescing, and skew handling, while DPP reduces I/O by pruning irrelevant fact‑table partitions at runtime.
This article explains how Spark SQL's Catalyst optimizer performs logical and physical planning, details the Tungsten engine's data‑structure and whole‑stage code generation improvements, compares them with the Volcano iterator model, and provides code examples and PDF resources for deeper study.
This article explains Spark SQL join classifications, the mechanics of Nested Loop Join, Sort‑Merge Join, and Hash Join, and describes how the Catalyst optimizer and Tungsten project improve query execution and memory efficiency in large‑scale data environments.
This article details JD Retail's strategic "Nirvana" product‑selection platform, describing the technical challenges of handling billions of items and hundreds of tags, and presenting a dual‑engine solution using ClickHouse and Elasticsearch with Spark‑driven data pipelines to achieve fast filtering, multidimensional analytics, and efficient storage.
This article explains the origins of Spark DataFrames, compares them with RDDs, describes how Spark SQL optimizes DataFrame execution, and provides detailed examples of creating DataFrames from RDDs, files, and JDBC sources along with common DataFrame operations and code snippets.
This article explains Spark’s unified memory management model, detailing the division between on‑heap and off‑heap memory, the roles of execution, storage, user, and reserved memory, configuration parameters, dynamic allocation, and how these concepts affect performance and resource utilization.
This article explains Spark's storage architecture, detailing the roles and interactions of BlockManager, MemoryStore, and DiskStore, including their initialization, data management mechanisms, code implementations, and eviction strategies, to help readers grasp how Spark efficiently handles in‑memory and on‑disk data.
This article provides a comprehensive overview of Spark's shuffle process, explaining its definition, internal mechanisms such as shuffle write and read, the evolution of shuffle managers, and practical optimization techniques including parameter tuning and broadcast variables, all aimed at improving performance in large‑scale data processing.
This article introduces Spark's core concepts, explains the RDD abstraction and its four main properties, and details the roles of DAGScheduler, SchedulerBackend, TaskScheduler, and ExecutorBackend, providing practical insights for beginners and intermediate users in big‑data processing.
This article compiles a comprehensive list of Chinese-language articles and tutorials covering big‑data technologies such as Flink, Spark, Hive, ClickHouse, data governance, and related interview preparation resources, providing a structured learning path for aspiring data engineers.
This article explains the concept of Iceberg catalogs, compares HiveCatalog and HadoopCatalog, and provides step‑by‑step Spark examples for downloading the Iceberg jar, creating tables, loading data, querying, and examining the underlying metadata and directory structures.
This article presents a detailed learning roadmap for aspiring big‑data experts, covering foundational programming languages, data structures, Linux basics, databases, distributed system theory, and essential frameworks such as Hadoop, Spark, Flink, Kafka, and provides curated B‑site video links and reference materials.
This article provides a comprehensive guide to designing and implementing a big‑data platform, covering architecture overview, data ingestion with Flume, storage on HDFS/Hive/HBase, processing engines such as Hive, Spark and Flink, scheduling solutions like Azkaban and Airflow, and the construction of self‑service analytics systems.
This article explains how Apache Spark 3.0 improves SQL workload performance through Adaptive Query Execution (AQE) and Dynamic Partition Pruning (DPP), detailing their design principles, runtime optimizations, configuration parameters, and practical examples that demonstrate reduced shuffle partitions, smarter join strategies, and handling of data skew.
This article explains Spark's memory model, identifies common driver and executor OOM scenarios, and provides detailed configuration and code‑level recommendations—including memory‑related parameters and shuffle‑tuning options—to prevent and resolve out‑of‑memory failures in Spark applications.
This article examines the strengths and weaknesses of Apache Iceberg, explains why Tencent selected it over alternatives, details Tencent’s own enhancements and integration with Flink, Spark, and other engines, and shares multiple real‑world implementations for building enterprise‑grade real‑time data lakes.
