Apache Hudi 使用文件聚类功能 (Clustering) 解决小文件过多的问题

本文详细阐述了在 “批处理后，流处理之前” 进行文件 Clustering 操作的方法。该方法可以将众多小文件合并成数量极少的大文件，从而防止过多小文件的产生。

在批处理结束后进行 Clustering 主要涉及如下几个步骤，它们主要都是通过 spark-submit 命令完成的：

#### 批处理数据结束

首先用 bulk\_insert 方式运行批处理任务。注意下面的操作都是在批处理任务完成后，接流之前进行。

查看表相关的 hdfs，可以发现由于使用了 bulk\_insert 的方式写入数据，导致文件数量非常多，而每个文件的 Size 非常小。我们希望将每个分区的1000多个小文件聚合成几个大文件，以免造成不必要的查询和系统维护开销。

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ hdfs dfs -count /flk_hudi/chdrpf_hudi_test03/*           7            7           32637997 /flk_hudi/chdrpf_hudi_test03/.hoodie           1         1067          571117942 /flk_hudi/chdrpf_hudi_test03/1           1         1071          716513820 /flk_hudi/chdrpf_hudi_test03/2           1         1072          644997032 /flk_hudi/chdrpf_hudi_test03/3           1         1072          507397985 /flk_hudi/chdrpf_hudi_test03/4           1         1069          730774472 /flk_hudi/chdrpf_hudi_test03/5           1         1067          586561261 /flk_hudi/chdrpf_hudi_test03/6           1         1063          557377359 /flk_hudi/chdrpf_hudi_test03/7           1         1070          483416155 /flk_hudi/chdrpf_hudi_test03/8           1         1071          587965407 /flk_hudi/chdrpf_hudi_test03/A           1         1071          570651877 /flk_hudi/chdrpf_hudi_test03/B           1         1068          796163049 /flk_hudi/chdrpf_hudi_test03/C           1         1064          732633320 /flk_hudi/chdrpf_hudi_test03/D           1         1067          524777141 /flk_hudi/chdrpf_hudi_test03/E           1         1070          550302848 /flk_hudi/chdrpf_hudi_test03/F           1         1076          540059544 /flk_hudi/chdrpf_hudi_test03/G           1         1071          590094172 /flk_hudi/chdrpf_hudi_test03/H           1         1076          505755100 /flk_hudi/chdrpf_hudi_test03/I           1         1068          606771875 /flk_hudi/chdrpf_hudi_test03/J           1         1068          495261290 /flk_hudi/chdrpf_hudi_test03/K           1         1067          516964732 /flk_hudi/chdrpf_hudi_test03/L           1         1060          482056347 /flk_hudi/chdrpf_hudi_test03/M           1         1054          607625266 /flk_hudi/chdrpf_hudi_test03/N           1         1077          551989638 /flk_hudi/chdrpf_hudi_test03/O           1         1076          590537140 /flk_hudi/chdrpf_hudi_test03/P           1         1069          536362956 /flk_hudi/chdrpf_hudi_test03/Q           1         1072          559723804 /flk_hudi/chdrpf_hudi_test03/R           1         1067          546042696 /flk_hudi/chdrpf_hudi_test03/S           1         1059          528438508 /flk_hudi/chdrpf_hudi_test03/T           1         1063          518288413 /flk_hudi/chdrpf_hudi_test03/U           1         1070          543146873 /flk_hudi/chdrpf_hudi_test03/V           1         1066          532588113 /flk_hudi/chdrpf_hudi_test03/W           1         1069          494606809 /flk_hudi/chdrpf_hudi_test03/X           1         1079          527128056 /flk_hudi/chdrpf_hudi_test03/Y           1         1068          477378497 /flk_hudi/chdrpf_hudi_test03/Z           1         1075          471848267 /flk_hudi/chdrpf_hudi_test03/a`

