HBaseConAsia2019 Keynote
- 1. THE COMMUNITY EVENT FOR APACHE HBASE™ THE COMMUNITY EVENT FOR APACHE HBASE™ July 20th, 2019 - Beijing, China THE COMMUNITY EVENT FOR APACHE HBASE™
- 2. THE COMMUNITY EVENT FOR APACHE HBASE™ THE COMMUNITY EVENT FOR APACHE HBASE™ July 20th, 2019 - Beijing, China
- 3. Opening Speech Baoqiu Cui | Xiaomi ⼩小⽶米集团副总裁 集团技术委员会主席
- 5. 开源 · 开放 · 平等 · 全球化 拥抱开源是⼩小⽶米的⼯工程⽂文化
- 8. 感谢 HBase 社区！ 感谢来⾃自全球的 HBase 贡献者！
- 9. The current status of HBase Duo Zhang | Xiaomi HBase PMC主席 Duo Zhang <[email protected]> Michael Stack <[email protected]>
- 10. Pervasive... ...distributed, scalable, big data store
- 11. In a nutshell... ➢...16,504 commits made by 391 contributors ➢...representing 862,469 lines of code ➢...mostly written in Java ➢...has a well established, mature codebase ➢...maintained by a very large development team ➢...with stable Y-O-Y commits ➢...took an estimated 240 years of effort (COCOMO model) ➢...starting with its first commit in April, 2007 (>10 years old!) Source https://www.openhub.net/p/hbase
- 13. Issues
- 14. Commits per month Source https://www.openhub.net/p/hbase
- 16. Our project ➢Apache HBase is an Open Source Apache project. ➢It's what we want to make of it. ➢No owners! ➢Anyone can help! ➢All welcome! ➢The more, the merrier!
- 17. From Our PMC Chair “The HBase project represents solid, useful computer science that solves problems and runs businesses every day. To keep that going, we need to keep bringing new ideas and approaches into the project...we need to continue to attract people from all backgrounds and parts of the world. I'd love to see more women, more people of color, and even more worldwide diversity. I'd like to see more contributions from people not employed by big data platform companies.” “If we continue to strive for a diversity of ideas and experiences, we'll keep innovating so that HBase remains relevant for years to come.” Misty Linville, Vice-President of the Apache HBase Project
- 18. Active branches Active Branches Latest Branch Release Release Manager branch-1.2 EOL'd 05/2019 Sean Busbey branch-1.3 1.3.5 (Yahoo) Francis Liu branch-1.4 1.4.10 (Current Stable) Andrew Purtell branch-1.5 Coming... Andrew Purtell branch-2.0 2.0.5 (EOL after 2.0.6) Michael Stack branch-2.1 2.1.5 Duo Zhang branch-2.2 2.2.0 Guanghao Zhang branch-3 Coming... Volunteer wanted!
- 19. HBase 2.2.0
- 20. In general ➢Redesign of the assignment related procedures ➢A TransitRegionStateProcedure rules all ➢No AssignProcedure, UnassignProcedure, MoveRegionProcedure ➢More proc-v2 based operations ➢Split WAL (experimental) ➢Upgrading ➢No effect on 1.x release ➢2.1.x & 2.0.x ➢The candidate of next stable pointer
- 21. TransitRegionStateProcedure ➢4 Types ➢Assign, Unassign, Move, Reopen ➢Sub procedures ➢OpenRegionProcedure ➢CloseRegionProcedure ➢One time, no retry ➢Why a new one? ➢MoveRegionProcedure does not work well with ServerCrashProcedure ➢Lots of conditions to deal with tons of corner cases, but still buggy
- 22. More improvements ➢More fencings ➢Master side on reportRegionStateTransition ➢RegionServer side on executeProcedures ➢WALProcedureStore ➢Fix bugs ➢Make it more robust
- 23. More proc-v2 based operations ➢Performance ➢Send request to RegionServer directly ➢Staged execution ➢Supported by the proc-v2 framework ➢ZooKeeper less ➢Ideally, only use ZooKeeper as an external storage ➢Support etcd, or other HA systems ➢Cloud Native
- 24. Upgrading to 2.2.0 ➢From 1.x ➢Upgrade to 1.4.x first for safety ➢Turn on zk-less assignment ➢Rolling upgrade region servers first ➢Upgrade masters ➢From 2.1.x/2.0.x ➢Make sure there are no RITs, then upgrade master first, and then rolling upgrade region servers ➢If there are old RIT procedures the new master will quit to make sure there are no damages to the cluster ➢There is an option for helping upgrading to 2.2.0, please see the upgrading section in our ref guide
- 25. 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
- 26. What’s Next
- 27. 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
- 28. 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) ➢…
- 29. 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
- 30. 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!
