IPC sink types and IPC file stream

[dexterduck]

IPC component of this PR: #876

Adds the following new types:

• io::ipc::read::file_async::FileStream - implements futures::Stream for IPC files.
• io::ipc::write::file_async::FileSink - implements futures::Sink for IPC files.
• io::ipc::write::file_async::StreamSink - implements futures::Sink for IPC streams.

[codecov]

Codecov Report

Merging #878 (3afd996) into main (66616dd) will decrease coverage by 0.03%.
The diff coverage is 66.78%.

@@            Coverage Diff             @@
##             main     #878      +/-   ##
==========================================
- Coverage   71.68%   71.64%   -0.04%     
==========================================
  Files         335      337       +2     
  Lines       18205    18452     +247     
==========================================
+ Hits        13050    13220     +170     
- Misses       5155     5232      +77     

Impacted Files	Coverage Δ
src/io/ipc/write/mod.rs	92.30% <ø> (ø)
src/io/ipc/write/common.rs	86.40% <58.33%> (-3.71%)	⬇️
src/io/ipc/read/file_async.rs	58.59% <58.59%> (ø)
src/io/ipc/write/stream_async.rs	72.41% <73.21%> (+16.85%)	⬆️
src/io/ipc/write/file_async.rs	75.29% <75.29%> (ø)
src/io/ipc/read/reader.rs	72.85% <100.00%> (+1.42%)	⬆️
src/io/ipc/write/common_async.rs	96.66% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 66616dd...3afd996. Read the comment docs.

[dexterduck]

@jorgecarleitao along the same lines as the Parquet integration, the Sink implementations could replace the file/stream writers for IPC as well, although it would require a wrapper type in order to preserve the ability to do per-chunk projections.

Potentially something to the effect of:

// Wrapper type
pub struct Projection {
  pub chunk: Chunk<Arc<dyn Array>>,
  pub fields: Option<Vec<IpcField>>,
}

impl From<Chunk<Arc<dyn Array>>> for Projection {
  fn from(chunk: Chunk<Arc<dyn Array>>) -> Self {
    Self { chunk, fields: None }
  }
}

// Usage
let mut sink = FileSink::new(..);
let chunks: Vec<Chunk<Arc<dyn Array>>> = vec![];
for chunk in chunks {
  sink.feed(chunk.into()).await?;
}

[jorgecarleitao]

• The IPCField is required to re-use dictionaries across IPC record batches. So, we do need it in the API.
• I am not very confortable with code duplication - could you either change the existing code (e.g. implement Sink for the structs we have) or remove them? Having 2 implementations of the same thing is not ideal imo

[dexterduck]

@jorgecarleitao I've added support for passing &[IpcField] to the stream and file sinks. I've removed StreamWriter and updated the integration tests to use the new implementations.

Per the example above, I've added a compatibility struct which allows calling the sinks as either sink.feed(chunk.into()) or sink.feed((chunk, Some(fields)).into()). The struct is currently io::ipc::write::Record. I presume there's a better name I could be using for it so just let me know if I should change it.

[jorgecarleitao]

Looks great! Thank you so much, @dexterduck - learned a lot in this PR!

[jorgecarleitao]

Curious: what is the purpose of Cow?

[dexterduck]

The idea is to allow passing the arguments either by value or by reference. Obviously, better to avoid unnecessary clones if possible, but also wanted to avoid requiring a reference since then if you want to pass owned values you have to use the somewhat non-ergonomic sink.feed((&chunk).into()) or sink.feed((&chunk, &fields[..]).into()). In contrast, with the chosen implementation you can use sink.feed(chunk.into()) or sink.feed((chunk, fields).into()) in all cases.

I originally used a generic implementation that parameterized Record to accept a type implementing Borrow<_>, but it turns out that having a single type that implements sink for multiple other types (e.g. FileSink implementing both Sink<Record<Chunk<_>>> and Sink<Record&<Chunk<_>>>) doesn't work because the compiler can't infer the correct generic type when you call sink.flush() or sink.close().

[jorgecarleitao]

Thank you for the explanation. Cool tip!