Advantages of MRS Compared with Self-Built Hadoop
MRS provides enterprise-level big data clusters on the cloud. Tenants can fully control the clusters and run big data components such as Hadoop, Spark, HBase, Kafka, and Storm with ease. MRS frees you from hardware purchase and maintenance. MRS is built based on Huawei FusionInsight big data enterprise-class platform, and has been deployed on tens of thousands of nodes in the industry, providing multi-level SLA assurance with professional Hadoop kernel service support. Compared with self-built Hadoop clusters, MRS has the following advantages:
- MRS supports one-click cluster creation, deletion, and scaling. You can use an elastic IP address (EIP) to access MRS Manager, making big data clusters easier to use.
- Self-built big data clusters pose problems such as high costs, long periods, difficult and inflexible O&M. To solve these problems, MRS provides one-click cluster creation, deletion, scale-out, and scale-in, allowing you to customize the cluster type, component range, number of nodes of each type, VM specifications, availability zones (AZs), VPC network, and authentication information. MRS can automatically create a cluster that meets the configuration requirements. In addition, you can quickly create multi-application clusters, for example, Hadoop analysis cluster, HBase cluster, and Kafka cluster. MRS supports heterogeneous cluster deployment. That is, VMs of different specifications can be combined in a cluster based on CPU types, disk capacities, disk types, and memory sizes .
- MRS provides an EIP-based secure channel for you to easily access the web UIs of components. This is more convenient than binding an EIP by yourself, and you can access the web UIs with a few clicks, avoiding the steps for logging in to a VPC, adding security group rules, and obtaining a public IP address.
- MRS provides custom bootstrap actions to flexibly configure your dedicated clusters. Third-party software that is not supported by MRS can be automatically installed, allowing you to perform custom operations such as modifying the cluster running environment.
- MRS supports the WrapperFS feature, provides the OBS translation capability (that is, access to OBS through address mapping) and can smoothly migrate data from HDFS to OBS. After migration, you can access the data stored in OBS from clients without modifying service code logic.
- MRS supports auto scaling, which is more cost-effective than the self-built Hadoop cluster.
MRS supports auto scaling to address peak and off-peak service loads. It applies for extra resources during peak hours and releases idle resources during off-peak hours, helping you save idle resources on the big data platform during off-peak hours, minimize costs, and focus on core services.
In big data applications, especially in periodic data analysis and processing, cluster computing resources need to be dynamically adjusted based on service data changes to meet service requirements. The auto scaling function of MRS enables clusters to be elastically scaled out or in based on cluster loads. In addition, if the data volume changes regularly and you want to scale out or in a cluster before the data volume changes, you can use the MRS resource plan feature. MRS supports two types of auto scaling policies: auto scaling rules and resource plans
- Auto scaling rules: You can increase or decrease Task nodes based on real-time cluster loads. Auto scaling will be triggered when the data volume changes but there may be some delay.
- Resource plans: If the data volume changes periodically, you can create resource plans to resize the cluster before the data volume changes, thereby avoiding a delay in increasing or decreasing resources.
Both auto scaling rules and resource plans can trigger auto scaling. You can configure both of them or configure one of them. Configuring both resource plans and auto scaling rules improves the cluster node scalability to cope with occasionally unexpected data volume peaks.
- MRS supports storage-compute decoupling, greatly improving the resource utilization of big data clusters.
In the traditional big data architecture where storage and compute resources are integrated, scaling-out is difficult and resources are not well-utilized. To solve these problems, MRS adopts a compute-storage separation architecture. Based on OBS, the storage achieves 99.999999999% reliability and unlimited capacity, supporting continuous growth of enterprise data. Computing resources can be elastically scaled in or out from 0 to N nodes. Hundreds of nodes can be quickly provisioned. With the new architecture, compute nodes can be elastically scaled. OBS-based cross-AZ data storage ensures higher reliability, frees you from worrying about emergencies such as earthquakes and fiber cuts. Storage and compute resources can be flexibly configured and elastically scaled as required. This makes resource allocation more accurate and reasonable, greatly improving the resource utilization of big data clusters and reducing the comprehensive analysis cost by 50%.
