[C++] Reading and Writing RLE data from/to Parquet

[asfimport]

Parquet already supports RLE data. So this would involve adding a new (optional) code path that this data can be directly read into the new Arrow RLE arrays, without decoding it.

Reporter: Tobias Zagorni / @zagto

_{Note: This issue was originally created as ARROW-17028. Please see the migration documentation for further details.}

[felipecrv]

RLE is now called REE.

[mapleFU]

Should we first implement REE write to Plain?
And when reading from Parquet, what can be used to build the REE Array? And how can we benifits from it?

(Maybe the best encoding is RLE in parquet, but it might need some complex logic)

[felipecrv]

Should we first implement REE write to Plain?

That would be great and a stop-gap solution would be to use RunEndDecode (expanding the array into memory first, then writing to Parquet).

If you want a more memory-efficient solution, you can look at the RunEndDecode code and write an adaptation that streams data into a plain Parquet array if required to be plain by the Parquet schema for some reason.

(Maybe the best encoding is RLE in parquet, but it might need some complex logic)

Yes, ideally we should write REE Arrow arrays as RLE Parquet arrays. It's on the level of complexity of RunEndDecode and I can help you extract common code from there if we find opportunities of re-use.

And when reading from Parquet, what can be used to build the REE Array? And how can we benifits from it?

I'm not familiar with the Parquet format, but AFAIU, it is able to store run-LENGTH encoded data, so reading that directly into a run-END encoded array would require less copying and less memory.

[mapleFU]

I think it should take some time to implement it.

1. Naive implemention: suitable for all kinds of encoder, writer does like decode from RunEndDecode and Put into encoder buffer
2. Special implemention for dict: for non-boolean type( I don't know whether string or binary is supported), REE can optimize for dictionary encoder.
3. Special implemention for RLE (RLE / bitpack): I guess it can be implemented, but currently I think it's a bit tricky

Maybe we could first implement (1), then regard (2) as an optimization @felipecrv ?

[felipecrv]

Maybe we could first implement (1), then regard (2) as an optimization @felipecrv ?

Yes. Sounds like a good plan.

This is the list of types supported by RunEndEncode and RunEndDecode so far. It doesn't include all nested types at the moment as I figured those would be more likely to be dict-encoded when compressed.

arrow/cpp/src/arrow/compute/kernels/vector_run_end_encode.cc

Lines 591 to 611 in 99dd998

	add_kernel(Type::NA);
	add_kernel(Type::BOOL);
	for (const auto& ty : NumericTypes()) {
	add_kernel(ty->id());
	}
	add_kernel(Type::DATE32);
	add_kernel(Type::DATE64);
	add_kernel(Type::TIME32);
	add_kernel(Type::TIME64);
	add_kernel(Type::TIMESTAMP);
	add_kernel(Type::DURATION);
	for (const auto& ty : IntervalTypes()) {
	add_kernel(ty->id());
	}
	add_kernel(Type::DECIMAL128);
	add_kernel(Type::DECIMAL256);
	add_kernel(Type::FIXED_SIZE_BINARY);
	add_kernel(Type::STRING);
	add_kernel(Type::BINARY);
	add_kernel(Type::LARGE_STRING);
	add_kernel(Type::LARGE_BINARY);

[mapleFU]

@jsjtxietian would you like to working on this?

[jsjtxietian]

@jsjtxietian would you like to working on this?

ok, will take a look this week

[mapleFU]

@jsjtxietian You can take a look at TypedColumnWriterImpl<DType>::WriteArrowDictionary, and do the same things. If you meet any problem you can ask me for that