[
https://issues.apache.org/jira/browse/LUCENE-9845?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17302649#comment-17302649
]
Robert Muir commented on LUCENE-9845:
-------------------------------------
{quote}
So, IndexedDISI would work OK if it provided a random access API. Graph search
tends to fetch values from all over the place, and it's not great to
continually reset the iterator or create a new one when you need to "go
backwards." I actually played around with extending IndexedDISI with random
access way back at the beginning of this HNSW patch:
https://issues.apache.org/jira/browse/LUCENE-9051, but dropped that idea in
the midst of just trying to get the basic feature working. Maybe we should
revisit that idea?
{quote}
I think we need to not waste a lick of time improving random access, and
instead prevent these algorithms from testing the user's hard drive seek speed,
and figure out how to make them work in hardware-friendly ways (means
SEQUENTIAL)
> Improve encoding of HNSW graph offsets
> --------------------------------------
>
> Key: LUCENE-9845
> URL: https://issues.apache.org/jira/browse/LUCENE-9845
> Project: Lucene - Core
> Issue Type: Wish
> Reporter: Michael Sokolov
> Priority: Major
>
> Today we use a simple {{long[]}} to encode the offset of each document's
> array of neighbors in the HNSW graph. Instead we should use a data structure
> that is optimized for encoding an increasing numeric array, like monotonic
> {{PackedInts}}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
