The advantages and disadvantages of hadoop hdfs replicating storage and erasure coding storage.

Hadoop 3.0.0-alpha1 及以上版本提供了纠删码(Erasure Coding)存储数据的支持,用户可以根据不同的场景和需求选择副本存储或EC存储方案,两种存储方案各有优缺点和适用场景。

1 副本存储

以文件大小:2.5G 写入HDFS为例;

$ ls –ltr
-rw-r--r-- 1 sywu sywu 2.5G Mar  4 14:15 data000

hdfs数据块大小:512M,默认副本数:3;2.5G文件被切分为5个数据块(B1,B2,B3,B4,B5)存储,单节点存储完成,数据块复制到其它节点,最终总的数据块数:15个,总消耗存储空间:7.5G。

HDFS 副本 hdfs 副本缺失 修复_hadoop

1.1 优点

  1. 副本保证数据可用性;当某个节点数据块丢失或破损,datanode发现后会自动从其它正常的节点中恢复数据块;
  2. 副本提高了作业运行并行度,作业可以同时在副本节点运行。

1.2 缺点

  1. 每个副本使用100%的存储开销,副本数越多,存储开销越大;
  2. 副本同步占用大量的网络和IO资源。

2 纠删码(Erasure Coding)存储

HDFS使用纠删码(Erasure Coding,以下简称EC)解决副本复制和副本存储所带来的空间和资源消耗问题,以EC代替副本,提供和副本存储相同级别的容错能力,并且存储开销不超过单副本存储的50%。

2.1 纠删码(Erasure Coding)组成结构

HDFS 副本 hdfs 副本缺失 修复_HDFS 副本_02

  1. EC由***数据(Data)和奇偶校验码(Parity)***两部分组成,数据存储时通过EC算法生成;生成的过程称为编码(encoding),恢复丢失数据块的过程称为解码(decoding)。
  2. 与HDFS文件基本构成单位:块(Block) 不同,EC的构成单位:块组(Block group)、块(Block)、单元(cell),每个块组存放与其它块组一样数量的数据块和奇偶校验码块;单元(cell)是EC内部最小的存储结构,多个单元组成条(Striping),存储在块(Block)里。
  3. EC写入方案有:连续布局(Contiguous Layout)和条形布局(Striping Layout)***;连续布局从 Hadoop 3.0.0-alpha1 版本开始提供支持,数据依次写入块中的单元(cell),一个块写满之后再写入下一个块;条形布局写入方案***目前还在开发阶段,按官方介绍,这种方案由若干个相同大小的单元(cell)序列构成条(stripe),数据被依次写入条的各个单元中,当一个条写满之后再写入下一个条,一个条的不同单元位于不同的数据块中
  4. 数据和奇偶校验码块数量由EC策略(Erasure coding policies)决定,比如上图的策略:RS-3-2-1024k,表示每个块组由3个数据块和2个校验块构成,每个单元(cell)大小为1024k,最大可丢失块2个,丢失超过2个则无法恢复。

2.2 Erasure Coding算法

目前支持的EC算法有XOR和Reed-Solomon两种;

2.2.1 XOR 算法

exclusive OR,基于异或运算的算法,从任意数量的单元(cell)中生成1个奇偶校验块。如果任何数据块丢失,则通过对其余位和奇偶校验位进行来恢复。

HDFS 副本 hdfs 副本缺失 修复_java_03


由于只有1个奇偶校验块,因此它只能容忍1个数据块故障,对于像HDFS这样需要处理多个故障的系统来说,不适用。

2.2.2 Reed-Solomon 算法

Reed-Solomon(简称:RS),该算法有两个参数k和m,记为 RS(k,m),RS算法将k个数据单元(cell)与生成器矩阵(Generator Matrix)相乘得到具有k个数据单元和m个奇偶校验码的扩展码字(extended codewords),最多可容忍m个数据块丢失;如果数据块丢失,则只要将 k+m 个单元格中的k个可用,通过将生成器矩阵的逆与扩展码字(extended codewords)相乘来恢复数据块。由于可以容忍多个数据块丢失,更适用于生产环境。