This article provides a comprehensive overview of Hive's execution engines—including MapReduce, Tez, and Spark—detailing their architectures, the six-stage Hive SQL compilation process, practical Explain syntax examples, and extensive tuning parameters for each engine to improve performance in big‑data environments.
The article describes how Yixin Group’s product team created an in‑house intelligent data warehouse using Hadoop, Flink/Spark, and standardized data services to transform scattered automotive‑finance data into a secure, scalable platform that supports real‑time analytics and drives business growth.
The "God of Big Data" project, launched in 2019, offers a detailed learning roadmap, systematic column resources covering Hadoop, Spark, Kafka, and more, and invites engineers transitioning from backend to big‑data development to follow curated articles, GitHub code, and CSDN tutorials.
This article breaks down a typical big data platform architecture into its four layers—data collection, storage and analysis, sharing, and real‑time computation—detailing the essential tools such as Flume, HDFS, Hive, Spark, DataX, and task scheduling systems that enable scalable, low‑latency data processing and delivery.
This article explains the typical architecture of a big‑data platform, detailing its four core layers—data collection, storage & analysis, data sharing, and application—and describing the key technologies such as Flume, DataX, HDFS, Hive, Spark, Spark Streaming, and task scheduling components.
This guide explains how to deploy Apache Kyuubi on a CDH 6 cluster, replace HiveServer2 with Kyuubi, integrate Spark 3, apply necessary patches, configure environment and Spark settings, and optimize engine sharing for various workloads, providing complete code snippets and step‑by‑step instructions.
Didi's data architecture team migrated high‑volume, real‑time funnel analysis from ClickHouse to StarRocks, built a multi‑layer pipeline with Kafka, Flink/Spark, Hive, and materialized views, and achieved sub‑3‑second query times on billions of rows, while outlining future enhancements.
This article explains how a 10 GB gzip file is stored and processed on HDFS, details the MapReduce split calculation using GzipCodec, and discusses why Spark reads such non‑splittable files with a single task, recommending file splitting or format conversion for better performance.
This article explains how bytecode and code‑generation improve Spark SQL performance, compares Java I/O and MapReduce InputFormats, reviews serialization choices in Spark and Flink, and describes reflection‑based DataFrame creation, storage‑memory eviction, fail‑fast design, and ConcurrentHashMap usage in big‑data frameworks.
This article describes how Ctrip built a data lineage system for its big data platform, covering the concept of data lineage, collection methods, open‑source tools such as Apache Atlas and DataHub, the in‑house table‑level and field‑level solutions, implementation details for Hive, Spark and Presto, storage in JanusGraph, and practical applications in data governance, metadata management, scheduling and sensitivity labeling.
This article outlines a typical big data platform architecture, detailing its core layers—data acquisition, storage and analysis, sharing, application, real‑time computation, and task scheduling—while introducing key technologies such as Flume, HDFS, Hive, Spark, DataX, and monitoring considerations.
The article outlines a typical big data platform architecture, detailing its core layers—data collection, storage and analysis, sharing, application, real-time computation, and task scheduling—while describing key technologies such as Flume, DataX, HDFS, Hive, Spark, Spark Streaming, and Redis.
This article provides a thorough overview of Apache Spark, covering its origins, comparison with MapReduce, core concepts such as RDD, DAG, Jobs, Stages, and Tasks, the submission process, Web UI, and detailed performance tuning techniques including data skew mitigation.
This article records an interview experience, presenting core questions on Java memory allocation, JVM parameters, Spark and MapReduce execution models, data skew causes and mitigation, real‑time framework scheduling, and system design for massive task scheduling, followed by analysis and learning recommendations.
This article presents a detailed collection of Spark performance‑tuning techniques—including submit‑script parameters, RDD and operator optimizations, parallelism and memory settings, broadcast variables, Kryo serialization, locality wait adjustments—as well as systematic methods for detecting and resolving data skew and common runtime issues such as shuffle failures, serialization errors, and JVM memory problems.
