Synthetic source should load fields sequentially

[iverase]

Synthetic source generally reads the values of a document from doc values. It does synthesise documents one by one by reading one doc value for each field at a time.

On the other hand doc values are optimised to read values sequentially so the access pattern of synthetic source is problematic as it will pollute the local cache and it can cause slow reads.

We should investigate the possibility of reading synthetic source fields sequentially to avoid cache issues.

[elasticsearchmachine]

Pinging @elastic/es-analytics-geo (Team:Analytics)

[elasticsearchmachine]

Pinging @elastic/es-analytics-geo (Team:Analytics)

[elasticsearchmachine]

Pinging @elastic/es-storage-engine (Team:StorageEngine)

[nik9000]

Synthetic source has some ability to do that. The interface is:

DocValuesLoader docValuesLoader(LeafReader leafReader, int[] docIdsInLeaf) throws IOException;

Some of the doc values implementations read the values out of doc values into an array up front and then replay them when they are building doc values. IIRC single valued numbers work that way but multivalued ones don't.