not panic when given all nones to FixedSizeBinaryArray::try_from_sparse_iter

[frolovdev]

Which issue does this PR close?

Closes #1390

[tustvold]

I'm really not sure about this, perhaps @alamb can weigh in here, skipping schema checks just because the arrays are empty feels likely to cause surprise down the line.

If we do anything in this vein, I think it would be better to special case just empty fixed size arrays with fixed size length 0, and allow them to compare equal to other fixed size arrays of the same type. This would both make the intent clear and limit the potential blast radius.

My personal preference would be to make the try_from_sparse_iter method take the length of the arrays as an argument, rather than adding workarounds in other parts of the codebase.

[HaoYang670]

My personal preference would be to make the try_from_sparse_iter method take the length of the arrays as an argument, rather than adding workarounds in other parts of the codebase

I personally agree with @tustvold. We could add an argument value_len: Option<usize>, such as (not tested):

    pub fn try_from_sparse_iter<T, U>(mut iter: T, value_len: Option<usize>) -> Result<Self>
    where
        T: Iterator<Item = Option<U>>,
        U: AsRef<[u8]>,

And the function will return an error if iter is all-nones and value_len is None. Maybe we could return something such as (not tested):

ArrowError::InvalidArgumentError("cannot infer the value length of the array")

[alamb]

My personal preference would be to make the try_from_sparse_iter method take the length of the arrays as an argument, rather than adding workarounds in other parts of the codebase

I think this sounds like the best of several non ideal options.

[tustvold]

Marking as draft to make it easier to see what is awaiting review, please feel free to unmark once ready for re-review 😄

[tustvold]

Closing this as #3054 has added support for this