This article explains the rapid growth of data, the limitations of traditional warehouses, and how a cloud‑based data lake built on object storage with Delta Lake format provides low‑cost, flexible, and ACID‑compliant analytics, followed by a step‑by‑step guide to ingest, manage, and analyze data using Alibaba Cloud DLF and Databricks DDI with Spark streaming and batch jobs.
This article compiles extensive interview questions and topics covering Hadoop, Spark, Flink, Hive, Kafka, MySQL, Redis, Java fundamentals, and algorithms, organized by companies such as Xiaomi, ByteDance, Alibaba, Shopee, Tencent, Meituan, NetEase, and Baidu, to help candidates prepare effectively for big‑data engineering roles.
This article introduces the most important and still mainstream components of the big data ecosystem—including Hadoop’s storage and compute framework, Hive data warehouse, HBase NoSQL database, Spark unified engine, Kafka messaging platform, and Elasticsearch search engine—explaining their core concepts, architectures, and typical use cases.
This article presents the Future Cloud Business Monitoring real-time dashboard solution, detailing its technical architecture, key challenges in massive log processing, storage choices, product considerations, experience sharing, future plans, and concrete case studies such as live classroom monitoring.
This article explains Hive's core components, execution architecture, how HiveQL is transformed into MapReduce jobs, the advantages of Tez over MapReduce in Hive 3.0+, and the integration of Spark with Hive for modern big‑data processing.
This article analyzes why Spark tasks fail with a "Task not serializable" exception when closures reference class members, demonstrates the issue with Scala code examples, and provides practical solutions such as using @transient annotations, moving functions to objects, and ensuring proper class serialization.
This extensive guide outlines a modern big‑data learning roadmap, covering essential programming languages, Linux, databases, distributed system theory, networking, offline and real‑time computation, message queues, data warehouses, algorithms, backend skills, interview preparation, and practical advice for building a personal knowledge system.
This article details the design and multi‑stage optimization of an A/B testing metric system, describing its product architecture, Spark‑based computation engine, ClickHouse OLAP layer, cumulative calculation improvements, and batch processing techniques that reduced processing time from hours to a few minutes for hundreds of experiments and metrics.
This article explains how to create a custom SparkListener in Apache Spark, provides Scala code examples for the listener and a main application, and details two registration approaches—via Spark configuration or SparkContext—along with a comprehensive list of listener event methods.
This article details Ctrip Finance's end‑to‑end data‑foundation architecture that uses MySQL binlog collection via Canal, Kafka streaming, Spark‑Streaming persistence to HDFS, and a merge process to produce timely MySQL‑Hive snapshots, addressing performance, consistency, and delete‑handling challenges.
This article provides an in‑depth overview of Apache Kylin, covering its history, mission, core MOLAP principles, technical architecture, step‑by‑step installation (Docker and Hadoop), performance tuning, advanced cube settings, and detailed case studies from major companies such as Baidu, Lianjia, and Didi.
This article reviews the advantages of Apache Iceberg for data lake storage, details Tencent’s custom optimizations and integration with Flink and Spark, and shares multiple real‑world implementations that demonstrate how Iceberg improves data consistency, reduces small‑file overhead, and enables near‑real‑time analytics in large‑scale big‑data environments.
This article explains how to efficiently move large volumes of data from Hive to HBase by leveraging HBase's bulkload mechanism, detailing the original MapReduce workflow, its performance bottlenecks, and a rewritten Spark‑based solution that simplifies ETL, improves partitioning, and achieves several‑fold speedup.
This article explains Apache Iceberg’s table‑format design, compares it with Hive’s limitations, details its snapshot‑based architecture and metadata handling, and describes how NetEase Cloud Music leveraged Iceberg to dramatically improve large‑scale log processing performance and stability.
This article provides a detailed collection of Spark interview questions covering deployment modes, performance advantages over MapReduce, shuffle mechanisms, RDD characteristics, optimization techniques, resource management, and various practical aspects of Spark on YARN, Mesos, and Kubernetes.
