Is your feature request related to a problem or challenge?
A common pattern is to GROUP BY a unique id and then get one arbitrary value of another column per group (for example, any tag value for each id). Which value comes back doesn't matter, so ordering is irrelevant.
DataFusion has no function for this today, and the workarounds are heavier than needed:
first_value(x) / last_value(x) can do this, but they do extra work any_value doesn't need.min(x) / max(x) are often used instead, but they don't really mean "any value", and they still compare every row of every batch.
What's missing is a function whose state is just "one value, captured once", so that once a group has a value, later rows for that group cost nothing.
ANY_VALUE is also a standard function (SQL:2023) and is already supported by spark, duckdb, postgres, and others.
Describe the solution you'd like
Add a built-in any_value(expression) aggregate that returns an arbitrary value from each group, ignoring NULLs (it returns NULL only when every value in the group is NULL).
Describe alternatives you've considered
No response
Additional context
ANY_VALUE is a standard aggregate (SQL:2023) and is widely supported, all returning an arbitrary non-NULL value from the group:
NULL handling varies slightly across engines; this proposal follows the ignore-NULLs behavior shared by the above (a NULL is returned only when every value in the group is NULL). RESPECT NULLS | IGNORE NULLS could be added later if needed.
Is your feature request related to a problem or challenge?
A common pattern is to
GROUP BYa unique id and then get one arbitrary value of another column per group (for example, any tag value for each id). Which value comes back doesn't matter, so ordering is irrelevant.DataFusion has no function for this today, and the workarounds are heavier than needed:
first_value(x)/last_value(x)can do this, but they do extra workany_valuedoesn't need.min(x)/max(x)are often used instead, but they don't really mean "any value", and they still compare every row of every batch.What's missing is a function whose state is just "one value, captured once", so that once a group has a value, later rows for that group cost nothing.
ANY_VALUEis also a standard function (SQL:2023) and is already supported by spark, duckdb, postgres, and others.Describe the solution you'd like
Add a built-in
any_value(expression)aggregate that returns an arbitrary value from each group, ignoring NULLs (it returns NULL only when every value in the group is NULL).Describe alternatives you've considered
No response
Additional context
ANY_VALUEis a standard aggregate (SQL:2023) and is widely supported, all returning an arbitrary non-NULL value from the group:NULL handling varies slightly across engines; this proposal follows the ignore-NULLs behavior shared by the above (a NULL is returned only when every value in the group is NULL).
RESPECT NULLS | IGNORE NULLScould be added later if needed.