查看当前 hdfs 路径下的文件个数。可以发现由于 bulk\_insert 导致小文件非常之多，这会显著影响查询的性能 (一次查询可能要做几千个 IO 操作)。

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ hdfs dfs -count /flk_hudi/chdrpf_hudi_test03/          43        37452        22269590565 /flk_hudi/chdrpf_hudi_test03`

#### Clustering

##### 配置清理策略

使用最简配置方法如下

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ cat /home/hadoop/hudi_clustering/clusteringjob.propertieshoodie.clustering.inline.max.commits=2hoodie.clustering.plan.strategy.max.num.groups=40`

添加高级配置项

`[hadoop@p0-tklfrna-tklrna-device02 ~]$ cat /home/hadoop/hudi_clustering/clusteringjob.properties   hoodie.clustering.inline=truehoodie.clustering.inline.max.commits=2hoodie.clustering.plan.strategy.max.num.groups=40hoodie.clustering.plan.strategy.target.file.max.bytes=1073741824hoodie.clustering.plan.strategy.max.bytes.per.group=2147483648hoodie.clustering.plan.strategy.small.file.limit=629145600`

##### Schedule

指定 Clustering 计划。计划制定完毕后 Hudi 对应 hdfs 的 Timeline 中会出现相应时间戳，以供执行计划。

`spark-submit \--master yarn \--class org.apache.hudi.utilities.HoodieClusteringJob \hdfs://nameservice1/utility_jars/hudi-utilities-bundle_2.12-0.10.0.jar \--schedule \--base-path hdfs://nameservice1/flk_hudi/chdrpf_hudi_test03 \--table-name chdrpf_hudi_test03 \--props file:///home/hadoop/hudi_clustering/clusteringjob.properties \--spark-memory 16g \> /home/hadoop/hudi_clustering/clusteringjob.log 2>&1`

查看 Hdfs 中的 Hudi 的 Timeline 获取时间戳。文件后缀为 `replacecommit.requested` 的时间戳即为我们需要的时间戳。复制我们需要的 `20220826105913373`，以便下一步粘贴。

`[hadoop@p0-tklfrna-tklrna-device02 ~]$ hdfs dfs -ls /flk_hudi/chdrpf_hudi_test03/.hoodie/Found 407 itemsdrwxr-xr-x   - hadoop supergroup          0 2022-08-26 10:10 /flk_hudi/chdrpf_hudi_test03/.hoodie/.auxdrwxr-xr-x   - hadoop supergroup          0 2022-08-26 14:53 /flk_hudi/chdrpf_hudi_test03/.hoodie/.temp-rw-r--r--   3 hadoop supergroup   18596070 2022-08-26 10:14 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101036547.commit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:10 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101036547.commit.requested-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:10 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101036547.inflight-rw-r--r--   3 hadoop supergroup   14041389 2022-08-26 10:16 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101404432.commit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:14 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101404432.commit.requested-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:14 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101404432.inflight...-rw-r--r--   3 hadoop supergroup    5685565 2022-08-26 10:59 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826105913373.replacecommit.requested...`

##### Execute

Clustering 执行需要使用刚才的时间戳配置 `--instant-time 20220826105913373` 于命令中即可执行。

`spark-submit \--master yarn \--class org.apache.hudi.utilities.HoodieClusteringJob \hdfs://nameservice1/utility_jars/hudi-utilities-bundle_2.12-0.10.0.jar \--instant-time 20220826105913373 \--base-path hdfs://nameservice1/flk_hudi/chdrpf_hudi_test03 \--table-name chdrpf_hudi_test03 \--props file:///home/hadoop/hudi_clustering/clusteringjob.properties \--spark-memory 16g \> /home/hadoop/hudi_clustering/clusteringjob_execution.log 2>&1`