Didi migrated over 10,000 Hive SQL tasks to Spark SQL, achieving 85% Spark task share, cutting execution time by 40%, and reducing CPU and memory usage by 21% and 49% respectively, through a systematic migration process that addressed syntax, UDF, performance, and functional differences between the two engines.
This article shares practical experience of building a high‑performance distributed prediction service by combining TensorFlow for Java with Spark‑Scala, covering framework selection, performance comparison, model training, loading, inference, deployment, and optimization techniques for large‑scale data processing.
This article systematically introduces NetEase Kyuubi, an open‑source high‑performance JDBC and SQL execution engine built on Apache Spark, covering its background, core architecture, service discovery, session and operation management, startup processes, and key source‑code implementations with detailed code examples.
The article details the development, challenges, and redesign of JD.com’s self‑operated rebate system, describing its early monolithic architecture, data‑intensive processing pipeline, migration to a modular, high‑availability platform built on Spark, Hive, and Elasticsearch, and the resulting performance and operational improvements.
This article describes how Linkflow migrated mutable customer data from MySQL to an Apache Hudi data lake using Debezium‑in‑Flink CDC, addressing challenges such as snapshot resumability, partial updates, row‑key merging, schema evolution, indexing, and concurrent writes to achieve minute‑level data freshness and improved offline processing performance.
This article explains the motivations behind Apache Iceberg, its design principles such as snapshot and MVCC, compares it with Hive, and describes how NetEase Cloud Music adopted Iceberg to improve metadata handling, query performance, and operational stability for massive daily log data.
This article explains how Spark coordinates master, worker, driver, and executor components to generate, submit, and run jobs, detailing the creation of logical and physical execution graphs, task allocation, result handling, and the shuffle read process with code examples and diagrams.
This guide compiles practical Spark tuning techniques, covering essential configuration parameters, programming best‑practices, detailed shuffle mechanics, and join optimization strategies, while also addressing common errors and mitigation steps, enabling developers to improve performance and resource utilization in large‑scale data processing jobs.
This article explains how Spark implements shuffle write and shuffle read, compares its high‑level and low‑level processes with Hadoop MapReduce, and details the internal data structures, memory‑disk trade‑offs, and configuration options that affect performance.
This article explains how Spark transforms a logical execution graph into a physical job by partitioning stages, applying pipeline concepts, and generating tasks, while illustrating the process with detailed code examples and the internal workflow of job submission.
This article explains how Spark builds the logical execution graph for RDDs, describes the four-step job processing pipeline, details the various dependency types such as NarrowDependency and ShuffleDependency, and reviews common transformations and their data‑flow characteristics.
This article explains Spark's cache and checkpoint mechanisms, detailing when to use each, how they are implemented internally, how cached and checkpointed RDDs are stored and retrieved, and the differences between caching, persisting, and checkpointing for reliable big‑data processing.
To meet iQIYI video production’s thousands‑QPS, petabyte‑scale, frequently‑updated data and large‑table join requirements, the team built a Spark‑plus‑ClickHouse real‑time warehouse that streams Kafka changes, joins HBase dimensions, and writes to ClickHouse, reducing reporting development time from days to hours while supporting both offline and real‑time analytics.
This article explains advanced Spark performance tuning techniques, focusing on diagnosing and resolving data skew and shuffle bottlenecks through stage analysis, key distribution inspection, and a variety of practical solutions such as Hive pre‑processing, key filtering, parallelism increase, two‑stage aggregation, map‑join, and combined strategies, while also covering ShuffleManager internals and related configuration parameters.
This article provides a comprehensive guide to Spark performance optimization, covering development‑level tuning principles, resource configuration parameters, practical code examples, and best‑practice recommendations to achieve high‑throughput big‑data processing.
The article explains the limitations of static shuffle partitions, execution‑plan estimation, and data skew in Spark SQL, and describes how Spark Adaptive Execution can automatically adjust shuffle partition numbers, switch join strategies, and mitigate skew through configurable parameters and code examples.
