Hi Brian, I just tried to explore the scenario you describe with the techproducts example and am able to see what you see:
# step 1: start solr with techproducts example and ltr enabled # step 2: upload one feature (originalScore) and one model using that feature # step 3: examine cache stats via the Admin UI (all zero to start with) # step 4: run a query which includes feature extraction e.g. [features] in fl # step 5: examine cache stats to see lookups but no inserts # step 6: run a query with feature extraction _and_ re-ranking using the model # step 7: examine cache stats to see both lookups and inserts Looking around the code the cache insert happens in FeatureLogger.java [1] which is called by the Rescorer [2] and this would allow the 'fl' feature logging to reuse the feature values calculated as part of the 'rq' re-ranking. However, if there was no feature value in the cache (because no 'rq' re-ranking happened) then the feature value is calculated by LTRFeatureLoggerTransformerFactory.java [3] and based on code inspection the results of that calculation are not added to the cache. It might be interesting to explore if/how that logic [3] could be changed. --Christine [1] https://github.com/apache/lucene-solr/blob/releases/lucene-solr/7.1.0/solr/contrib/ltr/src/java/org/apache/solr/ltr/FeatureLogger.java#L51-L60 [2] https://github.com/apache/lucene-solr/blob/releases/lucene-solr/7.1.0/solr/contrib/ltr/src/java/org/apache/solr/ltr/LTRRescorer.java#L185-L205 [3] https://github.com/apache/lucene-solr/blob/releases/lucene-solr/7.1.0/solr/contrib/ltr/src/java/org/apache/solr/ltr/response/transform/LTRFeatureLoggerTransformerFactory.java#L267-L280 ----- Original Message ----- From: solr-user@lucene.apache.org To: solr-user@lucene.apache.org At: 10/30/17 16:55:14 I'm still having this issue. Does anyone have LTR feature extraction successfully running and have cache inserts/hits? --Brian -----Original Message----- From: Brian Yee [mailto:b...@wayfair.com] Sent: Tuesday, October 24, 2017 12:14 PM To: solr-user@lucene.apache.org Subject: RE: LTR feature extraction performance issues Hi Alessandro, Unfortunately some of my most important features are query dependent. I think I found an issue though. I don't think my features are being inserted into the cache. Notice "cumulative_inserts:0". There are a lot of lookups, but since there appear to be no values in the cache, the hitratio is 0. stats: cumulative_evictions:0 cumulative_hitratio:0 cumulative_hits:0 cumulative_inserts:0 cumulative_lookups:215319 evictions:0 hitratio:0 hits:0 inserts:0 lookups:3303 size:0 warmupTime:0 My configs look are as follows: <cache name="QUERY_DOC_FV" class="solr.search.LRUCache" size="4096" initialSize="2048" autowarmCount="4096" regenerator="solr.search.NoOpRegenerator" /> <queryParser name="ltr" class="org.apache.solr.ltr.search.LTRQParserPlugin"/> <transformer name="features" class="org.apache.solr.ltr.response.transform.LTRFeatureLoggerTransformerFactory"> <str name="fvCacheName">QUERY_DOC_FV</str> <str name="defaultFormat">sparse</str> </transformer> Would anyone have any idea why my features are not being inserted into the cache? Is there an additional config setting I need? --Brian -----Original Message----- From: alessandro.benedetti [mailto:a.benede...@sease.io] Sent: Monday, October 23, 2017 10:01 AM To: solr-user@lucene.apache.org Subject: Re: LTR feature extraction performance issues It strictly depends on the kind of features you are using. At the moment there is just one cache for all the features. This means that even if you have 1 query dependent feature and 100 document dependent feature, a different value for the query dependent one will invalidate the cache entry for the full vector[1]. You may look to optimise your features ( where possible). [1] https://issues.apache.org/jira/browse/SOLR-10448 ----- --------------- Alessandro Benedetti Search Consultant, R&D Software Engineer, Director Sease Ltd. - www.sease.io -- Sent from: http://lucene.472066.n3.nabble.com/Solr-User-f472068.html