2.3 使用EC RS算法存储数据

Hadoop 3.0.0-alpha1 版本及以上版本默认已经启用EC,首先设置目录策略使用RS算法,以6个data块和3个奇偶校验码构成一个block group;

hdfs ec -setPolicy -path /data/ec -policy RS-6-3-1024k

将2.5G数据写入EC目录下,数据最终存储在一个block group下,共9个数据块,总消耗存储空间:2.5G。

HDFS 副本 hdfs 副本缺失 修复_大数据_04


检查状态;

/data/ec/split000 2684354560 bytes, erasure-coded: policy=RS-6-3-1024k, 1 block(s):  OK
0. BP-1016852637-192.168.1.192-1597819912196:blk_-9223372036854775792_4907534 len=2684354560 Live_repl=9  [blk_-9223372036854775792:DatanodeInfoWithStorage[192.168.1.192:1019,DS-3b10bc66-c5c9-47f8-b4ea-99441fc5df04,DISK], blk_-9223372036854775791:DatanodeInfoWithStorage[192.168.1.196:1019,DS-c618bb83-03f2-4007-a3a9-cbd11eb2a15a,DISK], blk_-9223372036854775790:DatanodeInfoWithStorage[192.168.1.193:1019,DS-6554db41-d285-4f31-9fde-2017cc211b0c,DISK], blk_-9223372036854775789:DatanodeInfoWithStorage[192.168.1.188:1019,DS-09a7d2b0-6018-43af-8a63-6be8ba9217e6,DISK], blk_-9223372036854775788:DatanodeInfoWithStorage[192.168.1.199:1019,DS-b24cf7ce-1235-478f-884d-19cde4a03c9e,DISK], blk_-9223372036854775787:DatanodeInfoWithStorage[192.168.1.187:1019,DS-43b686d2-e83c-49a3-b8e5-64c4e5edcb53,DISK], blk_-9223372036854775786:DatanodeInfoWithStorage[192.168.1.195:1019,DS-eb42e686-1ca9-44c3-9891-868c67d9d1fa,DISK], blk_-9223372036854775785:DatanodeInfoWithStorage[192.168.1.194:1019,DS-15635fea-314c-4703-a7ab-6f81db1e52cd,DISK], blk_-9223372036854775784:DatanodeInfoWithStorage[192.168.1.198:1019,DS-73bc78d2-e218-40a9-ae9c-d24706a0bc31,DISK]]

Status: HEALTHY
 Number of data-nodes:  10
 Number of racks:               1
 Total dirs:                    0
 Total symlinks:                0

Replicated Blocks:
 Total size:    0 B
 Total files:   0
 Total blocks (validated):      0
 Minimally replicated blocks:   0
 Over-replicated blocks:        0
 Under-replicated blocks:       0
 Mis-replicated blocks:         0
 Default replication factor:    3
 Average block replication:     0.0
 Missing blocks:                0
 Corrupt blocks:                0
 Missing replicas:              0

Erasure Coded Block Groups:
 Total size:    2684354560 B
 Total files:   1
 Total block groups (validated):        1 (avg. block group size 2684354560 B)
 Minimally erasure-coded block groups:  1 (100.0 %)
 Over-erasure-coded block groups:       0 (0.0 %)
 Under-erasure-coded block groups:      0 (0.0 %)
 Unsatisfactory placement block groups: 0 (0.0 %)
 Average block group size:      9.0
 Missing block groups:          0
 Corrupt block groups:          0
 Missing internal blocks:       0 (0.0 %)

尝试删除(blk_-9223372036854775792、blk_-9223372036854775791、blk_-9223372036854775790)三个数据块(最多允许3个数据块丢失)。

rm /data/current/BP-1016852637-192.168.1.192-1597819912196/current/finalized/subdir0/subdir0/blk_-9223372036854775792
rm /data/current/BP-1016852637-192.168.1.192-1597819912196/current/finalized/subdir0/subdir0/blk_-9223372036854775791
rm /data/current/BP-1016852637-192.168.1.192-1597819912196/current/finalized/subdir0/subdir0/blk_-9223372036854775790