In addition, the high performance compute-storage separation architecture breaks the limit of parallel computing of the integrated storage-compute architecture. It maximizes the high bandwidth and high concurrency of OBS, and optimizes the data access efficiency and in-depth parallel computing (such as metadata operation and write algorithm optimization) to improve higher performance.
- MRS supports Huawei-developed CarbonData and Superior Scheduler, delivering better performance.
- MRS supports self-developed CarbonData storage technology. CarbonData is a high-performance big data storage solution. It allows one data set to apply to multiple scenarios and supports features, such as multi-level indexing, dictionary encoding, pre-aggregation, dynamic partitioning, and quasi-real-time data query. This improves I/O scanning and computing performance and returns analysis results of tens of billions of data records in seconds.
- In addition, MRS supports self-developed Superior Scheduler, which enhances the scaling capability of a single cluster and is capable of scheduling over 10,000 nodes in a cluster. Superior Scheduler is a scheduling engine designed for the Hadoop Yarn distributed resource management system. It is a high-performance and enterprise-level scheduler designed for converged resource pools and multi-tenant service requirements. Superior Scheduler achieves all functions of open source schedulers, Fair Scheduler, and Capacity Scheduler. Compared with the open source schedulers, Superior Scheduler is enhanced in the enterprise multi-tenant resource scheduling policy, resource isolation and sharing by multiple users in a tenant, scheduling performance, system resource utilization, and cluster scalability, and is designed to replace open source schedulers.
- MRS optimizes software and hardware based on Kunpeng processors to fully release hardware computing power and achieve cost-effectiveness.
MRS supports Huawei-developed Kunpeng servers whose multi-core and high-concurrency capabilities are fully utilized to provide full-stack self-optimized chips, and uses Huawei-developed EulerOS, Huawei JDK, and data acceleration layer to ensure hardware performance, delivering high computing power for big data computing. With the similar performance, the cost of the end-to-end big data solution is reduced by 30%.
- MRS supports multiple isolation modes and multi-tenant permission management of enterprise-level big data, ensuring higher security.
- MRS supports resource deployment and isolation of physical resources in dedicated zones. You can flexibly combine computing and storage resources, such as dedicated computing resources + shared storage resources, shared computing resources + dedicated storage resources, and dedicated computing resources + dedicated storage resources. An MRS cluster supports multiple logical tenants. Permission isolation enables the computing, storage, and table resources of the cluster to be divided based on tenants.
- With Kerberos authentication, MRS provides role-based access control (RBAC) and sound audit functions.
- With Cloud Trace Service (CTS) being interconnected with MRS, you are provided with operation records of MRS resource operation requests and request results for querying, auditing, and backtracking. You can use CTS to audit and trace all cluster operations.
- It is proved that with Host Security Service (HSS) interconnected with MRS, service security is enhanced without deteriorating functions and performance.
- MRS supports unified user login based on web UI. MRS Manager provides user authentication, which grants you permission to access a cluster.
- MRS supports data storage encryption, encrypted storage of all user accounts and passwords, encrypted transmission of data channels, and bidirectional certificate authentication for cross-trusted-zone data access of service modules.
- MRS big data clusters provide a complete multi-tenant solution for enterprise-level big data. Multi-tenant refers to a collection of multiple resources (each resource set is a tenant) in an MRS big data cluster. It can allocate and schedule resources, including computing and storage resources. Multi-tenant isolates the resources of a big data cluster into resource sets. Users can lease desired resource sets to run applications and jobs and store data. In a big data cluster, multiple resource sets can be deployed to meet diverse requirements of multiple users.
- MRS supports fine-grained permission management. With the fine-grained authorization capability provided by HUAWEI CLOUD IAM, MRS can specify the operations, resources, and request conditions of specific services. This mechanism allows for more flexible policy-based authorization, meeting requirements for secure access control. For example, you can grant MRS users only the permissions for performing specified operations on MRS clusters, such as creating a cluster and querying a cluster list rather than deleting a cluster. In addition, MRS supports fine-grained permission management of OBS for multiple tenants. Permissions to access OBS buckets and objects in the buckets are differentiated based on user roles, so that MRS users can each control a different directory in OBS buckets.
