juliuszsompolski opened a new pull request #28671: URL: https://github.com/apache/spark/pull/28671
### What changes were proposed in this pull request? Timestamp literals in Spark are interpreted as timestamps in local timezone spark.sql.session.timeZone. If JDBC client is e.g. in TimeZone UTC-7, and sets spark.sql.session.timeZone to PST, and sends a query "SELECT timestamp '2020-05-20 12:00:00'", and the JVM timezone of the Spark cluster is e.g. UTC+2, then what currently happens is: * The timestamp literal in the query is interpreted as 12:00:00 UTC-7, i.e. 19:00:00 UTC. * When it's returned from the query, it is collected as a java.sql.Timestamp object with Dataset.collect(), and put into a Thriftserver RowSet. * Before sending it over the wire, the Timestamp is converted to String. This happens in explicitly in ColumnValue for RowBasedSet, and implicitly in ColumnBuffer for ColumnBasedSet (all non-primitive types are converted toString() there). The conversion toString uses JVM timezone, which results in a "21:00:00" (UTC+2) string representation. * The client JDBC application parses gets a "21:00:00" Timestamp back (in it's JVM timezone; if the JDBC application cares about the correct UTC internal value, it should set spark.sql.session.timeZone to be consistent with its JVM timezone) The problem is caused by the conversion happening in Thriftserver RowSet with the generic toString() function, instead of using HiveResults.toHiveString() that takes care of correct, timezone respecting conversions. This PR fixes it by converting the Timestamp values to String earlier, in SparkExecuteStatementOperation, using that function. This fixes SPARK-31861. Thriftserver also did not work spark.sql.datetime.java8API.enabled, because the conversions in RowSet expected an Timestamp object instead of Instant object. Using HiveResults.toHiveString() also fixes that. For this reason, we also convert Date values in SparkExecuteStatementOperation as well - so that HiveResults.toHiveString() handles LocalDate as well. This fixes SPARK-31859. Thriftserver also did not correctly set the active SparkSession. Because of that, configuration obtained using SQLConf.get was not the correct session configuration. This affected getting the correct spark.sql.session.timeZone. It is fixed by extending the use of SparkExecuteStatementOperation.withSchedulerPool to also set the correct active SparkSession. When the correct session is set, we also no longer need to maintain the pool mapping in a sessionToActivePool map. The scheduler pool can be just correctly retrieved from the session config. "withSchedulerPool" is renamed to "withLocalProperties" and moved into a mixin helper trait, because it should be applied with every operation. This fixes SPARK-31863. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org