现在尝试读取文件;

hdfs dfs -get  /data/ec/split000 .

datanode首先抛未找到数据块异常;

21/03/05 16:55:39 WARN impl.BlockReaderFactory: I/O error constructing remote block reader.
java.io.IOException: Got error, status=ERROR, status message opReadBlock BP-1016852637-192.168.1.192-1597819912196:blk_-9223372036854775792_4907534 received exception java.io.FileNotFoundException: BlockId -9223372036854775792 is not valid., for OP_READ_BLOCK, self=/192.168.1.192:42168, remote=/192.168.1.192:1019, for file /data/ec/split000, for pool BP-1016852637-192.168.1.192-1597819912196 block -9223372036854775792_4907534
        at org.apache.hadoop.hdfs.protocol.datatransfer.DataTransferProtoUtil.checkBlockOpStatus(DataTransferProtoUtil.java:134)
        at org.apache.hadoop.hdfs.protocol.datatransfer.DataTransferProtoUtil.checkBlockOpStatus(DataTransferProtoUtil.java:110)
        at org.apache.hadoop.hdfs.client.impl.BlockReaderRemote.checkSuccess(BlockReaderRemote.java:447)
        at org.apache.hadoop.hdfs.client.impl.BlockReaderRemote.newBlockReader(BlockReaderRemote.java:415)
        at org.apache.hadoop.hdfs.client.impl.BlockReaderFactory.getRemoteBlockReader(BlockReaderFactory.java:860)
        at org.apache.hadoop.hdfs.client.impl.BlockReaderFactory.getRemoteBlockReaderFromTcp(BlockReaderFactory.java:756)
        at org.apache.hadoop.hdfs.client.impl.BlockReaderFactory.build(BlockReaderFactory.java:390)
        at org.apache.hadoop.hdfs.DFSInputStream.getBlockReader(DFSInputStream.java:657)
        at org.apache.hadoop.hdfs.DFSStripedInputStream.createBlockReader(DFSStripedInputStream.java:256)
        at org.apache.hadoop.hdfs.StripeReader.readChunk(StripeReader.java:293)
        at org.apache.hadoop.hdfs.StripeReader.readStripe(StripeReader.java:323)
        at org.apache.hadoop.hdfs.DFSStripedInputStream.readOneStripe(DFSStripedInputStream.java:318)
        at org.apache.hadoop.hdfs.DFSStripedInputStream.readWithStrategy(DFSStripedInputStream.java:391)
        at org.apache.hadoop.hdfs.DFSInputStream.read(DFSInputStream.java:828)
        at java.io.DataInputStream.read(DataInputStream.java:100)
        at org.apache.hadoop.io.IOUtils.copyBytes(IOUtils.java:94)
        at org.apache.hadoop.io.IOUtils.copyBytes(IOUtils.java:68)
        at org.apache.hadoop.io.IOUtils.copyBytes(IOUtils.java:129)
        at org.apache.hadoop.fs.shell.CommandWithDestination$TargetFileSystem.writeStreamToFile(CommandWithDestination.java:497)
        at org.apache.hadoop.fs.shell.CommandWithDestination.copyStreamToTarget(CommandWithDestination.java:419)
        at org.apache.hadoop.fs.shell.CommandWithDestination.copyFileToTarget(CommandWithDestination.java:354)
        at org.apache.hadoop.fs.shell.CommandWithDestination.processPath(CommandWithDestination.java:289)
        at org.apache.hadoop.fs.shell.CommandWithDestination.processPath(CommandWithDestination.java:274)
        at org.apache.hadoop.fs.shell.Command.processPaths(Command.java:331)
        at org.apache.hadoop.fs.shell.Command.processPathArgument(Command.java:303)
        at org.apache.hadoop.fs.shell.CommandWithDestination.processPathArgument(CommandWithDestination.java:269)
        at org.apache.hadoop.fs.shell.Command.processArgument(Command.java:285)
        at org.apache.hadoop.fs.shell.Command.processArguments(Command.java:269)
        at org.apache.hadoop.fs.shell.CommandWithDestination.processArguments(CommandWithDestination.java:240)
        at org.apache.hadoop.fs.shell.FsCommand.processRawArguments(FsCommand.java:120)
        at org.apache.hadoop.fs.shell.Command.run(Command.java:176)
        at org.apache.hadoop.fs.FsShell.run(FsShell.java:328)
        at org.apache.hadoop.util.ToolRunner.run(ToolRunner.java:76)
        at org.apache.hadoop.util.ToolRunner.run(ToolRunner.java:90)
        at org.apache.hadoop.fs.FsShell.main(FsShell.java:391)