##### 文件聚类完毕后

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ hdfs dfs -count /flk_hudi/chdrpf_hudi_test03/*           7           10           39759457 /flk_hudi/chdrpf_hudi_test03/.hoodie           1         1068          644693330 /flk_hudi/chdrpf_hudi_test03/1           1         1072          912384991 /flk_hudi/chdrpf_hudi_test03/2           1         1073          783040567 /flk_hudi/chdrpf_hudi_test03/3           1         1073          535431665 /flk_hudi/chdrpf_hudi_test03/4           1         1070          938545286 /flk_hudi/chdrpf_hudi_test03/5           1         1068          676230669 /flk_hudi/chdrpf_hudi_test03/6           1         1064          625387487 /flk_hudi/chdrpf_hudi_test03/7           1         1071          494572949 /flk_hudi/chdrpf_hudi_test03/8           1         1072          675599389 /flk_hudi/chdrpf_hudi_test03/A           1         1072          643710911 /flk_hudi/chdrpf_hudi_test03/B           1         1069         1056860522 /flk_hudi/chdrpf_hudi_test03/C           1         1065          940690081 /flk_hudi/chdrpf_hudi_test03/D           1         1068          563929957 /flk_hudi/chdrpf_hudi_test03/E           1         1071          606406555 /flk_hudi/chdrpf_hudi_test03/F           1         1077          589463777 /flk_hudi/chdrpf_hudi_test03/G           1         1072          682564783 /flk_hudi/chdrpf_hudi_test03/H           1         1077          529816271 /flk_hudi/chdrpf_hudi_test03/I           1         1069          712917512 /flk_hudi/chdrpf_hudi_test03/J           1         1069          514668751 /flk_hudi/chdrpf_hudi_test03/K           1         1068          550874973 /flk_hudi/chdrpf_hudi_test03/L           1         1061          495250431 /flk_hudi/chdrpf_hudi_test03/M           1         1055          716887761 /flk_hudi/chdrpf_hudi_test03/N           1         1078          612144859 /flk_hudi/chdrpf_hudi_test03/O           1         1077          679350316 /flk_hudi/chdrpf_hudi_test03/P           1         1070          586176818 /flk_hudi/chdrpf_hudi_test03/Q           1         1073          625760986 /flk_hudi/chdrpf_hudi_test03/R           1         1068          603042997 /flk_hudi/chdrpf_hudi_test03/S           1         1060          576062292 /flk_hudi/chdrpf_hudi_test03/T           1         1064          555764103 /flk_hudi/chdrpf_hudi_test03/U           1         1071          598050377 /flk_hudi/chdrpf_hudi_test03/V           1         1066          532588113 /flk_hudi/chdrpf_hudi_test03/W           1         1069          494606809 /flk_hudi/chdrpf_hudi_test03/X           1         1079          527128056 /flk_hudi/chdrpf_hudi_test03/Y           1         1068          477378497 /flk_hudi/chdrpf_hudi_test03/Z           1         1075          471848267 /flk_hudi/chdrpf_hudi_test03/a`

##### 运行清理

在进行完 Clustering 操作后，很多小文件都被合并进大文件了。由于 Hudi 不会主动删除过期和不必要的文件，因此需要利用手动清理策略来对过期文件进行清理删除。