- 33. The Apache Way ➢Independence ➢Only PMC can control the direction of the project ➢Diversity ➢Community Over Code ➢A healthy community is a higher priority than good code
- 35. The advantages and technology trend of HBase on the cloud Apache HBase PMC 阿⾥里里巴巴 资深专家 阿⾥里里云 HBase负责⼈人 沈沈春辉 Chunhui Shen | Alibaba
- 36. HBase在阿⾥里里的使⽤用情况 The current status of HBase at Alibaba
- 37. 使用场景 Ali-HBase Alipay Bills Cainiao Logistics Lo g Monitor, Log, Tracking, IoT Data…Message, Orders, Feeds … AI Storage Ant Intelligent Security Recommendation Search, BI Report… Intelligent Customer Service Core Scenarios
- 38. 使用规模 10000+ Nodes 100+ Clusters 300+ 200+ PB Data 9000+ UsersMillion OPS Use Scale
- 39. Begin： 1.Physical machine 2.One HBase per App 3.Exclusive ZK&HDFS per HBase Phase 1 2010~2012 Change： 1.Multi-App on Shared HBase 2.Use RSGroup for isolation 3.HBase on Shared ZK&HDFS Change： 1. Exclusive ZK&HDFS per HBase Change： 1. Run on Cloud Infrastructure Phase 2 2012~2014 Phase 3 2014~Now Phase 4 2018~Now 部署演进 Evolution of deployment Pain: Maintenance Cost 、 Resource Fragmentation 、Waste of MetaServer Pain: stability Pain: move machines Weak： Storage Fragmentation Challenge: Be Cloud Native
- 41. 当HBase运行在云上时 “云”的能力如何发挥 HBase的技术如何演进 When HBase runs on the cloud How to unleash the energy of Cloud The technology challenge and trend of HBase
- 42. 当HBase在云上运⾏行行时 优势⼀一：超弹性 When HBase runs on the Cloud Advantage 1：Hyper Elasticity
- 43. ⼤大促扩容 突发流量量 异常隔离 需要弹性 Need elasticity Expansion for shopping day Burst traffic Anomaly isolation
- 44. 扩容时间 = 资源准备 + 环境准备 + 节点加⼊入 + 节点服务 HBase： 1 ~ 5 MinsTraditional Datacenter： Hours ~ Days 弹性的衡量#1 Non-matched Measurement of elasticity #1 node join in node in full serviceresource preparation environment preparation Time of capacity expansion
- 45. 云上的资源创建 Physical Resource Pool Resource Scheduling Management Compute virtualization Storage virtualization Network virtualization VM Instance VM Instance VM Instance VM Instance Create instance on the cloud 虚拟化 + 池化 快速创建实例例 Virtualization Resource pool Quickly create an instance
- 46. 扩容时间 = 资源准备 + 环境准备 + 节点加⼊入 + 节点服务 HBase： 1 ~ 5 MinsCloud Platform： 1 ~ 5 Mins 弹性的衡量#2 Matched Measurement of elasticity #1 node join in node in full serviceresource preparation environment preparation Time of capacity expansion
- 47. 弹性的价值-成本 Resource Timeline Actual Resource Requirement Actual System Capacity Double 11 Shopping day Waste of Resource Potential Optimization by Elasticity The value of elasticity - cost Daily promotionNew feature Mid-year sales promotion
- 48. 系统负载报警 Load Warning 业务流量报警 Application Traffic Warning 自动扩容 Auto scaling 系统平稳 Back to healthy 人工排查 Take a check 资源释放 Release temporary resource 弹性的价值-稳定#1 业务受损 Business suffers damage 业务恢复 Application recovery The value of elasticity – stability #1
- 49. RS4t4-region1 t4-region2 Abnormal Group t1-region1 t2-region1 t3-region1 t2-region2 t1-region2 t3-region2 RS2 RS3 RS1 Normal Group t4-region1 弹性的价值-稳定#2 1 T4表请求出现异常 3 隔离T4表 2 扩容资源 The value of elasticity – stability #2 Abnormal request to table t4 Add new node RS4 Move Table T4 to the new group
- 50. 当HBase在云上运⾏行行时 优势⼆二：低成本 When HBase runs on the Cloud Advantage 2：Low Cost
