improve: reuse `Arc<dyn Array>` in parquet record batch reader.

[RinChanNOWWW]

Both in arrow_reader and async_reader, if there are predicates, the reader will first read arrays (wrapped in RecordBatch) to evaluate the predicates and get a row selection. And then the reader will use the row selection to output the final needed arrays.

If some arrays in the final output are contained in the prefetched arrays, we also have to deserialize them again. This is quite wasteful.

We should have a reasonable way to reuse the arrays.

[RinChanNOWWW]

In my opinion, we can store the primitive arrays once we read in the parquet record batch reader, and combine them into a RecordBatch when needed.

[tustvold]

The challenge is bounding the memory usage of the reader and balancing this against the cost of decoding, which for primitives is relatively low. I don't think it is as simple as "keep the arrays around", it will probably require some sort of heuristic

[RinChanNOWWW]

The challenge is bounding the memory usage of the reader and balancing this against the cost of decoding
I don't think it is as simple as "keep the arrays around"

We can only keep the arrays we need to output and release the others.

which for primitives is relatively low

Besides decoding there is also decompression. And it may become a large cost for variable-length type like string.

it will probably require some sort of heuristic

Yes. Such as all the output arrays are contained in predicate arrays.

[tustvold]

We can only keep the arrays we need to output and release the others.

That is still potentially a non-trivial amount of data, it could theoretically be an entire column chunk worth, which could easily blow your memory budget 😅

Perhaps we could add a configurable threshold of the maximum number of bytes to keep around in this way, and fallback to decoding columns again if this threshold is exceeded?

[alamb]

🎣 -- I believe @XiangpengHao was thinking about working on this (I know you have a paper to write too, etc lol)

[alamb]

@XiangpengHao came up with a way to cache just 2 pages, see #7363

[alamb]

I am actively working on this approach. Details on

• [EPIC] Faster performance for parquet predicate evaluation for non selective filters #7456