[
https://issues.apache.org/jira/browse/LUCENE-7110?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16334147#comment-16334147
]
Adrien Grand commented on LUCENE-7110:
--------------------------------------
If I understand correctly, leaf nodes would store the entire shape. One thing
that concerns me is that for all leaves whose MBR intersects the query, there
would be a lot of per-document work since the shape would need to be
deserialized and then compared with the query. I'm afraid this would scale
poorly when the number of indexed shapes increases?
Using a grid like the prefix tree strategy from lucene-spatial seems like it
could at least run INTERSECTS queries efficiently. Other relations sound more
complicated however since they require knowledge of the entire shape, so I'm
afraid they can't be implemented without a full scan?
In the end I'm wondering that the R-tree doesn't seem to bring much compared to
indexing shapes with a grid approach to have a good approximation of shapes
that intersect the query, and potentially storing shapes (or metadata about the
shapes) in a doc-value field in order to support other relations that
INTERSECTS using a two-phase iterator? I'm mentioning this in order to get
feedback about whether I'm missing something, but this per-document cost still
worries me and leads me to thinking that it might be saner to only support
intersection?
> Add Shape Support to BKD (extend to an R*/X-Tree data structure)
> ----------------------------------------------------------------
>
> Key: LUCENE-7110
> URL: https://issues.apache.org/jira/browse/LUCENE-7110
> Project: Lucene - Core
> Issue Type: New Feature
> Reporter: Nicholas Knize
> Priority: Major
>
> I've been tinkering with this off and on for a while and its showing some
> promise so I'm going to open an issue to (eventually) add this feature to a
> future release.
> R*/X-Tree is a data structure designed to support Shapes (2D, 3D, nD) where,
> like the internal node, the key for each leaf node is the Minimum Bounding
> Range (MBR - sometimes "incorrectly" referred to as Minimum Bounding
> Rectangle) of the shape. Inserting a shape then boils down to the best way of
> optimizing the tree structure. This optimization is driven by a set of
> criteria for choosing the appropriate internal key (e.g., minimizing overlap
> between siblings, maximizing "squareness", minimizing area, maximizing space
> usage). Query is then (a bit oversimplified) a two-phase process:
> * recurse each branch that overlaps with the MBR of the query shape
> * compute the relation with the leaf node(s) - in higher dimensions (3+)
> this becomes an increasingly difficult computational geometry problem
> The current BKD implementation is a special simplified case of an R*/X tree
> where, for Point data, it is always guaranteed there will never be overlap
> between sibling nodes (because you're using the point data as the keys). By
> exploiting this property the tree algorithms (split, merge, etc) are
> relatively cheap (hence their performance boost over postings based
> numerics). By modifying the key data, and extending the tree generation
> algorithms BKD logic can be extended to support Shape data using the MBR as
> the Key and modifying split and merge based on the criteria needed for
> optimizing a shape-based data structure.
> The initial implementation (based on limitations of the GeoAPI) will support
> 2D shapes only. Once the GeoAPI can performantly handle 3D shapes the change
> is relatively trivial to add the third dimension to the tree generation code.
> Like everything else, this feature will be created in sandbox and, once
> mature, will graduate (likely to lucene-spatial or lucene-spatial-extras
> depending on the library needs).
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
