Handling constant folding errors

[pramodsatya]

Description

In the FusionNext (SPI) project, we are trying to make the Presto Native engine more correct by doing constant folding with the native worker code. Right now, the co-ordinator uses Presto Java code for constant folding. This could have compatibility issues with the native engine. When investigating this work we saw an issue for cases where constant folding evaluations could throw exceptions.

Velox constant folding code is implemented by the function tryFoldIfConstant. If constant folding results in a VeloxUserError, such as Divide with zero exception for say divide(0, 0), the expression compiler in Velox suppresses this error and returns the input expression as is. The runtime evaluates the expression to throw the exception thereafter.

Presto has a different behavior. Any exceptions encountered during constant folding are replaced by a FailFunction expression (refer to function processWithExceptionHandling). This FailFunction could be used to trigger further optimizations in the co-ordinator.

For evaluating constant folding with the native worker, we introduced a new Native Worker REST endpoint which takes as input a set of Presto expressions and returns constant folded versions of them to the caller. The NativeExpressionOptimizer prestodb/presto#24126 which does the conversion invokes the Velox ExprCompiler to compile constants. As a result this code returns the original expression if the constant folding can throw an exception. The expression is evaluated during runtime and it throws an exception then. While this still results in correct behavior, we would like to retain the Presto optimization with FailFunction to enable more optimizations.

We propose extending the Velox Expression Compiler to aid this conversion. We want to add an option where the Expression Compiler could substitute a FailFunction for the cases where the constant expression evaluation throws. FailFunction could provide more context about the constant folding error, such as an ErrorCode and ErrorMessage. We will trigger this behavior based on a specific config and retain the default behavior to return the original expression.
#12376 models the code we would like.

This code change really simplifies any expression manipulation in the Prestissimo code for NativeExpressionOptimizer as we would otherwise need to re-implement much of the Velox expression compilation there.

Co-authored by: @aditi-pandit @czentgr

[aditi-pandit]

@pedroerp @mbasmanova @bikramSingh91 : We encountered this issue when attempting constant folding with Prestissimo native side-car. Would be great to hear your thoughts.

[mbasmanova]

@pramodsatya Sounds reasonable overall.

While this still results in correct behavior, we would like to retain the Presto optimization with FailFunction to enable more optimizations.

Would you elaborate on the types of optimizations this enables? A few examples would be helpful. Thanks.

[pramodsatya]

Thanks @mbasmanova. Currently the Presto optimizer does not use FailFunction but it could be used to trigger optimizations in the future. Such as consider the following switch conditional expression:

CASE 5 WHEN 1 THEN 1 WHEN 0 / 0 THEN 2 END

The first case here will return false and the second case will throw an exception, which would now be detected by the FailFunction. So at query planning we could replace this switch expression with a FailFunction. The primary goal currently is to achieve consistency with Presto when constant folding expressions with the native engine (via the sidecar).

[mbasmanova]

@pramodsatya In this example, we replace constant expression CASE.... with a simpler expression fail("...error message.."), right? This is fine, but it is not clear to me how this is helping optimizer the query overall. Would you elaborate some more?

[PingLiuPing]

This code change really simplifies any expression manipulation in the Prestissimo code for NativeExpressionOptimizer as we would otherwise need to re-implement much of the Velox expression compilation there.

@pramodsatya Have same questions with Maria, suppose velox tryFoldIfConstant return FailFunction instead of the same original expression, how would this simplifiy the expression evaluation/manipulation in Prestissimo? The experssion needs to be evaluated still. Can you please explain more what does expression manipulation mean and how a FailFunction could help here? Thank you.