Big Data HBase

HBASECON Asia 2019

Community Over Code！

Posted by HH on July 22, 2019

这次HBase Conference Asia 2019主要分为两个部分，一部分是Key Note,另一部分是HBase Internal

下面我为大家主要介绍一下这次会议的亮点和未来的趋势。

HBase的现状及发展趋势

HBase目前的稳定版本还是在1.4.x。2.x版本优化了很多的新特性，比如offheap的读，region状态合理化，procedure v2架构，减少对hdfs和zk的依赖，等等等等。

发展方向

HBase项目目前由小米公司主导了2.x版本，没错就是小米。HBCK2这个非常重要的运维工具由Cloudera开发。所以对于HBase这个产品在出现一些故障时是通过procedure v2架构自恢复还是使用HBCK2由人工界定来恢复一直处于拉锯状态。目前来看是小米占了上风（刚诞生一个主席，更加有话语权）。

云原生HBase

对于HBase云原生和最佳迁移方案，阿里云走在前面。相关的详细资料和产品走传送门可以了解。腾讯云提供云主机给客户来搭建和运维，与我们自建差异不大。

SQL的支持

HBase是KV存储，不支持SQL和创建二级索引。社区有Pheonix提供支持，性能不好。对此各场有了自己研发的建立索引和搜索引擎。快手的BitBase和华为的Lemon SQL都是通过BitMap来实现的，BitBase近期应该会开源，可以期待一下。

硬件方法面的进展

来自Intel的工程师，主要集中在两个地方，一个是FPGA，一个是Persistent memory，一个侧重性能优化，一个侧重降低成本。

最后

下面是HBase Conference Asia 2019我参与了的部分PPT原文，从拍的照片手打出来的，后续社区应该会放出原版，我会在这里贴下载链接。提取码：i60j。

HBase拥抱云原生，逐渐脱离HDFS和zk在当时的现场是有些担忧和怀疑的，通过后面的深入了解这或许就是趋势。最后引一下全场最让我感动的地方“Community Over Code”。一个健康的社区比好的代码更重要，代码不好可以重构，脱离社区一个产品也将走向尽头。

Key Note

The current status of HBase BY Duo Zhang

The stable pointer

  • Currently still on 1.4.x
  • AM-v2 is still not stable enough
  • AM-v1 is not stable either

HBCK2

  • Can fix basic region assignment problems
  • A bit different from HBCK1
  • Still not powerful enough compare to HBCK1
  • Cluster status report

New projects

  • HBase Connectors
    • Kafka Proxy
    • Spark
    • 1.0.0
  • HBase Filesystem
    • HBOSS,HBase OSS(S3) Adaptor
  • HBase Operator Tools
    • HBCK2
  • HBase Native Client
    • HBase-14850

HBase 3.0.0

  • Plan to cut branch-3 by the last quarter
  • Need to have a ‘stable’ 2.X release line first
  • ‘New’ features
    • Fold namespace table into meta table(HBASE-21154)
    • Synchronous replication(HBASE-19064)
    • Off-Heap read starting from DFSClient(HBASE-21879)
    • Proc-v2 based ACL(HBASE-21602)
    • Reimplement sync client on top of async client(HBASE-21512)

Delayed

  • SQL Engine
    • Like CQL
    • Phoenix
    • Upstream of in-house solution
  • Splitable meta
    • Scalability
    • Optimistic to have this in 4.0.0
  • WAL abstraction
    • Remove the last unavoidable HDFS dependency
    • Cloud Native
    • HBASE-20952

Github PR is now available

  • Still need a jira account to file an issue…
  • The commit message for the PR should start with the issue number, for example”HBASE-12345 Test Github PR”
  • All other things can be done on Github

The Apache Way

  • Independency
    • Only PMC can control the direction of the project
    • Diversity
  • Community Over Code
    • A healthy community is a higher priority than good code

The advantages and technology trend of HBase on the cloud BY Chunhui Shen

Ali-HBase Core Scenarios

  • Message,Orders,Feeds
  • Monitor,Log,Tracking,IoT Data
  • AI Storage
  • Recommendation Search, BI Report

Use Scale

  • 10000+ Nodes
  • 100+ Clusters
  • 300+ Million OPS
  • 200+ PB Data
  • 9000+ Users