- 51. 服务器采购成本 The cost of buying machines Average buy cost of one X86 server Number of x86 server to buy Cost of Alibaba Cloud 1001 500 2,000 10,000 50,00030,000 规模越⼤大 单位成本越低 The larger of scale The lower of cost
- 52. 物理部署下的常见情况 Cluster A CPU：5% Storage：80% Cluster B CPU：40% Storage ：30% Cluster C CPU：20% Storage ：70% HDFS HBASE HBASE HDFS HBASE HDFS 不不同集群(workload)使⽤用同配置的服务器器 Common situation when deploy on physical machines Different clusters(workload) have servers with the same hardware configuration
- 53. 上云之后#1 Cluster A CPU：20% Storage ：70% Cluster B CPU：20% Storage ：70% Cluster C CPU：20% Storage ：70% HBASE HDFS HBASE HDFS HBASE HDFS After deploy on the cloud #1
- 54. 上云之后#2 不不同集群(workload)使⽤用不不同配置的虚拟机 灵活的存储计算比 存储计算分离 同集群使⽤用相同配置，让运维保持简单 After deploy on the cloud #2 Flexible ratio between storage and computation Separation of storage and computation Different clusters could have servers with the different hardware configuration The same hardware configuration in one cluster， keep the maintenance be easy
- 55. HDFS异构存储#1 HDD SSD HDD HDD SSD HDD HDD SSD HDD HDFS Cluster Datanode Datanode Datanode One_SSD 1 Replica on SSD, others on HDDTable A A A A All_HDD All Replicas on HDD Table B B B B Advantage： 1. Cost-effective 2. Allocate storage according to scenario Disadvantage： 1. Personalized machine type 2. The capacity ratio of SSD to HDD HDFS heterogeneous storage #1
- 56. 异构存储介质#2 HDD云盘 SSD云盘 HDD云盘 HDD云盘 SSD云盘 HDD云盘 HDD云盘 SSD云盘 HDD云盘 HDFS Cluster Datanode Datanode Datanode One_SSD 1 Replica on SSD, others on HDDTable A A A All_HDD All Replicas on HDD Table B B B HDFS heterogeneous storage #2 Advantage： 1. Cost-effective 2. Allocate storage according to scenario Disadvantage： 1. Personalized machine type 2. The capacity ratio of SSD to HDD
- 57. 当HBase在云上运⾏行行时 优势三：⾼高稳定性 When HBase runs on the Cloud Advantage 1：High Stability
- 58. 阿里云ECS的稳定性特征 机器器过保⽆无感知 极低物理理宕机率 机房迁移⽆无感知 99.95% Availability SLA Never Retired Stability characteristics of Alibaba Cloud‘s ECS Not aware of machine replacement Not aware of datacenter migration The very low probability of machine downtime
- 59. 物理机房-Auto Recovery，But not sustainable HBase Cluster Region1 Region2 Region3 Region4 Region5 Region6 Region7 Region8 HBase Cluster Region1 Region2 Region3 Region4 Region5 Region6 Region7 Region8 HBase Cluster Region1 Region2 Region3 Region4 Region5 Region6 Region7 Region8 1 2 3 HBase Cluster Region1 Region3 Region4Region5 4 Region2 Region7 Region6 Region8 N久后，⼈人⼯工介⼊入 负载⻛风险 When deploy on physical datacenter load risk After a long period of time, manual intervention
- 60. 上云之后-Auto recovery forever HBase Cluster Region1 Region2 Region3 Region4 Region5 Region6 Region7 Region8 HBase Cluster Region1 Region2 Region3 Region4 Region5 Region6 Region7 Region8 1 2 Region1 Region3 Region4Region7 3 Region5 Region6 Region8 Region2 Restart the processes automatically HBase Cluster After deploy on the cloud Instance recovers automatically
- 61. Long Compaction Queue Replication Delay Too Much Failover Too Slow 其他常见的问题 Other common problems
- 62. 没有什么问题是扩容和重启不能解决的 在云上，问题处理理变得更更简单 No problem that can't be solved by increasing capacity and rebooting On the cloud, problem handling becomes easier
