lucene-solr-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Timothy Potter <thelabd...@gmail.com>
Subject Re: external values source
Date Sat, 20 Apr 2013 00:02:20 GMT
Hi Maciek,

I think a custom ValueSource is definitely what you want because you
need to compute some derived value based on an indexed field and some
external value.

The trick is figuring how to make the lookup to the external data
very, very fast. Here's a rough sketch of what we do:

We have a table in a database that contains a numeric value for a user
and an organization, such as query:

select num from table where userId='bob' and orgId=123 (similar to
what you stated in question #4)

On the Solr side, documents are indexed with user_id_s field, which is
half of what I need to do my lookup. The orgId is determined by the
Solr client at query construction time, so is passed to my custom
ValueSource (aka function) in the query. In our app, users can be
associated with many different orgIds and changes frequently so we
can't index the association.

To do the lookup to the database, we have a custom ValueSource,
something like: dbLookup(user_id_s, 123)

(note: user_id_s is the name of the field holding my userID values in
the index and 123 is the orgId)

Behind the scenes, the ValueSource will have access to the user_id_s
field values using FieldCache, something like:

final BinaryDocValues dv =
FieldCache.DEFAULT.getTerms(reader.reader(), "user_id_s");

This gives us fast access to the user_id_s value for any given doc
(question #1 above) So now we can return an IntDocValues instance by
doing:

    @Override
    public FunctionValues getValues(Map context, AtomicReaderContext
reader) throws IOException {
        final BytesRef br = new BytesRef();
        final BinaryDocValues dv =
FieldCache.DEFAULT.getTerms(reader.reader(), fieldName);
        return new IntDocValues(this) {
            @Override
            public int intVal(int doc) {
                dv.get(doc,br);
                if (br.length == 0)
                    return 0;

                final String user_id_s = br.utf8ToString(); // the
indexed userID for doc
                int val = 0;
                // todo: do custom lookup with orgID and user_id_s to
compute int value for doc
                return val;
            }
        }
        ...
    }

In this code, fieldName is set in the constructor (not shown) by
parsing it out of the parameters, something like:

            this.fieldName =
((org.apache.solr.schema.StrFieldSource)source).getField();

The user_id_s field comes into your ValueSource as a StrFieldSource
(or whatever type you use) ... here is how the ValueSource gets
constructed at query time:

public class MyValueSourceParser extends ValueSourceParser {
    public void init(NamedList namedList) {}

    public ValueSource parse(FunctionQParser fqp) throws SyntaxError {
        return new MyValueSource(fqp.parseValueSource(), fqp.parseArg());
    }
}

There is one instance of your ValueSourceParser created per core. The
parse method gets called for every query that uses the ValueSource.

At query time, I might use the ValueSource to return this computed
value in my fl list, such as:

fl=id,looked_up:dbLookup(user_id_l,123),...

Or to sort by:

sort=dbLookup(user_id_s,123) desc

The data in our table doesn't change that frequently, so we export it
to a flat file in S3 and our custom ValueSource downloads from S3,
transforms it into an in-memory HashMap for fast lookups. We thought
about just issuing a query to load the data from the db directly but
we have many nodes and the query is expensive and result set is large
so we didn't want to hammer our database with N Solr nodes querying
for the same data at roughly the same time. So we do it once and post
the compressed results to a shared location. The data in the table is
"sparse" as compared to the number of documents and userIds we have.

We simply poll S3 for changes every few minutes, which is good enough
for us. This happens from many nodes in a large Solr Cloud cluster
running in EC2 so S3 works well for us as a distribution mechanism.
Admittedly polling kind of sucks so we tried using Zookeeper to notify
our custom watchers when a znode changes but a ValueSource doesn't get
notified when a core is reloaded so we ended up having many weird
issues with Zookeeper watchers in our custom ValueSource. For example,
new ValueSourceParsers get created when a core is reloaded but the
previous instance doesn't get notified that it's going out of service.
So this gives you an idea of how we load external data into a fast
lookup data structure in Solr (~question #2)

When filtering, we use PostFilter to tell Solr that our filter is
expensive so should be applied last (after all other criteria have
run), something like:

fq={!frange l=2 u=8 cost=200 cache=false}dbLookup(user_id_s,123)

This computes a function range query using our custom ValueSource but
tells Solr that it is expensive (cost >= 100) so apply it after all
other filters have been applied.
http://yonik.wordpress.com/tag/post-filter/

Lastly, as for speed, the user_id_s field gets loaded into FieldCache
and the lookup happens in a HashMap, albeit a large one. We're pretty
happy with the performance of this function and the only rub is that
our lookup table has to fit in memory.

Hope this helps or at least gives you some info for further questions.

Cheers,
Tim

On Fri, Apr 19, 2013 at 4:03 PM, Maciej Li┼╝ewski <maciej.lizewski@3e.pl> wrote:
> I need some explanation on how ValuesSource and related classes work.
>
> There are already implemented ExternalFileField, example on how to load data
> from database (
> <http://sujitpal.blogspot.com/2011/05/custom-sorting-in-solr-using-external.
> html>
> http://sujitpal.blogspot.com/2011/05/custom-sorting-in-solr-using-external.h
> tml)
>
> But they all fetch ALL data into memory which may consume large amounts of
> this resource. Also documents are referenced by 'doc' integer value.
>
>
>
> My questions:
>
> 1)      Is the 'doc' value pointing to document in whole index? If so - how
> to get value of such documents field (for example: field named 'id')?
>
> 2)      Is there possibility to create ValuesSource, FieldType (or similar
> interface which will provide external data to sort and in query results)
> which will work only on some subset of documents and use external source
> capabilities to fetch document related data?
>
> 3)      How does it all work (memory consumption, hashtable access speed,
> etc), when there is a lot of documents in index (tens of millions for
> example)?
>
> 4)      Are there any other examples on loading external data from database
> (I want to have numerical 'rate' from simple table having two columns:
> 'document unique key' string, 'rate' integer/float) which are not just proof
> of concept but real-life examples?
>
>
>
> Any help and hints appreciated
>
> TIA
>
>
>
> --
>
> Maciek
>

Mime
View raw message