清理策略的配置文件

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ cat /home/hadoop/hudi_clustering/hudi_cleaning.properties    # hudi_cleaning.properties  # When enabled, the cleaner table service is invoked immediately after each commit, to delete older file slices     hoodie.clean.automatic=true  # Only applies when hoodie.clean.automatic is turned on.      # When turned on runs cleaner async with writing, which can speed up overall write performance.   hoodie.clean.async=true  # # This policy has the effect of keeping N number of file versions irrespective of time.      # # This policy is useful when it is known how many MAX versions of the file does one want to keep at any given time.   # # hoodie.cleaner.policy=KEEP_LATEST_COMMITShoodie.cleaner.policy=KEEP_LATEST_COMMITS  # # Number of commits to retain, without cleaning.     # # This will be retained for num_of_commits * time_between_commits (scheduled).   # # hoodie.cleaner.commits.retained=3   # When KEEP_LATEST_FILE_VERSIONS cleaning policy is used,    # the minimum number of file slices to retain in each file group, during cleaning.hoodie.cleaner.commits.retained=1  # When set to true, cleaner also deletes the bootstrap base file when it's skeleton base file is cleaned.     hoodie.cleaner.delete.bootstrap.base.file=false   # Only if the log file size is greater than the threshold in bytes, the file group will be compacted.hoodie.commits.archival.batch=60hoodie.archive.merge.small.file.limit.bytes=104857600  # When set to true, compaction service is triggered after each write.      # While being simpler operationally, this adds extra latency on the write path.   hoodie.compact.inline=falsehoodie.parquet.small.file.limit=124857600hoodie.cleaner.parallelism=800  hoodie.cleaner.incremental.mode=true  # Archiving service moves older entries from timeline into an archived log after each write,      # to keep the metadata overhead constant, even as the table size growshoodie.keep.max.commits=3hoodie.keep.min.commits=2`

利用命令执行清理策略

`spark-submit \--class org.apache.hudi.utilities.HoodieCleaner \hdfs://nameservice1/utility_jars/hudi-utilities-bundle_2.12-0.10.0.jar \--props file:///home/hadoop/hudi_clustering/hudi_cleaning.properties \--target-base-path hdfs://nameservice1/flk_hudi/chdrpf_hudi_test03 \> /home/hadoop/hudi_clustering/clusteringjob_cleaning.log 2>&1`

##### 接流处理任务

此时，可以将流处理任务接至该 Hudi 表中。文件清理的效果会在 Hudi 接流后显现。

##### 清理后文件个数

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ hdfs dfs -count /flk_hudi/chdrpf_hudi_test03/*          39         2818           61047630 /flk_hudi/chdrpf_hudi_test03/.hoodie           1            5          295730057 /flk_hudi/chdrpf_hudi_test03/1           1            5          581449403 /flk_hudi/chdrpf_hudi_test03/2           1            5          541564433 /flk_hudi/chdrpf_hudi_test03/3           1            5          113526185 /flk_hudi/chdrpf_hudi_test03/4           1            5          819123981 /flk_hudi/chdrpf_hudi_test03/5           1            5          361258893 /flk_hudi/chdrpf_hudi_test03/6           1            4          205559110 /flk_hudi/chdrpf_hudi_test03/7           1            4           33721101 /flk_hudi/chdrpf_hudi_test03/8           1            5          352884732 /flk_hudi/chdrpf_hudi_test03/A           1            5          294248033 /flk_hudi/chdrpf_hudi_test03/B           1            5          771533591 /flk_hudi/chdrpf_hudi_test03/C           1            5          614827884 /flk_hudi/chdrpf_hudi_test03/D           1            5          157676833 /flk_hudi/chdrpf_hudi_test03/E           1            5          226004511 /flk_hudi/chdrpf_hudi_test03/F           1            5          198656601 /flk_hudi/chdrpf_hudi_test03/G           1            5          372307018 /flk_hudi/chdrpf_hudi_test03/H           1            5           97041611 /flk_hudi/chdrpf_hudi_test03/I           1            5          427390894 /flk_hudi/chdrpf_hudi_test03/J           1            5           78296341 /flk_hudi/chdrpf_hudi_test03/K           1            5          136428423 /flk_hudi/chdrpf_hudi_test03/L           1            5           53218521 /flk_hudi/chdrpf_hudi_test03/M           1            5          439899957 /flk_hudi/chdrpf_hudi_test03/N           1            5          242278011 /flk_hudi/chdrpf_hudi_test03/O           1            5          357549763 /flk_hudi/chdrpf_hudi_test03/P           1            5          200702230 /flk_hudi/chdrpf_hudi_test03/Q           1            5          265952714 /flk_hudi/chdrpf_hudi_test03/R           1            5          229783530 /flk_hudi/chdrpf_hudi_test03/S           1            5          191817537 /flk_hudi/chdrpf_hudi_test03/T           1            5          151138760 /flk_hudi/chdrpf_hudi_test03/U           1            5          221236895 /flk_hudi/chdrpf_hudi_test03/V           1         4112         2060894265 /flk_hudi/chdrpf_hudi_test03/W           1         4117         1910706738 /flk_hudi/chdrpf_hudi_test03/X           1         4169         2042792364 /flk_hudi/chdrpf_hudi_test03/Y           1         2221          995253322 /flk_hudi/chdrpf_hudi_test03/Z           1         1075          472877437 /flk_hudi/chdrpf_hudi_test03/a`

