Iceberg unit tests, support Iceberg + nonhive catalogs, Iceberg Kryo Serializer #993
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -239,9 +239,15 @@ case object Iceberg extends Format { | |
| override def partitions(tableName: String, partitionColumns: Seq[String])(implicit | ||
| sparkSession: SparkSession): Seq[Map[String, String]] = { | ||
| sparkSession.sqlContext | ||
| .sql(s"SELECT partition FROM $tableName" ++ ".partitions") | ||
|
There was a problem hiding this comment. ooc does this work for regular hive tables? There was a problem hiding this comment. This is for iceberg, Hive support is here
Collaborator
Author
There was a problem hiding this comment. it should work for hive tables, and the internal target I'm hitting are more or less "regular hive tables". Iceberg abstracts itself from the catalog implementation, so as long as your iceberg has an interface to your catalog implementation, it will work. |
||
| .collect() | ||
| .map { row => | ||
| val partitionStruct = row.getStruct(0) | ||
| partitionStruct.schema.fieldNames.zipWithIndex.map { | ||
| case (fieldName, idx) => | ||
| fieldName -> partitionStruct.get(idx).toString | ||
| }.toMap | ||
| } | ||
| } | ||
|
|
||
| private def getIcebergPartitions(tableName: String, | ||
|
|
@@ -395,7 +401,14 @@ case class TableUtils(sparkSession: SparkSession) { | |
| rdd | ||
| } | ||
|
|
||
|
Collaborator
There was a problem hiding this comment. curious, does the old method not work for iceberg?
Collaborator
Author
There was a problem hiding this comment. Good question! It does work for non Iceberg tables IF your underlying catalog supports this operation. This line of code is querying the catalog directly, but the more idiomatic thing to do with Iceberg is to use its built in partition APIs, which will be agnostic to your underlying catalog https://iceberg.apache.org/docs/latest/spark-queries/#spark-queries (note here that this also documents the point I made that Iceberg doesn't support DSv1)
Collaborator
Author
There was a problem hiding this comment. I'm sure most setups use a catalog that works with DSv1, but ours does not. I have read that there's better pushdown in V2 sources but I can't really be a good source for that benchmark considering my setup doesn't work with V1
Collaborator
There was a problem hiding this comment. ok, this should be fine and they are equivalent anyways. |
||
| def tableExists(tableName: String): Boolean = { | ||
| try { | ||
| sparkSession.sql(s"DESCRIBE TABLE $tableName") | ||
| true | ||
| } catch { | ||
| case _: AnalysisException => false | ||
| } | ||
| } | ||
|
|
||
| def loadEntireTable(tableName: String): DataFrame = sparkSession.table(tableName) | ||
|
|
||
|
|
@@ -973,17 +986,38 @@ case class TableUtils(sparkSession: SparkSession) { | |
| partitions: Seq[String], | ||
| partitionColumn: String = partitionColumn, | ||
| subPartitionFilters: Map[String, String] = Map.empty): Unit = { | ||
| // TODO this is using datasource v1 semantics, which won't be compatible with non-hive catalogs | ||
|
Collaborator
There was a problem hiding this comment. could you explain more on the dsv1 and dsv2? Do you have a pointer?
Collaborator
Author
There was a problem hiding this comment. Sure! It's largely historical, but the tl;dr was that at some point the way you build a datasource connector was redone to support more sink formats with better performance. This is a good article: https://blog.madhukaraphatak.com/spark-datasource-v2-part-1 and this pdf: https://issues.apache.org/jira/browse/SPARK-15689. The most notable reason this is coming up is that Iceberg is not integrated with DataSourceV1. |
||
| // notably, the unit test iceberg integration uses hadoop because of | ||
| // https://github.com/apache/iceberg/issues/7847 | ||
| if (partitions.nonEmpty && tableExists(tableName)) { | ||
| val dropSql = tableFormatProvider.readFormat(tableName) match { | ||
| // really this is Dsv1 vs Dsv2, not hive vs iceberg, | ||
| // but we break this way since only Iceberg is migrated to Dsv2 | ||
| case Iceberg => | ||
| // Build WHERE clause: (ds='2024-05-01' OR ds='2024-05-02') [AND k='v' AND …] | ||
| val mainPred = partitions | ||
| .map(p => s"$partitionColumn='${p}'") | ||
| .mkString("(", " OR ", ")") | ||
|
|
||
| val extraPred = subPartitionFilters | ||
| .map { case (k, v) => s"$k='${v}'" } | ||
| .mkString(" AND ") | ||
|
|
||
| val where = Seq(mainPred, extraPred).filter(_.nonEmpty).mkString(" AND ") | ||
|
|
||
| s"DELETE FROM $tableName WHERE $where" | ||
| case _ => | ||
abbywh marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| val partitionSpecs = partitions | ||
| .map { partition => | ||
| val mainSpec = s"$partitionColumn='$partition'" | ||
| val specs = mainSpec +: subPartitionFilters.map { | ||
| case (key, value) => s"${key}='${value}'" | ||
| }.toSeq | ||
| specs.mkString("PARTITION (", ",", ")") | ||
| } | ||
| .mkString(",") | ||
| s"ALTER TABLE $tableName DROP IF EXISTS $partitionSpecs" | ||
| } | ||
| sql(dropSql) | ||
| } else { | ||
| logger.info(s"$tableName doesn't exist, please double check before drop partitions") | ||
|
|
||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
how is working with metals relative to intellij? does the debugger work as well?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It's really good actually. The debugged worked out of the box, I found it comparable to IntelliJ overall.
I'd recommend it to anyone who has remote dev boxes since VSCode's integration is far better in my experience. All the tests run a lot faster and I got in way more dev cycles. I probably would only recommend over IntelliJ with a dev box though.