More advanced techniques to read parquet files efficiently

[jiacai2050]

Describe This Problem

Usually the IO part of a query is the most time consuming, so reducing time spent on this would improve query latency quietly a lot.

In current implementation, we have already applies some tricks to optimize this, to name a few:

1. concurrent reads even for one file
2. min/max prune
3. custom bloom filter prune

There is an awesome blog written by @tustvold and @alamb introducing some more advanced techniques to further improve read speed, which is definitely a must-read for developer in Arrow ecosystem.

Proposal

Explore ideas introduced in Querying Parquet with Millisecond Latency, Some notable ideas are:

• Page prune
• Late materialization
• Decode optimization, especially dictionary encoding

Additional Context

No response