[SPARK-21660] Yarn ShuffleService failed to start when the chosen directory become read-only

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 2.1.1
• Component/s: Shuffle, Spark Core, YARN
• Assignee: Unassigned
• Reporter: ShuMing Li
• Created: 08/Aug/17 03:50
• Updated: 12/Dec/22 18:11
• Resolved: 23/Aug/17 04:02

Description

Background

In our production environment，disks corrupt to `read-only` status almost once a month. Now the strategy of Yarn ShuffleService which chooses an available directory(disk) to store Shuffle info(DB) is as below(https://github.com/apache/spark/blob/master/common/network-yarn/src/main/java/org/apache/spark/network/yarn/YarnShuffleService.java#L340)：
1. If NameNode's recoveryPath not empty and shuffle DB exists in the recoveryPath, return the recoveryPath;
2. If recoveryPath empty and shuffle DB exists in `yarn.nodemanager.local-dirs`, set recoveryPath as the existing DB path and return the path;
3. If recoveryPath not empty(shuffle DB not exists in the path) and shuffle DB exists in `yarn.nodemanager.local-dirs`， mv the existing shuffle DB to recoveryPath and return the path;
4. If all above don't hit, we choose the first disk of `yarn.nodemanager.local-dirs`as the recoveryPath;

All above strategy don't consider the chosen disk(directory) is writable or not, so in our environment we meet such exception:

2017-06-25 07:15:43,512 ERROR org.apache.spark.network.util.LevelDBProvider: error opening leveldb file /mnt/dfs/12/yarn/local/registeredExecutors.ldb. Creating new file, will not be able to recover state for existing applications
at org.apache.spark.network.util.LevelDBProvider.initLevelDB(LevelDBProvider.java:48)
at org.apache.spark.network.shuffle.ExternalShuffleBlockResolver.<init>(ExternalShuffleBlockResolver.java:116)
at org.apache.spark.network.shuffle.ExternalShuffleBlockResolver.<init>(ExternalShuffleBlockResolver.java:94)
at org.apache.spark.network.shuffle.ExternalShuffleBlockHandler.<init>(ExternalShuffleBlockHandler.java:66)
at org.apache.spark.network.yarn.YarnShuffleService.serviceInit(YarnShuffleService.java:167)
2017-06-25 07:15:43,514 WARN org.apache.spark.network.util.LevelDBProvider: error deleting /mnt/dfs/12/yarn/local/registeredExecutors.ldb
2017-06-25 07:15:43,515 INFO org.apache.hadoop.service.AbstractService: Service spark_shuffle failed in state INITED; cause: java.io.IOException: Unable to create state store
at org.apache.spark.network.util.LevelDBProvider.initLevelDB(LevelDBProvider.java:77)
at org.apache.spark.network.shuffle.ExternalShuffleBlockResolver.<init>(ExternalShuffleBlockResolver.java:116)
at org.apache.spark.network.shuffle.ExternalShuffleBlockResolver.<init>(ExternalShuffleBlockResolver.java:94)
at org.apache.spark.network.shuffle.ExternalShuffleBlockHandler.<init>(ExternalShuffleBlockHandler.java:66)
at org.apache.spark.network.yarn.YarnShuffleService.serviceInit(YarnShuffleService.java:167)
at org.apache.spark.network.util.LevelDBProvider.initLevelDB(LevelDBProvider.java:75)

Consideration

1. For many production environment, `yarn.nodemanager.local-dirs` always has more than 1 disk, so we can make a better chosen strategy to avoid the problem above;
2. Can we add a strategy to check the DB directory we choose is writable, so avoid the problem above?

Issue Links

• duplicates: SPARK-17321 YARN shuffle service should use good disk from yarn.nodemanager.local-dirs