lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Erick Erickson" <erickerick...@gmail.com>
Subject Re: retrieve all docs efficiently - just one field
Date Wed, 11 Jun 2008 14:05:47 GMT
<<<I have read a little about the hitcollector class and the fieldselector
api,
but I am still not sure how they may help me or even if they can.>>>

I infer from this that you're using a Hits object to get your IDs to insert
in
your temporary table. Here's the problem with Hits... It re-executes
the query every 100 (200?) hits. So you can think of it as

while (more hits) {
   if ((count % 100) == 0) execute the search and throw away the first
<count> items
   work with the document
}

It can be a major bottleneck to re-execute the query every 100 hits you look
at. HitCollector avoids this re-execution, and can result in very
significant
speedups when iterating through many documents.

FieldSelector will allow lazy fetching. That is, when you do something
like Reader.document(idx, selector) you'll be able to only load those
fields from the document that you specify with the selector. In your case,
you would only load the ID you care about and insert that in your temporary
table. This can also result in very significant savings, especially if you
only want to load a very small field from a document that has very large
fields. See a writeup I did for one of my projects on the Lucene Wiki

http://wiki.apache.org/lucene-java/FieldSelectorPerformance?highlight=(FieldSelector)


Hope this helps
Erick



On Tue, Jun 10, 2008 at 6:35 PM, 1world1love <jd_cowan@yahoo.com> wrote:

>
> Greetings all. I have read many posts concerning similar use cases, but I
> am
> still a little hazy on the best way to achieve what I need to do. Here is
> the background:
>
> 2 million documents with multiple sections, some sections contain
> structured
> data, some unstructured.
>
> We parse the docs and place the structured stuff in oracle where each
> section is a table and one master table to relate them all.
>
> We index the unstructured sections with lucene where each section is a
> document (meaning a total of about ~30 million documents) with extra fields
> including one for the primary key of the master table and then some meta
> fields to describe the section - type, date, etc.
>
> For a common use case, say we have a table called demographics with a
> number
> field that represents age (overly simplistic but gets the point across).
>
> So say we want all people over the age of 50 who may have visited Panama:
>
> --
> We have our lucene index and we want to search the section text for the
> word
> "panama"
>
> AND
>
> We want to select from the demographics table where age > 50.
> --
>
> Now I need to intersect the master table IDs from my lucene hits and my
> table results.
>
> I have a java stored procedure that runs the lucene query and creates a
> temporary table with a single column where I insert the master id from the
> hits of my lucene query. I then can do a join with my structured query
> results.
>
> The problem here is obviously the speed of iterating through the hits to
> extract the single field that I need.
>
> Notes:
> - I must be able to get a full set of results, though I only need the one
> id
> field
> - We originally went with Oracle text which was simple, but limited and
> quite slow for most queries
>
>
> I have read a little about the hitcollector class and the fieldselector
> api,
> but I am still not sure how they may help me or even if they can.
>
> I have also tooled around with the idea of using termdocs, but the queries
> may get a little complex with various ors/ands/nots, though probably not
> spans and so forth.
>
> Any suggestions will be greatly apreciated.
>
> Thanks,
>
> J
>
> --
> View this message in context:
> http://www.nabble.com/retrieve-all-docs-efficiently---just-one-field-tp17766268p17766268.html
> Sent from the Lucene - Java Users mailing list archive at Nabble.com.
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-user-help@lucene.apache.org
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message