SimpleFacets: Performance Boost for Tokenized Fields for smaller DocSet using Term Vectors

• Key: SOLR-711
• Type: Improvement
• Status: Closed
• Resolution: Fixed
• Priority: Major
• Assignee: Unassigned
• Reporter: Fuad Efendi
• Component/s: search
• Affects Version/s: 1.3
• Fix Version/s: 1.4

Description

From http://www.nabble.com/SimpleFacets%3A-Performance-Boost-for-Tokenized-Fields-td19033760.html:

Scenario:

• 10,000,000 documents in the index;
• 5-10 terms per document;
• 200,000 unique terms for a tokenized field.

Obviously calculating sizes of 200,000 intersections with FilterCache is 100 times slower than traversing 10 - 20,000 documents for smaller DocSets and counting frequencies of Terms.

Not applicable if size of DocSet is close to total number of unique tokens (200,000 in our scenario).

See SimpleFacets.java:

public NamedList getFacetTermEnumCounts(
  SolrIndexSearcher searcher, 
  DocSet docs, ...

Comments

[Toby Cole]

We've seen this problem with our dataset, we have around 10m small records and were trying to facet on several multi-valued strings. Two of which had over 40k unique values (around 10 values per record).
If we can come up with a plan I don't mind volunteering to implement it.

[Shalin Shekhar Mangar]

What should be a good criteria to switch between the current and proposed strategy?

Fuad, did you run any tests to find a magic ratio between unique tokens and DocSet size?

With the new performance changes in faceting with SOLR-475, is this issue still relevant?

[Fuad Efendi]

Thanks Shalin for pointing to SOLR-475 which is very advanced solution to term counting approach.