Vadim Ivanov commented on SOLR-4787:

How to upgrade to the latest version of SOLR and be able to perform fast joins 
(Bjoin or something as fast as it)
SOLR is developing and 6.x version is already available but I'm stuck on 4.10 
because of joins
I’m using bjoins to join collections with 2-3 billion documents (set of 
products) to collections of availability (less than 10 million documents)  and 
managed to get response time about 200 ms, which is acceptable. 
But, using regular joins, leads to response time greater than 10 sec
Are there any changes in 6x versions allowing join to be as fast as bjoin?

> Join Contrib
> ------------
>                 Key: SOLR-4787
>                 URL: https://issues.apache.org/jira/browse/SOLR-4787
>             Project: Solr
>          Issue Type: New Feature
>          Components: search
>    Affects Versions: 4.2.1
>            Reporter: Joel Bernstein
>            Priority: Minor
>             Fix For: 6.0
>         Attachments: SOLR-4787-deadlock-fix.patch, 
> SOLR-4787-pjoin-long-keys.patch, SOLR-4787-with-testcase-fix.patch, 
> SOLR-4787.patch, SOLR-4787.patch, SOLR-4787.patch, SOLR-4787.patch, 
> SOLR-4787.patch, SOLR-4787.patch, SOLR-4787.patch, SOLR-4787.patch, 
> SOLR-4787.patch, SOLR-4787.patch, SOLR-4787.patch, SOLR-4787.patch, 
> SOLR-4787.patch, SOLR-4787.patch, SOLR-4787.patch, 
> SOLR-4797-hjoin-multivaluekeys-nestedJoins.patch, 
> SOLR-4797-hjoin-multivaluekeys-trunk.patch
> This contrib provides a place where different join implementations can be 
> contributed to Solr. This contrib currently includes 3 join implementations. 
> The initial patch was generated from the Solr 4.3 tag. Because of changes in 
> the FieldCache API this patch will only build with Solr 4.2 or above.
> *HashSetJoinQParserPlugin aka hjoin*
> The hjoin provides a join implementation that filters results in one core 
> based on the results of a search in another core. This is similar in 
> functionality to the JoinQParserPlugin but the implementation differs in a 
> couple of important ways.
> The first way is that the hjoin is designed to work with int and long join 
> keys only. So, in order to use hjoin, int or long join keys must be included 
> in both the to and from core.
> The second difference is that the hjoin builds memory structures that are 
> used to quickly connect the join keys. So, the hjoin will need more memory 
> then the JoinQParserPlugin to perform the join.
> The main advantage of the hjoin is that it can scale to join millions of keys 
> between cores and provide sub-second response time. The hjoin should work 
> well with up to two million results from the fromIndex and tens of millions 
> of results from the main query.
> The hjoin supports the following features:
> 1) Both lucene query and PostFilter implementations. A *"cost"* > 99 will 
> turn on the PostFilter. The PostFilter will typically outperform the Lucene 
> query when the main query results have been narrowed down.
> 2) With the lucene query implementation there is an option to build the 
> filter with threads. This can greatly improve the performance of the query if 
> the main query index is very large. The "threads" parameter turns on 
> threading. For example *threads=6* will use 6 threads to build the filter. 
> This will setup a fixed threadpool with six threads to handle all hjoin 
> requests. Once the threadpool is created the hjoin will always use it to 
> build the filter. Threading does not come into play with the PostFilter.
> 3) The *size* local parameter can be used to set the initial size of the 
> hashset used to perform the join. If this is set above the number of results 
> from the fromIndex then the you can avoid hashset resizing which improves 
> performance.
> 4) Nested filter queries. The local parameter "fq" can be used to nest a 
> filter query within the join. The nested fq will filter the results of the 
> join query. This can point to another join to support nested joins.
> 5) Full caching support for the lucene query implementation. The filterCache 
> and queryResultCache should work properly even with deep nesting of joins. 
