Parquet: Read/write support for PageIndex (ColumnIndex / OffsetIndex)

[alamb]

Implementation tracker:

• Parquet: Support read process for PageIndex (ColumnIndex / OffsetIndex)  #1749
• support write page index #1777

Is your feature request related to a problem or challenge? Please describe what you are trying to do.
Suggested by @ParadoxShmaradox ON #1191

In analytics systems, parquet files are passed around between different systems, so the more metadata is written by the rust implementation of parquet, the better other systems may be able to handle it.

One such type of metadata is PageIndex https://github.com/apache/parquet-format/blob/master/PageIndex.md

Specifically, I believe this is called a ColumnIndex and OffsetIndex in parquet-format:

https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L938-L971

https://github.com/apache/parquet-format/blob/master/src/main/thrift/parquet.thrift#L926-L932

Which is used to speed up queries in certain systems -- for example https://blog.cloudera.com/speeding-up-select-queries-with-parquet-page-indexes/

Describe the solution you'd like

1. Write PageIndex structures into files created by parquet-rs (specif
2. Provide access to reading PageIndex from the parquet metadata structures

Note that if creating the PageIndex structures is too expensive, their creation could be optional and controlled via options.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Original request here #1191 (comment)

[Ted-Jiang]

@alamb @tustvold
Our team is building a query engine which is sensitive to time-to-first-result latency(bottleneck is scan and agg). So we want to support PageIndex in arrow-rs avoid useless IO costs. I will follow up before discussions, Try to give a design of this (according to parquet-mr). If you can give any advice, would be grateful ❤️.

[tustvold]

Hi @Ted-Jiang, I think the first step will be to support reading and writing the page index. Once that is done, we can look to include support in #1605. I'm currently actively working on object store IO, so if you wanted to handle the index encode/decode we can probably parallelize this effort 😃

[Ted-Jiang]

Hi @Ted-Jiang, I think the first step will be to support reading and writing the page index. Once that is done, we can look to include support in #1605. I'm currently actively working on object store IO, so if you wanted to handle the index encode/decode we can probably parallelize this effort 😃

Sure！💪 I prefer to support reading(from java version) first, check the effect on performance.

[alamb]

I think the idea of parallelizing the implementation (as @tustvold suggests to implement the encode/decode of the page index) would be great

If creating the page index on write consumes too much time, perhaps we can make writing optional (aka in https://docs.rs/parquet/14.0.0/parquet/file/properties/struct.WriterProperties.html) -- same could be done for reading the page index if it is too expensive

[tustvold]

I think this is now complete