Details
Major Description
The following returns the wrong answer:
set spark.sql.codegen.wholeStage=false; set spark.sql.codegen.factoryMode=NO_CODEGEN; select max(col1), max(col2) from values (cast(null as decimal(27,2)), cast(null as decimal(27,2))), (cast(77.77 as decimal(27,2)), cast(245.00 as decimal(27,2))) as data(col1, col2); +---------+---------+ |max(col1)|max(col2)| +---------+---------+ |null |239.88 | +---------+---------+
This is because InterpretedMutableProjection inappropriately uses InternalRow#setNullAt to set null for decimal types with precision > Decimal.MAX_LONG_DIGITS.
The path to corruption goes like this:
Unsafe buffer at start:
offset/len for offset/len for
1st decimal 2nd decimal
offset: 0 8 16 (0x10) 24 (0x18) 32 (0x20)
data: 0300000000000000 0000000018000000 0000000028000000 0000000000000000 0000000000000000 0000000000000000 0000000000000000
When processing the first incoming row ([null, null]), InterpretedMutableProjection calls setNullAt for the decimal types. As a result, the pointers to the storage areas for the two decimals in the variable length region get zeroed out.
Buffer after projecting first row (null, null):
offset/len for offset/len for
1st decimal 2nd decimal
offset: 0 8 16 (0x10) 24 (0x18) 32 (0x20)
data: 0300000000000000 0000000000000000 0000000000000000 0000000000000000 0000000000000000 0000000000000000 0000000000000000
When it's time to project the second row into the buffer, UnsafeRow#setDecimal uses the zero offsets, which causes UnsafeRow#setDecimal to overwrite the null-tracking bit set with decimal data:
null-tracking
bit area
offset: 0 8 16 (0x10) 24 (0x18) 32 (0x20)
data: 5db4000000000000 0000000000000000 0200000000000000 0000000000000000 0000000000000000 0000000000000000 0000000000000000
The null-tracking bit set is overwritten with 239.88 (0x5db4) rather than 245.00 (0x5fb4) because setDecimal indirectly calls setNotNullAt(1), which turns off the null-tracking bit associated with the field at index 1.
In addition, the decimal at field index 0 is now null because of the corruption of the null-tracking bit set.
When a decimal type with precision > Decimal.MAX_LONG_DIGITS is null, InterpretedMutableProjection should write a null Decimal value rather than call setNullAt (see.)
This bug could get exercised during codegen fallback. Take for example this case where I forced codegen to fail for the Greatest expression:
spark-sql> select max(col1), max(col2) from values (cast(null as decimal(27,2)), cast(null as decimal(27,2))), (cast(77.77 as decimal(27,2)), cast(245.00 as decimal(27,2))) as data(col1, col2); 22/12/05 08:18:54 ERROR CodeGenerator: failed to compile: org.codehaus.commons.compiler.CompileException: File 'generated.java', Line 58, Column 1: ';' expected instead of 'if' org.codehaus.commons.compiler.CompileException: File 'generated.java', Line 58, Column 1: ';' expected instead of 'if' at org.codehaus.janino.TokenStreamImpl.compileException(TokenStreamImpl.java:362) at org.codehaus.janino.TokenStreamImpl.read(TokenStreamImpl.java:149) at org.codehaus.janino.Parser.read(Parser.java:3787) ... 22/12/05 08:18:56 WARN MutableProjection: Expr codegen error and falling back to interpreter mode java.util.concurrent.ExecutionException: org.codehaus.commons.compiler.CompileException: File 'generated.java', Line 43, Column 1: failed to compile: org.codehaus.commons.compiler.CompileException: File 'generated.java', Line 43, Column 1: ';' expected instead of 'boolean' at com.google.common.util.concurrent.AbstractFuture$Sync.getValue(AbstractFuture.java:306) at com.google.common.util.concurrent.AbstractFuture$Sync.get(AbstractFuture.java:293) at org.apache.spark.sql.catalyst.expressions.codegen.CodeGenerator$$anon$1.load(CodeGenerator.scala:1583) at org.apache.spark.sql.catalyst.expressions.codegen.CodeGenerator$$anon$1.load(CodeGenerator.scala:1580) at com.google.common.cache.LocalCache$LoadingValueReference.loadFuture(LocalCache.java:3599) at com.google.common.cache.LocalCache$Segment.loadSync(LocalCache.java:2379) ... 36 more ... NULL 239.88 <== incorrect result, should be (77.77, 245.00) Time taken: 6.132 seconds, Fetched 1 row(s) spark-sql>
