Copy-On-Write Support to Support Mutating/Updating Arrays in Place

[tustvold]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.

Certain use-cases may benefit from being able to mutate arrays in place, without copying them. Fortunately we have most of the pieces to support this, it just requires some plumbing work.

Describe the solution you'd like

A high-level outline can be found below:

• Add COW conversion from Buffer to MutableBuffer likely using Arc::try_unwrap
• Add COW conversion from T: Array to their corresponding builder, i.e. PrimitiveArray::into_builder
• Add conversion from ArrayRef to Arc<T> - https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=abb7fcf97e33ba83d44b17f6c89e8f0b

It should then be possible to do something like (likely encapsulated in a nicer interface)

fn cow_builder(col: ArrayRef) -> Int32Builder {
    let col: Arc<Int32Array> = downcast_array(col).unwrap();
    let col: Int32Array = Arc::try_unwrap(col).unwrap_or_else(|_| col.as_ref().clone());
    col.into_builder()
}

Additional context

Arrow2 recently merged a form of support for this.

[alamb]

Related arrow2 code: Example https://github.com/jorgecarleitao/arrow2/blob/main/examples/cow.rs#L23

Looks like it was added by @jorgecarleitao in jorgecarleitao/arrow2#1042 (which was a bit hard to find as it is described in terms of the implementation rather than the end user API change) 👍

[avantgardnerio]

Instead of copying-on-write (or maybe in addition to), how would we feel about an AppendOnlyArrayBuilder that never calls any method that would cause MutableBuffer to realloc() and move any pointers?

If we had that, we could have AppendableRecordBatchs, and a borrow() -> &RecordBatch (as opposed to .finish()) that would allow us to append (but not mutate) data, while still being able to do all the fun stuff we can do with RecordBatches today (query in DataFusion, export to parquet, etc).

[tustvold]

I think that is a separate feature, would you mind creating a new issue along with the intended use-case? I suspect it concerns implementing some sort of append only memtable, something I have thoughts on having implemented something similar for IOx (spoiler mem copies are far from the bottleneck).

[avantgardnerio]

@tustvold @alamb @andygrove I filed an issue with PoC implementation: #3142

[viirya]

For COW, the high-level outline I think we have corresponding implementation now. Is there anything I miss?

[tustvold]

Theoretically it might be nice to provide into_buffer for ByteArray, but otherwise I think we can close this

[viirya]

For into_buffer, do you mean a MutableBuffer of a ByteArray so we can write ByteArrayType data into it?

[tustvold]

Sorry I meant into_builder for ByteArray

[alamb]

🎉 thank you @viirya -- cc @avantgardnerio