This article describes how Soul's data engineering team replaced nightly batch ETL with real-time Delta Lake ingestion on EMR, detailing the motivations, comparative analysis of Delta, Hudi, Iceberg, the implementation architecture, encountered issues such as data skew and schema evolution, and the solutions adopted to improve performance and reliability.
This article examines why many developers favor Spark's mapPartitions over map, analyzes the underlying source code, highlights common pitfalls such as complexity and OOM risks, and provides practical guidelines and code examples for correctly using mapPartitions in both simple and advanced scenarios.
This guide presents a complete Spark performance optimization handbook covering development‑time best practices, resource‑parameter tuning, detailed data‑skew detection and mitigation techniques, advanced shuffle‑engine configurations, and practical code examples to help engineers build faster, more reliable Spark jobs.
This article describes how a data engineering team replaced nightly batch ETL with a Delta Lake‑based real‑time log ingestion pipeline on EMR, detailing the motivations, architecture, implementation steps, encountered issues such as data skew and schema evolution, and the practical solutions they applied to achieve low‑latency, reliable data delivery.
This article details how Youzan's offline Spark computing platform was transformed for the cloud‑native era by migrating from YARN to Kubernetes, introducing containerization, storage‑compute separation, dynamic allocation, deployment optimizations, and a collection of practical lessons to reduce cost and improve resource utilization.
Youzan migrated its large‑scale offline Spark workloads from YARN to a cloud‑native Kubernetes architecture, separating storage and compute with Ceph FS, adding dynamic executor allocation and remote shuffle services, and applying numerous Spark and deployment tweaks that yielded elastic scaling, higher resource utilization, reduced costs, and valuable operational lessons.
This article presents JD's self‑developed Remote Shuffle Service for Spark, detailing its architecture, goals, implementation details, performance benchmarks, and real‑world production case studies that demonstrate its impact on shuffle efficiency and system stability in large‑scale data processing.
This article provides a comprehensive overview of Apache Iceberg, covering its origins, key features, practical Spark and Flink code examples, notable deployments at Alibaba and Tencent, and its future role as a universal table format for big‑data analytics.
This article shares a non‑CS graduate's interview experience, study methods, and a detailed list of common interview questions covering Java fundamentals, data‑warehouse concepts, Spark, Kafka, Zookeeper, HBase, and Elasticsearch, along with personal reflections on advanced interview expectations.
This article describes JD's research and production deployment of a self‑developed Remote Shuffle Service for Spark, covering its motivations, architectural design, cloud‑native features, monitoring, performance benchmarks against external shuffle solutions, and a real‑world promotion‑period case study that demonstrates improved stability and resource efficiency.
This article reviews the evolution, architecture, and key components of major Chinese big‑data platforms—including those of Taobao, Didi, Meituan, 360, Kuaishou, and JD—highlighting data ingestion, storage, processing engines, scheduling systems, and service‑oriented designs that underpin their large‑scale data operations.
This article presents a detailed case study of how Meituan's in‑store dining sales team identified severe efficiency issues in their Apache Kylin‑based OLAP system, dissected the construction process, and applied a step‑by‑step optimization roadmap—including engine migration, dimension pruning, resource configuration, and Spark‑based layered building—to boost query performance and achieve near‑perfect SLA.
The author recounts a month‑long, four‑round technical interview at Alibaba Mama, detailing phone, on‑site, and HR stages, with deep discussions on Spark, Kafka, Hadoop, platform design, and backend fundamentals, while sharing resource links for big‑data interview preparation.
This article describes a high‑throughput solution for calculating hourly feature coverage, positive‑sample ratio and negative‑sample ratio on billions of records by streaming data with Spark, indexing per experiment‑hour in Elasticsearch, and executing parallel aggregation tasks with Java code.
This guide demonstrates how to build a Docker‑based Spark standalone environment, configure Apache Zeppelin to connect to it, and perform data analysis on local CSV files, HDFS, and streaming sources such as Twitter and Kafka, with complete code examples.
