[
https://issues.apache.org/jira/browse/SPARK-34298?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17278846#comment-17278846
]
Steve Loughran commented on SPARK-34298:
----------------------------------------
root dirs are special in that they always exist. Normally apps like spark and
hive don't notice this as nobody ever runs jobs which write to the base of
file:// or hdfs:// ; object stores are special there.
You might the commit algorithms get a bit confused too. In which case:
* fixes for the s3a committers welcome;
* there is a serialized test phase in hadoop-aws where we do stuff against the
root dir;
* and a PR for real integration tests can go in
https://github.com/hortonworks-spark/cloud-integration
* anything related to the classic MR committer will be rejected out of fear of
going near it; not safe for s3 anyway
Otherwise: workaround is to write into a subdir. Sorry
> SaveMode.Overwrite not usable when using s3a root paths
> --------------------------------------------------------
>
> Key: SPARK-34298
> URL: https://issues.apache.org/jira/browse/SPARK-34298
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 3.1.2
> Reporter: cornel creanga
> Priority: Minor
>
> SaveMode.Overwrite does not work when using paths containing just the root eg
> "s3a://peakhour-report". To reproduce the issue (an s3 bucket + credentials
> are needed):
> {color:#0033b3}val {color}{color:#000000}out {color}=
> {color:#067d17}"s3a://peakhour-report"{color}
> {color:#0033b3}val {color}{color:#000000}sparkContext{color}:
> {color:#000000}SparkContext {color}=
> {color:#000000}SparkContext{color}.getOrCreate()
> {color:#0033b3}val {color}{color:#000000}someData {color}=
> {color:#871094}Seq{color}(Row({color:#1750eb}24{color},
> {color:#067d17}"mouse"{color}))
> {color:#0033b3}val {color}{color:#000000}someSchema {color}=
> {color:#871094}List{color}(StructField({color:#067d17}"age"{color},
> {color:#000000}IntegerType{color},
> {color:#0033b3}true{color}),StructField({color:#067d17}"word"{color},
> {color:#000000}StringType{color},{color:#0033b3}true{color}))
> {color:#0033b3}val {color}{color:#000000}someDF {color}=
> {color:#871094}spark{color}.createDataFrame(
>
> {color:#871094}spark{color}.sparkContext.parallelize({color:#000000}someData{color}),StructType({color:#000000}someSchema{color}))
> {color:#000000}sparkContext{color}.hadoopConfiguration.set({color:#067d17}"fs.s3a.access.key"{color},
> accessK{color:#000000}ey{color}))
> {color:#000000}sparkContext{color}.hadoopConfiguration.set({color:#067d17}"fs.s3a.secret.key"{color},
> {color:#000000}secretKey{color}))
> {color:#000000}sparkContext{color}.hadoopConfiguration.set({color:#067d17}"fs.s3a.aws.credentials.provider"{color},
>
> {color:#067d17}"org.apache.hadoop.fs.s3a.SimpleAWSCredentialsProvider"{color})
> {color:#000000}sparkContext{color}.hadoopConfiguration.set({color:#067d17}"fs.s3a.impl"{color},
> {color:#067d17}"org.apache.hadoop.fs.s3a.S3AFileSystem"{color})
> {color:#000000}someDF{color}.write.format({color:#067d17}"parquet"{color}).partitionBy({color:#067d17}"age"{color}).mode({color:#000000}SaveMode{color}.{color:#871094}Overwrite{color})
> .save({color:#000000}out{color})
>
> Error stacktrace:
> Exception in thread "main" java.lang.IllegalArgumentException: Can not create
> a Path from an empty string
> at org.apache.hadoop.fs.Path.checkPathArg(Path.java:168)[....]
> at org.apache.hadoop.fs.Path.suffix(Path.java:446)
> at
> org.apache.spark.sql.execution.datasources.InsertIntoHadoopFsRelationCommand.deleteMatchingPartitions(InsertIntoHadoopFsRelationCommand.scala:240)
>
> If you change out from {color:#0033b3}val {color}{color:#000000}out {color}=
> {color:#067d17}"s3a://peakhour-report"{color} to {color:#0033b3}val
> {color}{color:#000000}out {color}=
> {color:#067d17}"s3a://peakhour-report/folder" {color:#172b4d}the code
> works.{color}{color}
> {color:#067d17}{color:#172b4d}There are two problems in the actual code from
> InsertIntoHadoopFsRelationCommand.deleteMatchingPartitions: {color}{color}
> {color:#067d17}{color:#172b4d}a) it uses org.apache.hadoop.fs.Path.suffix
> method that doesn't work on root paths
> {color}{color}
> {color:#067d17}{color:#172b4d}b) it tries to delete the root folder directly
> (in our case the s3 bucket name) and this is prohibited (in the S3AFileSystem
> class){color}{color}
> {color:#067d17}{color:#172b4d}I think that there are two
> choices:{color}{color}
> {color:#067d17}{color:#172b4d}a) throw an explicit error when using overwrite
> mode for root folders {color}{color}
> {color:#067d17}{color:#172b4d}b)fix the actual issue. don't use the
> Path.suffix method and change the clean up code from
> InsertIntoHadoopFsRelationCommand.deleteMatchingPartitions to list the root
> folder content and delete the entries one by one.{color}{color}
> I can provide a patch for both choices, assuming that they make sense.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
