[LUCENE-5542] Explore making DVConsumer sparse-aware

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Component/s: core/codecs
• Lucene Fields: New
• Assignee: Unassigned
• Reporter: Shai Erera
• Created: 20/Mar/14 21:28
• Updated: 28/Aug/22 14:03
• Resolved: 22/Sep/16 21:23

Description

Today DVConsumer API requires the caller to pass a value for every document, where null means "this doc has no value". The Codec can then choose how to encode the values, i.e. whether it encodes a 0 for a numeric field, or encodes the sparse docs. In practice, from what I see, we choose to encode the 0s.

I wonder if we e.g. added an Iterable<Number> to DVConsumer.addXYZField(), if that would make a better API. The caller only passes <doc,value> pairs and it's up to the Codec to decide how it wants to encode the missing values. Like, if a user's app truly has a sparse NDV, IndexWriter doesn't need to "fill the gaps" artificially. It's the job of the Codec.

To be clear, I don't propose to change any Codec implementation in this issue (w.r.t. sparse encoding - yes/no), only change the API to reflect that sparseness. I think that if we'll ever want to encode sparse values, it will be a more convenient API.

Thoughts? I volunteer to do this work, but want to get others' opinion before I start.

Issue Links

• is related to: LUCENE-7407 Explore switching doc values to an iterator API
• relates to: LUCENE-7253 Make sparse doc values and segments merging more efficient