然后自动通过RS算法恢复数据;

21/03/05 16:55:39 hdfs.DFSClient: refreshLocatedBlock for striped blocks, offset=0. Obtained block LocatedStripedBlock{BP-1016852637-192.168.1.192-1597819912196:blk_-9223372036854775792_4907534; getBlockSize()=2684354560; corrupt=false; offset=0; locs=[DatanodeInfoWithStorage[192.168.1.192:1019,DS-3b10bc66-c5c9-47f8-b4ea-99441fc5df04,DISK], DatanodeInfoWithStorage[192.168.1.196:1019,DS-c618bb83-03f2-4007-a3a9-cbd11eb2a15a,DISK], DatanodeInfoWithStorage[192.168.1.193:1019,DS-6554db41-d285-4f31-9fde-2017cc211b0c,DISK], DatanodeInfoWithStorage[192.168.1.188:1019,DS-09a7d2b0-6018-43af-8a63-6be8ba9217e6,DISK], DatanodeInfoWithStorage[192.168.1.199:1019,DS-b24cf7ce-1235-478f-884d-19cde4a03c9e,DISK], DatanodeInfoWithStorage[192.168.1.187:1019,DS-43b686d2-e83c-49a3-b8e5-64c4e5edcb53,DISK], DatanodeInfoWithStorage[192.168.1.195:1019,DS-eb42e686-1ca9-44c3-9891-868c67d9d1fa,DISK], DatanodeInfoWithStorage[192.168.1.194:1019,DS-15635fea-314c-4703-a7ab-6f81db1e52cd,DISK], DatanodeInfoWithStorage[192.168.1.198:1019,DS-73bc78d2-e218-40a9-ae9c-d24706a0bc31,DISK]]; indices=[0, 1, 2, 3, 4, 5, 6, 7, 8]}, idx=0
21/03/05 16:55:39 WARN hdfs.DFSClient: [DatanodeInfoWithStorage[192.168.1.192:1019,DS-3b10bc66-c5c9-47f8-b4ea-99441fc5df04,DISK]] are unavailable and all striping blocks on them are lost. IgnoredNodes = null
21/03/05 16:55:39 hdfs.DFSClient: refreshLocatedBlock for striped blocks, offset=0. Obtained block LocatedStripedBlock{BP-1016852637-192.168.1.192-1597819912196:blk_-9223372036854775792_4907534; getBlockSize()=2684354560; corrupt=false; offset=0; locs=[DatanodeInfoWithStorage[192.168.1.192:1019,DS-3b10bc66-c5c9-47f8-b4ea-99441fc5df04,DISK], DatanodeInfoWithStorage[192.168.1.196:1019,DS-c618bb83-03f2-4007-a3a9-cbd11eb2a15a,DISK], DatanodeInfoWithStorage[192.168.1.193:1019,DS-6554db41-d285-4f31-9fde-2017cc211b0c,DISK], DatanodeInfoWithStorage[192.168.1.188:1019,DS-09a7d2b0-6018-43af-8a63-6be8ba9217e6,DISK], DatanodeInfoWithStorage[192.168.1.199:1019,DS-b24cf7ce-1235-478f-884d-19cde4a03c9e,DISK], DatanodeInfoWithStorage[192.168.1.187:1019,DS-43b686d2-e83c-49a3-b8e5-64c4e5edcb53,DISK], DatanodeInfoWithStorage[192.168.1.195:1019,DS-eb42e686-1ca9-44c3-9891-868c67d9d1fa,DISK], DatanodeInfoWithStorage[192.168.1.194:1019,DS-15635fea-314c-4703-a7ab-6f81db1e52cd,DISK], DatanodeInfoWithStorage[192.168.1.198:1019,DS-73bc78d2-e218-40a9-ae9c-d24706a0bc31,DISK]]; indices=[0, 1, 2, 3, 4, 5, 6, 7, 8]}, idx=1
21/03/05 16:55:39 hdfs.DFSClient: refreshLocatedBlock for striped blocks, offset=0. Obtained block LocatedStripedBlock{BP-1016852637-192.168.1.192-1597819912196:blk_-9223372036854775792_4907534; getBlockSize()=2684354560; corrupt=false; offset=0; locs=[DatanodeInfoWithStorage[192.168.1.192:1019,DS-3b10bc66-c5c9-47f8-b4ea-99441fc5df04,DISK], DatanodeInfoWithStorage[192.168.1.196:1019,DS-c618bb83-03f2-4007-a3a9-cbd11eb2a15a,DISK], DatanodeInfoWithStorage[192.168.1.193:1019,DS-6554db41-d285-4f31-9fde-2017cc211b0c,DISK], DatanodeInfoWithStorage[192.168.1.188:1019,DS-09a7d2b0-6018-43af-8a63-6be8ba9217e6,DISK], DatanodeInfoWithStorage[192.168.1.199:1019,DS-b24cf7ce-1235-478f-884d-19cde4a03c9e,DISK], DatanodeInfoWithStorage[192.168.1.187:1019,DS-43b686d2-e83c-49a3-b8e5-64c4e5edcb53,DISK], DatanodeInfoWithStorage[192.168.1.195:1019,DS-eb42e686-1ca9-44c3-9891-868c67d9d1fa,DISK], DatanodeInfoWithStorage[192.168.1.194:1019,DS-15635fea-314c-4703-a7ab-6f81db1e52cd,DISK], DatanodeInfoWithStorage[192.168.1.198:1019,DS-73bc78d2-e218-40a9-ae9c-d24706a0bc31,DISK]]; indices=[0, 1, 2, 3, 4, 5, 6, 7, 8]}, idx=6