Evolution of deployment

  • Physical machine, One HBase per APP,Exclusive ZK&HDFS per HBase
  • Multi-APP on Shared HBase,Use RSGroup for isolation,HBase on shared ZK&HDFS
  • Exclusive zk&hdfs per hbase
  • Relay on cloud infrastructure

Key points in deployment

  • Stability
  • Elasticity
  • Cost
  • Efficiency

When HBase runs on the cloud

  • How to unleash the energy of Cloud
  • The technology challenge and trend of HBase

Need elasticity

  • Expansion for shopping day
  • Burst traffic
  • Anomaly isolatic (Hot Key)

Measurement of elasticity Before Cloud Native

\[Time Of Capacity Expansion=Resource Preparation+Enviroment Preparation+Node Join In+Node In Service\] \[Resource Preparation+Enviroment Preparation=Traditional Datacenter:Hours To Days\] \[Node Join In+Node In Service=HBase:1To5 Hours\]

Create instance on the cloud

  • Virtualization Resource pool
  • Quickly create an instance

cloud native

Measurement of elasticity Aefore Cloud Native

\[Time Of Capacity Expansion=Resource Preparation+Enviroment Preparation+Node Join In+Node In Service\] \[Resource Preparation+Enviroment Preparation=Cloud Platform:1To5Hours\] \[Node Join In+Node In Service=HBase:1To5 Hours\]

The value of elasticity - Cost

\[Waste Of Resource=Actual System Capacity-Actual Resource Requirement\]

The value of elasticity-Stability

  • Abnormal request to table t4
  • Add new node RS4
  • Move Table to the new group

Common situation when deploy on physical machines

Cluster A Cluster B Cluster C
CPU 5% CPU 40% CPU 20%
Storage 80% Storage 30% Storage 70%

Different clusters (workload) have servers with the same hardware configuration.

After deploy on the cloud

  • Flexible ratio between storage and computation
  • Different clusters could have servers with the different hardware configuration
  • Separation of storage and computation
  • The same hardware configuration in one cluster, keep the maintenance be easy

HDFS heterogeneous storage

Table A Table B
One_SSD
1 Replica on SSD
others on HDD		 All_HDD
All Replica on HDD

Stability characters of Alibaba Cloud’s ECS

  • Not aware of machine replacement(99.95% Availability SLA)
  • Not aware of datacenter migration(Never Retired)
  • The very low proability of machine downtime

Other common problems

  • Long compaction Queue
  • Replication Delay Too Much
  • Failover Too Slow

Multiple active masters at global scale with HBASE

No problem that can’t be solved by increasing capacity and rebooting.On the cloud,problem handling becomes easier.

HBase Replication + Globally Distributed Datacenters=Multiple active masters at global scale with HBASE

Sepatate Storage of Log and File

Storage Write Performance Read Performance
Cloud SSD Disk 1 7
Cloud HDD Disk 1 3.5
Object Storage(OSS) 1 1

Make use fo shared block storage

  • Shared block storage supports concurrent access from multiple ECS instances
  • SFS-A: Cloud Native Distribute File System Based On Cloud Shared bLOCK

Compaction On FPGA

HBase Internal

HBCK2 in a nutshell

  • Simpler tool
    • less fix commands
    • No diagnosis command
    • Requires deeper HBase internal working from operators
  • Shipped independently from hbase
    • packaged with hbase-operators project
    • Can evolve on its own pace
    • New versions can be run without needing whole hbase upgrade
  • Master oriented(more later)
  • Master not completing initialisation
    • Meta/Namespace table “Not Online” issues
  • Table RIT issues
  • Procedures stuck
  • Table in wrong state
  • Missing regions in META
    • User induced via imcompatible OfflineMetaRepair tool

HBCK2 For Operators: How do I get and run it?

http://github.com/apache/hbase-operator-tools

Further GC optimization: Reading HFileBlock into offheap directly

BDS: A data synchronization platform for HBase

The Procedure v2 implementation of WAL splitting and acl

Distributed Bitmap Index Solution And Lightweight SQL Engine

CATALOG
    FEATURED TAGS
    Mac 终端 Git Java token 语言基础 Apollo
    FRIENDS