> Only the queryResultCache comes into play with the PostFilter implementation 
> because PostFilters are not cacheable in the filterCache.
> The syntax of the hjoin is similar to the JoinQParserPlugin except that the 
> plugin is referenced by the string "hjoin" rather then "join".
> fq=\{!hjoin fromIndex=collection2 from=id_i to=id_i threads=6 
> fq=$qq\}user:customer1&qq=group:5
> The example filter query above will search the fromIndex (collection2) for 
> "user:customer1" applying the local fq parameter to filter the results. The 
> lucene filter query will be built using 6 threads. This query will generate a 
> list of values from the "from" field that will be used to filter the main 
> query. Only records from the main query, where the "to" field is present in 
> the "from" list will be included in the results.
> The solrconfig.xml in the main query core must contain the reference to the 
> hjoin.
> <queryParser name="hjoin" 
> class="org.apache.solr.joins.HashSetJoinQParserPlugin"/>
> And the join contrib lib jars must be registed in the solrconfig.xml.
>  <lib dir="../../../contrib/joins/lib" regex=".*\.jar" />
> After issuing the "ant dist" command from inside the solr directory the joins 
> contrib jar will appear in the solr/dist directory. Place the the 
> solr-joins-4.*-.jar  in the WEB-INF/lib directory of the solr webapplication. 
> This will ensure that the top level Solr classloader loads these classes 
> rather then the core's classloaded. 
> *BitSetJoinQParserPlugin aka bjoin*
> The bjoin behaves exactly like the hjoin but uses a BitSet instead of a 
> HashSet to perform the underlying join. Because of this the bjoin is much 
> faster and can provide sub-second response times on result sets of tens of 
> millions of records from the fromIndex and hundreds of millions of records 
> from the main query.
> But there are limitations to how the bjoin can be used. The bjoin treats the 
> join keys as addresses in a BitSet and uses the Lucene OpenBitSet 
> implementation which performs very well but is not sparse. So the BitSet 
> memory is dictated by the size of the join keys. For example a bitset with a 
> max join key of 200,000,000 will need 25 MB of memory. For this reason the 
> BitSet join does not support long join keys. In order to keep memory usage 
> down the join keys should also be packed at the low end, for example from 1 
> to 50,000,000. 
> Below is a sampe bjoin:
> fq=\{!bjoin fromIndex=collection2 from=id_i to=id_i threads=6 
> fq=$qq\}user:customer1&qq=group:5
> To register the bjoin the solrconfig.xml in the main query core must contain 
> the reference to the bjoin.
> <queryParser name="bjoin" 
> class="org.apache.solr.joins.BitSetJoinQParserPlugin"/>
> *ValueSourceJoinParserPlugin aka vjoin*
> The second implementation is the ValueSourceJoinParserPlugin aka "vjoin". 
> This implements a ValueSource function query that can return a value from a 
> second core based on join keys and limiting query. The limiting query can be 
> used to select a specific subset of data from the join core. This allows 
> customer specific relevance data to be stored in a separate core and then 
> joined in the main query.
> The vjoin is called using the "vjoin" function query. For example:
> bf=vjoin(fromCore, fromKey, fromVal, toKey, query)
> This example shows "vjoin" being called by the edismax boost function 
> parameter. This example will return the "fromVal" from the "fromCore". The 
> "fromKey" and "toKey" are used to link the records from the main query to the 
> records in the "fromCore". The "query" is used to select a specific set of 
> records to join with in fromCore.
> Currently the fromKey and toKey must be longs but this will change in future 
> versions. Like the pjoin, the "join" SolrCache is used to hold the join 
> memory structures.
> To configure the vjoin you must register the ValueSource plugin in the 
> solrconfig.xml as follows:
> <valueSourceParser name="vjoin" 
> class="org.apache.solr.joins.ValueSourceJoinParserPlugin" />

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org

Reply via email to