[
https://issues.apache.org/jira/browse/LUCENE-7563?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Michael McCandless updated LUCENE-7563:
---------------------------------------
Attachment: LUCENE-7563.patch
Initial patch, compacting (at indexing time) the long[] leafBlockFPs
and byte[] splitPackedValues in the current BKD index, into a smaller
packed byte[] structure.
I still have a number of nocommits but the core idea seems to be
working ... Lucene's tests pass at least once.
At first I prototyped using an FST to compact the binary tree index,
and it reduced quite a bit, but then I realized even FST has some
annoying inefficiency for this usage, so I made this dedicated packed
byte[] structure instead which compressed even better.
On our nightly 20M NYC taxis benchmark
(https://home.apache.org/~mikemccand/lucenebench/sparseResults.html)
this reduces heap usage by ~60%.
It compresses N>1 dimension too but I haven't tested by how much...
The structure should work very well for apps that e.g. index a short
as if it were a long, prefix-coding away all those leading 0s. It
takes advantage of how the BKD tree is traversed at search time,
always starting at root and pushing down to children.
Separately, I think we could consider increasing the max leaf block size
from 1024 points to maybe 2048 or 4096 points ... I'll open a new
issue for that.
> BKD index should compress unused leading bytes
> ----------------------------------------------
>
> Key: LUCENE-7563
> URL: https://issues.apache.org/jira/browse/LUCENE-7563
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Michael McCandless
> Fix For: master (7.0), 6.4
>
> Attachments: LUCENE-7563.patch
>
>
> Today the BKD (points) in-heap index always uses {{dimensionNumBytes}} per
> dimension, but if e.g. you are indexing {{LongPoint}} yet only use the bottom
> two bytes in a given segment, we shouldn't store all those leading 0s in the
> index.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
