[ 
https://issues.apache.org/jira/browse/LUCENE-9583?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17222245#comment-17222245
 ] 

Jim Ferenczi commented on LUCENE-9583:
--------------------------------------

I wonder if the graph approach really needs random access. For each node we 
need to access the list of neighbors so this pattern doesn't require random 
access because the list is already sorted by doc ids. So instead of adding 
another interface I wonder what do you think of adding a reset method in 
VectorValues ? For each node, the pattern to access would be to reset the 
iterator first and then move it to the first neighbor. We can make optimization 
internally to provide fast reset so that we don't need two implementations for 
the first two approaches that we foresee ?

 

> Without it, we would need to load all vectors into RAM while 
>flushing/merging, as we currently do in {{BinaryDocValuesWriter.BinaryDVs}}. I 
>wonder if it's worth paying this cost for the simpler API

 

We only do that when flushing so the amount of data is limited. 

The only case where this could be an issue is if you're sorting your index 
after the fact. That's a new feature we added recently and in this specific 
case we have to load the entire binary field in memory. However this is not 
specific to vectors, any binary field is affected so it's a more generic 
problem that we don't need to solve now or maybe never ;).

> How should we expose VectorValues.RandomAccess?
> -----------------------------------------------
>
>                 Key: LUCENE-9583
>                 URL: https://issues.apache.org/jira/browse/LUCENE-9583
>             Project: Lucene - Core
>          Issue Type: Improvement
>            Reporter: Michael Sokolov
>            Priority: Major
>          Time Spent: 10m
>  Remaining Estimate: 0h
>
> In the newly-added {{VectorValues}} API, we have a {{RandomAccess}} 
> sub-interface. [~jtibshirani] pointed out this is not needed by some 
> vector-indexing strategies which can operate solely using a forward-iterator 
> (it is needed by HNSW), and so in the interest of simplifying the public API 
> we should not expose this internal detail (which by the way surfaces internal 
> ordinals that are somewhat uninteresting outside the random access API).
> I looked into how to move this inside the HNSW-specific code and remembered 
> that we do also currently make use of the RA API when merging vector fields 
> over sorted indexes. Without it, we would need to load all vectors into RAM  
> while flushing/merging, as we currently do in 
> {{BinaryDocValuesWriter.BinaryDVs}}. I wonder if it's worth paying this cost 
> for the simpler API.
> Another thing I noticed while reviewing this is that I moved the KNN 
> {{search(float[] target, int topK, int fanout)}} method from {{VectorValues}} 
>  to {{VectorValues.RandomAccess}}. This I think we could move back, and 
> handle the HNSW requirements for search elsewhere. I wonder if that would 
> alleviate the major concern here? 



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to