lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Tony Ma <...@opentext.com>
Subject Re: [EXTERNAL] - Re: Is docvalue sorted by value?
Date Tue, 06 Mar 2018 02:28:25 GMT
Hi Erick,

I raise this question is about the sorting scenario as you mentioned in #2.

If the hit docs are about 100, and my query just want top 2. If the values are not sorted,
it has to iterate all 100 docs and find top2 in a priority queue. If the values are already
sorted, it just need to iterate first 2. If the query is unselective, the hit doc might be
huge, pre-sort or not will have big differences.

I understand your thinking that if the doc values are not persisted with doc id sequence,
it is unable to retrieve field value by doc id.

Actually, I am just wondering how lucene handle the sorting scenario, is iterating all values
of all docs unavoidable? 


On 3/6/18, 6:50 AM, "Erick Erickson" <erickerickson@gmail.com> wrote:

    I think there are two issues here that are being conflated
    1> _within_ a document, i.e. for a multi-valued field the values are
    stored as Dominik says as a SORTED_SET. Not only will they be returned
    (if you return from docValues rather than stored) in lexical order,
    but identical values will be collapsed
    
    2> across multiple documents, the question about  "...persisted with
    order of values, not document id..." really makes no sense. The point
    of DocValues is to answer the question "for document X what is the
    value of field Y". X here is the _internal_ document ID. Now consider
    a search. There are two documents that are hits, doc 35 and doc 198
    (internal lucene doc ID). To sort them by field Y you have to know
    what the value in that field is for those two docs is. How would
    "pre-ordering" the values help here? If I have the _values_ in order,
    I have no clue what docs are associated with them. That question is
    what the "inverted index" is there to answer.
    
    So I have doc 35 and 198. Think of DocValues as a large array indexed
    by internal doc id. To know how these two docs sort all I have to do
    is index into the array. It's slightly more complicated than that, but
    conceptually that's what happens.
    
    Best,
    Erick
    
    On Mon, Mar 5, 2018 at 11:29 AM, Dominik Safaric
    <dominiksafaric@gmail.com> wrote:
    >> So, can doc values be persisted with order of values, not document id? This should
be fast in sort scenario that the values are pre-ordered instead of scan/sort at runtime.
    >
    >
    > No, unfortunately doc values cannot be persisted in order. Lucene stores this values
internally as a DocValuesType.SORTED_SET, where the values are being stored using for example
Long.compareTo().
    >
    > If you'd like to retrieve the values in insertion order, use stored instead of doc
values instead of. Then you might access the values in order using the LeafReader's document
function. However, beware that may induce performance issues because it requires loading the
document from disk.
    >
    > If you require to store and retrieve multiple numeric values per document in order,
you might consider using PointValues. PointValues are internally indexed with KD-trees. But,
beware that PointValues have a limited dimensionality, in terms that you can for example store
values in 8 dimensions, each of max 16 bytes.
    >
    >> On 5 Mar 2018, at 15:33, Tony Ma <tma@opentext.com> wrote:
    >>
    >> Per my understanding, doc values (binary doc values / numeric doc values) are
stored with sequence of document id. Sorted numeric doc values just means if a document has
multiple values, the values will be sorted for same document, but for different documents,
the value is still ordered by document id. Is that true?
    >> So, can doc values be persisted with order of values, not document id? This should
be fast in sort scenario that the values are pre-ordered instead of scan/sort at runtime.
    >
    >
    > ---------------------------------------------------------------------
    > To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
    > For additional commands, e-mail: java-user-help@lucene.apache.org
    >
    
    ---------------------------------------------------------------------
    To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
    For additional commands, e-mail: java-user-help@lucene.apache.org
    
    

Mime
View raw message