[
https://issues.apache.org/jira/browse/LUCENE-7656?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Adrien Grand updated LUCENE-7656:
---------------------------------
Attachment: LUCENE-7656.patch
Here is an updated patch that should address your comments:
- It still uses shifting, but there are no lost cells anymore. As a
side-effect, this also means we do not compute geo distances for those lost
cells at initialization time so it initialization might have got a bit faster.
- I wanted to look into initializing the distance predicate lazily but
remembered that {{IndexSearcher}} might call {{Weight.scorer}} from multiple
threads so this has some complexity that I'd like to delay to another issue. I
did more benchmarking to check how much of an issue that is, and initialization
constantly runs in less than half a millisecond on my machine, so it is not
that annoying to do it twice I think.
- I moved the distance predicate from Query to Weight. It has some memory
footprint, so it is better if queries do not hold a reference to it (eg. if
they are used as cache keys).
- I replaced the {{Relation[]}} array with a {{byte[]}} that stores ordinals
of the enum constants. This looked like a safe win as this 4x smaller array
could be more friendly with the cpu cache.
> Implement geo box and distance queries using doc values.
> --------------------------------------------------------
>
> Key: LUCENE-7656
> URL: https://issues.apache.org/jira/browse/LUCENE-7656
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Adrien Grand
> Priority: Minor
> Attachments: LUCENE-7656.patch, LUCENE-7656.patch, LUCENE-7656.patch
>
>
> Having geo box and distance queries available as both point and
> doc-values-based queries means we could use them with
> {{IndexOrDocValuesQuery}}.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
