handling overflow for the integer types

[kmitchener]

Is your feature request related to a problem or challenge? Please describe what you are trying to do. A clear and concise description of what the problem is. Ex. I'm always frustrated when [...] (This section helps Arrow developers understand the context and why for this feature, in addition to the what)

Describe the solution you'd like A clear and concise description of what you want to happen.

I'm opening this issue to get consensus on what the desired DataFusion behavior should be when overflowing numeric types in DataFusion. All tests below done on master as of time of issue creation.

Current situation for overflow:

DataType	Test SQL	DataFusion (release)	Postgres
Int8	select 127::tinyint + 1::tinyint;	wraps	-
Int16	select 32767::smallint + 1::smallint;	wraps	ERROR: smallint out of range
Int32	select 2147483647::int + 1::int;	wraps	ERROR: integer out of range
Int64	select 9223372036854775807::bigint + 1::bigint;	wraps	ERROR: bigint out of range
UInt8	select 255::tinyint unsigned + 1::tinyint unsigned;	wraps	-
UInt16	select 65535::smallint unsigned + 1::smallint unsigned;	wraps	-
UInt32	select 4294967295::int unsigned + 1::int unsigned;	wraps	-
UInt64	select power(2,64)::bigint unsigned;	wraps	-

Current situation for attempting to cast an oversized number:

DataType	Test SQL	DataFusion (release)	Postgres
Int8	select 128::tinyint;	null	-
Int16	select 32768::smallint;	null	ERROR: smallint out of range
Int32	select 2147483648::int;	null	ERROR: integer out of range
Int64	select 9223372036854775808::bigint;	null	ERROR: bigint out of range
UInt8	select 256::tinyint unsigned;	null	-
UInt16	select 65536::smallint unsigned;	null	-
UInt32	select 4294967296::int unsigned;	null	-
UInt64	select 18446744073709551615::bigint unsigned;	null for values even less than 2^64. some weird behavior here.	-

I think the behavior between casting a "too big" number, and overflowing should be the same.

My proposal would be to make 2 changes:

• return an error during overflow situations, rather than wrapping around silently. principle of least surprise. I believe most databases throw errors in case of overflow and users would be surprised if DF silently returns "bad" data.
• we should make cast operations on out of range data (over- or under-sized) return an error

My proposals are based on years of Oracle and Postgres use though, I have no Spark experience. What other thoughts and opinions are out there? How does Spark behave in these cases?

Describe alternatives you've considered A clear and concise description of any alternative solutions or features you've considered.

Additional context Add any other context or screenshots about the feature request here.

[liukun4515]

@kmitchener @alamb @andygrove I think this is an issue about behavior when overflow. We can do more investigation about this, and decide the behavior by default. We also can change the behavior with the config or option.

[alamb]

I don't have a huge preference -- when in doubt I think we have tried to follow the postgres semantics, for consistency.

In terms of checking for overflows, etc I would also say we should try and avoid slowing things down too much, if possible

[alamb]

My proposal would be to make 2 changes:

I think those proposals are very reasonable

[liukun4515]

My proposals are based on years of Oracle and Postgres use though, I have no Spark experience. What other thoughts and opinions are out there? How does Spark behave in these cases?

Like cast, if we convert a value to another type which is overflow, the default result is NULL.

For the mathematical operations, we should add the option for that.

I think the two behavior is ok, but we should make them consistent.

@kmitchener

In the spark, If we don't set the special parameter, spark will not throw the error, and just return the wrapping value.

You can try it.

the doc ref in the spark: https://spark.apache.org/docs/latest/sql-ref-ansi-compliance.html https://spark.apache.org/docs/latest/sql-ref-ansi-compliance.html#arithmetic-operations

cc @alamb

[alamb]

An option to control the behavior also seems reasonable to be (though I suspect it would add some non trivial complexity to the implementation, so perhaps we can only do this if/when a user has a compelling usecase 🤔 )