After some more digging, I narrowed it down to filtering. Without any filters,
the MLT is back to it's normal performance (8ms average response time for our
case). The issue goes away with 6.0 upgrade.
The hot method is Lucene's DisiPriorityQueue downHeap(), which takes 5X more
calls in 5.5.2 compared to 6.0. I am guessing that some of the Solr filters
refactoring fixed it for 6.0 release. I am not sure which.
As a work-around, for now I just refactored the custom MLT handler to convert
the filters into boolean clauses, which takes care of the issue.
Any insights into why this is happening in Solr 5.5.2?
Our configuration:
1. mlt.maxqt=1002. There is an additional filter passed as a parameter3. <field
name="some_mlt" type="text_en" indexed="true" stored="true" multiValued="true"
omitNorms="false" termVectors="true"/>4. text_en is a pretty standard text
fieldType.
Thanks,
Ivan
On Monday, October 31, 2016 5:10 PM, Ivan Provalov <iprov...@yahoo.com>
wrote:
I noticed a 3X performance degradation for MoreLikeThis between 4.6.1 and
5.5.2. Our configuration:
<field name="some_mlt" type="text_en" indexed="true" stored="true"
multiValued="true" omitNorms="false" termVectors="true"/>
where text_en is a pretty standard text fieldType.
Any pointers?
Thanks,
Ivan Provalov