- MRS supports enterprise project management. The enterprise project is one way of managing cloud resources. Enterprise Management provides comprehensive management services for enterprise customers, such as cloud resources, personnel, permissions, and financial statuses. Common management consoles are oriented to the control and configuration of individual cloud products. The Enterprise Management console, in contrast, is more focused on resource management. It is designed to help enterprises manage cloud-based resources, personnel, permissions, and finances, in a hierarchical management manner, such as management of companies, departments, and projects. MRS allows users who have enabled Enterprise Project Management Service (EPS) to configure enterprise projects for a cluster during cluster creation and use EPS to manage MRS resources by group. This feature is applicable to scenarios where you need to manage multiple resources by group and perform operations such as permission control and project-based fee query on enterprise projects.
- MRS implements HA for all management nodes and supports comprehensive reliability mechanism, making the system more reliable.
Based on Apache Hadoop open source software, MRS optimizes and improves the reliability of main service components.
- HA for all management nodes
In the Hadoop open source version, data and compute nodes are managed in a distributed system, in which a single point of failure (SPOF) does not affect the operation of the entire system. However, a SPOF may occur on management nodes running in centralized mode, which becomes the weakness of the overall system reliability.
MRS provides similar double-node mechanisms for all management nodes of the service components, such as MRS Manager, Presto, HDFS NameNodes, Hive Servers, HBase HMasters, Yarn ResourceManagers, Kerberos Servers, and Ldap Servers. All of them are deployed in active/standby mode or configured with load sharing, effectively preventing SPOFs from affecting system reliability.
- Comprehensive reliability mechanism
By reliability analysis, the following measures to handle software and hardware exceptions are provided to improve the system reliability:
- After power supply is restored, services are running properly regardless of a power failure of a single node or the whole cluster, ensuring data reliability in case of unexpected power failures. Key data will not be lost unless the hard disk is damaged.
- Health status checks and fault handling of the hard disk do not affect services.
- The file system faults can be automatically handled, and affected services can be automatically restored.
- The process and node faults can be automatically handled, and affected services can be automatically restored.
- The network faults can be automatically handled, and affected services can be automatically restored.
- HA for all management nodes
- MRS provides a visualized big data cluster management interface in a unified manner, making O&M easier.
- On the big data cluster management interface, service startup and stopping, configuration modification, and health check are available. MRS also provides visualized and convenient cluster management, monitoring, and alarm functions. Additionally, you can check and audit the system health status in one click, ensuring normal system running and lowering system O&M costs.
- After Simple Message Notification (SMN) is configured, MRS can send real-time cluster health status information, including cluster changes and component alarms in real time to you through SMS messages or emails, facilitating O&M, real-time monitoring, and real-time alarm sending.
- MRS supports rolling patch upgrade and provides visualized patch release information and one-click patch installation without manual intervention, ensuring long-term stability of user clusters.
- If a problem occurs when you use an MRS cluster, you can initiate O&M authorization on the MRS management console. O&M personnel can help you quickly locate the problem, and you can revoke the authorization at any time. You can also initiate log sharing on the MRS management console to share a specified log scope with O&M personnel, so that O&M personnel can locate faults without accessing the cluster.
- MRS supports to dump logs about cluster creation failures to OBS for O&M personnel to obtain and analyze the logs.
- MRS has an open ecosystem and supports seamless interconnection with peripheral services, allowing you to quickly build a unified big data platform.
- Based on MRS, a full-stack big data service, enterprises can build a unified big data platform for data access, data storage, data analysis, and value mining, and interconnect with DAYU and data visualization services to help customers easily resolve difficulties in data channel cloudification, big data job development and scheduling, and data display. Thereby, customers are free from complex big data platform construction and professional big data optimization and maintenance, and they can focus on industry applications and use one piece of data in multiple service scenarios. DAYU is a one-stop operation platform for entire data lifecycle management. It provides data integration, data development, data governance, data service, and data visualization functions. MRS data can be connected to DAYU, which greatly reduces the threshold for using big data. Based on the visualized development GUI, diverse data development types (script development and job development), fully hosted job scheduling and O&M monitoring capabilities, and built-in industry data processing pipelines, it helps you quickly build big data processing centers with visualized development processes and online collaborative operations. In this way, data can be well-managed and scheduled for real profits quicker.
- MRS is fully compatible with the open source big data ecosystem. With abundant data and application migration tools, MRS helps you quickly migrate data from your own platforms without code modification and service interruption.