This article details a real‑world investigation of Spark SQL job latency on a YARN cluster, explains how switching the scheduler to FAIR mode, creating resource pools, and consolidating small Parquet files dramatically reduced scheduler delay and cut execution time from over 100 seconds to under 20 seconds.
This article provides a comprehensive guide to Spark performance tuning, covering essential configuration parameters, best‑practice programming recommendations, detailed shuffle mechanics, join optimization strategies, and common error troubleshooting for big‑data workloads.
This article examines the pervasive small‑file issue in big‑data environments, explains its impact on storage and processing performance, and presents a comprehensive set of solutions—including file merging, Hadoop archives, SequenceFiles, HBase, CombineFileInputFormat, and Spark/Flink strategies—to mitigate metadata overhead and improve I/O efficiency.
This article details JD's Golden Eye business's adoption of ClickHouse for offline and real‑time traffic data analysis, covering system architecture, data ingestion pipelines, high‑availability design, monitoring, performance optimizations, and practical trade‑offs, offering insights for large‑scale analytical database deployments.
This tutorial introduces Apache Zeppelin, explains REPL and Jupyter concepts, outlines its core features and project structure, and provides step‑by‑step instructions for installing Zeppelin, creating notebooks, and connecting to databases, Spark, and Flink with practical code examples.
This article explains what data skew is, why it occurs in large‑scale Hadoop and Spark jobs, illustrates typical scenarios that cause it, and provides practical strategies and platform‑specific optimizations to detect, mitigate, and prevent skew in big‑data processing pipelines.
This article explains how Spark can be combined with MongoDB for large‑scale data processing, covering Spark fundamentals, comparisons with HDFS, practical integration patterns, performance benefits, real‑world case studies, and detailed code examples for deployment and analytics.
This article explains the concept and benefits of data lakes, outlines the storage and acceleration challenges they pose, presents an ideal checklist for selecting a data lake solution, and evaluates Alibaba Cloud's JindoFS against that checklist, highlighting its capabilities for big‑data and AI workloads.
This article shares practical experiences of building an industrial data middle‑platform with DeltaLake, covering heterogeneous distributed stream handling, batch‑stream unified analytics, and transactional/algorithm support to improve data timeliness, reliability, and operational efficiency in manufacturing environments.
This article outlines the purpose, timing, procedures, tools, and optimization techniques for big data performance testing, providing detailed guidance on test planning, execution, metric collection, and analysis to ensure reliable and efficient big data system deployments.
Meituan’s sales system “Qingtian” boosted OLAP performance by migrating Apache Kylin’s build engine from MapReduce to Spark, consolidating Hive files, refining dictionary creation, applying a By‑layer algorithm, and bulk‑loading cuboid files to HBase, cutting resource consumption and halving build time, ultimately reaching a 100 % SLA.
This article explains what data skew is, why it occurs in large‑scale Hadoop and Spark jobs, how to recognize its symptoms such as stuck reducers or OOM executors, and presents practical strategies—including business‑level adjustments, code refactoring, and platform‑specific tuning—to mitigate the problem.
This article explains why Spark Streaming introduced backpressure, how the dynamic rate‑control mechanism works, and provides a detailed walkthrough of the relevant source code, including the RateController class, its registration, and the execution flow that adjusts ingestion rates to match processing capacity.
The article presents a comprehensive big‑data solution that stores massive GPS trajectory records in HBase, processes them with Spark to identify individuals who visited a pandemic source region, and visualizes their spatio‑temporal distribution in target cities to support precise epidemic control measures.
This article introduces the Actor model, its fundamental rules, and how it underpins Flink and Spark RPC mechanisms, then explains the Akka framework, its actor hierarchy, supervision, lifecycle, and dispatcher, providing a concise foundation for distributed big‑data processing.
This article explains the zero‑copy data transfer technique, compares it with traditional read/write approaches, shows Java NIO code examples, and discusses its use in high‑performance systems such as Kafka and Spark, highlighting the reductions in context switches and memory copies.