- 63. 当HBase在云上运⾏行行时 优势四：全球部署 When HBase runs on the Cloud Advantage 4：Global Deployment
- 64. APP 全球多活 全球服务 Global service 异地容灾 Resilience of regional failure 就近访问 Localized access Multi-master at global scale HBase APP HBase APP HBase
- 65. HBase Replication HBase Cluster HBase Cluster HBase Cluster Read&WriteRead&Write Read Only Replication 特点： 1. ⽀支持多活 Support for multi-master 2. ⾃自由拓拓扑 Free topology 3. ⽀支持循环链路路 Support for circular data links 4. 在线添加/删除链路路 Add and remove data links online
- 66. HBase Replication + Globally Distributed Datacenters 上云之后 Multiple active masters at global scale with HBase Build replication on the cloud
- 67. 当HBase在云上运⾏行行时 技术的挑战及趋势 When HBase runs on the Cloud Challenge and Trend of Technology
- 68. Log与File分离存储 Cloud SSD Disk Cloud HDD Disk Object Storage(OSS) Cost comparison between hbase on cloud disk and hbase on object storage 0 0.5 1 1.5 2 0.2 0.7 2 Cloud SSD Disk with 2 Replicas Cloud HDD Disk with 2 Replicas Object Storage : 1 11 : Write Performance : 3.5 17 : The above data comes from the HBase test on Alibaba Cloud，only for reference Table3Table2 HLog Table1 Alibaba Cloud HBase Separate Storage of Log and File 元/GB/Month Read Performance
- 69. 使用共享块存储 MetaNode MetaNode FileNode FileNode FileNode Shared Disk1 Shared Disk2 Shared Disk3 Shared Disk4 Shared Disk5 Shared Disk6 Shared Disk Pool SFS Cluster SFS Client API HDFS Compatible API HDFS Compatible Shell 共享块存储: ⼀一种⽀支持多台ECS实例例并发读写访问的数据块级存储设备 Shared block storage supports concurrent access from multiple ECS instances SFS:基于共享云盘的云原⽣生分布式⽂文件系统 SFS-A Cloud Native Distributed File System based on cloud shared block storage HDFS on Normal Cloud Disk At least 2 replicas SFS on Shared Block Storage Only 1 replica Advantage of SFS： 1. 200% 存储容量量提升 More available storage than HDFS 2. 130% 吞吐性能提升 Better throughput 3. 兼容HDFS客户端 HDFS compatible 4. ⽔水平扩展⾄至PB级存储 Scale out to PB level 5. 单节点故障持续可⽤用 Automatic fault tolerance Make use of shared block storage
- 70. Compaction on FPGA At the same cost , FPGA shows 3X higher compaction performance than CPU Eliminate the impact of compaction on online requests HFile HFile RS-1 HFile HFile RS-2 HFile HFile RS-3 Query Query Query Compaction Compaction Worker Client Client … Client Compaction Worker Cloud accelerates landing the new hardware
- 71. 使用阿里云HBase，获得最佳云上体验 阿里云数据库HBasehttps://cn.aliyun.com/product/hbase 自建HBase 核心指标 吞吐性能 是社区版的3-7倍 N/A 毛刺 P99延迟是社区版的1/10 N/A 压缩率 是社区版的1.5-2倍 N/A 可用性 SLA保障，单集群99.9%，双集群99.99% N/A 主备容灾 成熟、支持跨版本、主无影响、且支持透明切换 无优化，不支持切换 企业特性 数据搬迁 支持在线、跨版本、自动化、高效搬迁，应用0影响、0改造 自己用工具、 效率&正确性无保障 备份恢复 支持数据备份至OSS及恢复 不支持 冷热分离 支持，冷数据存储在OSS，成本减少70% 不支持 多租户 支持全局Quota限流 只支持单Server 生态体验 MySQL数据同步 提供产品化功能、支持在线增量同步 自己用工具、不支持在线增量 Spark分析 产品化深度集成，支持高性能分析、增量归档、结果回流等 无优化，数据集成需要较大开发 功能 二级索引 支持强一致、最终一致、本地索引等多形态，性能大幅优化 依赖外部组件 全文索引 产品化关联外部搜索引擎Solr、Elasticsearch，自动数据同步 不支持 专业运维 支持HBase-Manager (表管理、数据查询、热点识别、自动合并、大scan识别等) 无 技术团队 专家 拥有HBase领域专家数十人 Apache社区4 PMC、6 Committer - 社区 中国HBase技术社区发起者 定期主办HBase Con、Meetup等中大型会议 - 实践经验 内部打磨9年，支持天猫双十一，阿里部署10000+台，公有云600+集群 -
- 73. THE COMMUNITY EVENT FOR APACHE HBASE™ THE COMMUNITY EVENT FOR APACHE HBASE™ July 20th, 2019 - Beijing, China