最终3个丢失的数据块被恢复,文件正常读取。

2.4 优点

  1. 相比副本存储方式大大降低了存储资源和IO资源的使用;
  2. 通过XOR和RS算法保证数据安全,有效解决允许范围内数据块破碎和丢失导致的异常;

2.5 缺点

  1. 恢复数据时需要去读其它数据块和奇偶校验码块数据,需要消耗IO和网络资源;
  2. EC算法编码和解密计算需要消耗CPU资源;
  3. 存储大数据,并且数据块较集中的节点运行作业负载会较高;
  4. EC文件不支持hflush, hsync, concat, setReplication, truncate, append 等操作。

3 副本存储和EC存储读写对比

写文件;

文件大小

三副本存储(秒)

EC存储(秒)

2.5G

6.492

9.093

5G

29.778

17.245

12G

65.156

30.989

读文件;

文件大小

三副本存储(秒)

EC存储(秒)

2.5G

4.915

4.755

5G

7.737

7.119

12G

14.493

13.775

4 总结

综合副本存储和EC存储优缺点,EC存储更适合用于存储备份数据和使用频率较少的非热点数据,副本存储更适用于存储需要追加写入和经常分析的热点数据。

参考文献

  • https://hadoop.apache.org/docs/r3.1.0/hadoop-project-dist/hadoop-common/release/3.0.0-alpha1/RELEASENOTES.3.0.0-alpha1.html - Apache Hadoop 3.0.0-alpha1 Release Notes
  • https://issues.apache.org/jira/browse/HDFS-7285 - Erasure Coding Support inside HDFS
  • https://issues.apache.org/jira/browse/HDFS-8030 - HDFS Erasure Coding Phase II – EC with contiguous layout
  • https://www.sciencedirect.com/topics/engineering/reed-solomon-code - Reed-Solomon Code