可以看到每个分区内的小文件已经被聚合成大文件，并随着流数据的进入，文件数量的增长速度也在合理范围内。

Ps: 我们把后几个分区作为对照组没有进行文件聚合。可以通过在 Clustering 的配置文件中调大 `hoodie.clustering.plan.strategy.max.num.groups=30` 的值来增加 SparkJob 的 parallelism 从而把所有分区涵盖进行，进行文件聚合。

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ hdfs dfs -count /flk_hudi/chdrpf_hudi_test03/          76        19050        17396389394 /flk_hudi/chdrpf_hudi_test03`

##### Timeline 观察

`20220826105913373.replacecommit`表示进行完毕聚类操作的时刻

`20220826114108591.clean`表示进行完毕清理操作的时刻

`20220826114317026.commit`表示进行完毕新数据写入操作的时刻

`[hadoop@p0-tklfrna-tklrna-device02 hudi_clustering]$ hdfs dfs -ls /flk_hudi/chdrpf_hudi_test03/.hoodieFound 30 itemsdrwxr-xr-x   - hadoop supergroup          0 2022-08-26 10:10 /flk_hudi/chdrpf_hudi_test03/.hoodie/.auxdrwxr-xr-x   - hadoop supergroup          0 2022-08-26 11:46 /flk_hudi/chdrpf_hudi_test03/.hoodie/.temp-rw-r--r--   3 hadoop supergroup   18596070 2022-08-26 10:14 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101036547.commit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:10 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101036547.commit.requested-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:10 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101036547.inflight-rw-r--r--   3 hadoop supergroup   14041389 2022-08-26 10:16 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101404432.commit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:14 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101404432.commit.requested-rw-r--r--   3 hadoop supergroup          0 2022-08-26 10:14 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826101404432.inflight-rw-r--r--   3 hadoop supergroup    1435895 2022-08-26 11:09 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826105913373.replacecommit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:03 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826105913373.replacecommit.inflight-rw-r--r--   3 hadoop supergroup    5685565 2022-08-26 10:59 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826105913373.replacecommit.requested-rw-r--r--   3 hadoop supergroup    1009885 2022-08-26 11:37 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826113342082.commit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:33 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826113342082.commit.requested-rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:33 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826113342082.inflight-rw-r--r--   3 hadoop supergroup    3811303 2022-08-26 11:40 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826113740364.commit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:37 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826113740364.commit.requested-rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:37 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826113740364.inflight-rw-r--r--   3 hadoop supergroup    2940587 2022-08-26 11:43 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114026452.commit-rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:40 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114026452.commit.requested-rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:40 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114026452.inflight-rw-r--r--   3 hadoop supergroup    5005100 2022-08-26 11:41 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114108591.clean-rw-r--r--   3 hadoop supergroup    4260649 2022-08-26 11:41 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114108591.clean.inflight-rw-r--r--   3 hadoop supergroup    4260649 2022-08-26 11:41 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114108591.clean.requested-rw-r--r--   3 hadoop supergroup    2867542 2022-08-26 11:46 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114317026.commit  -rw-r--r--   3 hadoop supergroup          0 2022-08-26 11:43 /flk_hudi/chdrpf_hudi_test03/.hoodie/20220826114317026.commit.requested  `