bersprockets opened a new pull request, #39349:
URL: https://github.com/apache/spark/pull/39349
### What changes were proposed in this pull request?
Change `InterpretedUnsafeProjection#getElementSize` to choose the
appropriate element size for the underlying SQL type of a UDT, rather than
simply using the the default size of the underlying SQL type.
### Why are the changes needed?
Consider this query:
```
// create a file of vector data
import org.apache.spark.ml.linalg.{DenseVector, Vector}
case class TestRow(varr: Array[Vector])
val values = Array(0.1d, 0.2d, 0.3d)
val dv = new DenseVector(values).asInstanceOf[Vector]
val ds = Seq(TestRow(Array(dv, dv))).toDS
ds.coalesce(1).write.mode("overwrite").format("parquet").save("vector_data")
// this works
spark.read.format("parquet").load("vector_data").collect
sql("set spark.sql.codegen.wholeStage=false")
sql("set spark.sql.codegen.factoryMode=NO_CODEGEN")
// this will get an error
spark.read.format("parquet").load("vector_data").collect
```
The failures vary, incuding
* `VectorUDT` attempting to deserialize to a `SparseVector` (rather than a
`DenseVector`)
* negative array size (for one of the nested arrays)
* JVM crash (SIGBUS error).
This is because `InterpretedUnsafeProjection` initializes the outer-most
array writer with an element size of 17 (the size of the UDT's underlying
struct), rather than an element size of 8, which would be appropriate for an
array of structs.
When the outer-most array is later accessed, `UnsafeArrayData` assumes an
element size of 8, so it picks up a garbage offset/size tuple for the second
element.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
New unit test.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
