[SOLR-2888] FSTSuggester refactoring: utf8 storage, external sorts (OOM prevention), code cleanups

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.6, 4.0-ALPHA
• Component/s: spellchecker
• Assignee: Robert Muir
• Reporter: Dawid Weiss
• Created: 10/Nov/11 16:36
• Updated: 10/May/13 10:41
• Resolved: 17/Jan/12 04:02

Description

This issue incorporates several problems:

• utf16 was used previously to store and lookup terms, now it is utf8
• the construction would OOM with large number of terms because of the need to sort entries. Sorter APIs have been added and an implementation of external (on-disk) sorting is also added (Robert Muir).
• the FSTLookup class has been split and refactored into FSTCompletion and FSTCompletionBuilder, FSTCompletionLookup remains a facade connecting all the pieces and implements Lookup interface. For large inputs use FSTCompletionBuilder directly (and pre-bucket your input weights).
• Automatic bucketing in FSTCompletionLookup has been changed from linear min/max discretization into dividing into ranges after all values have been sorted. This empirically handles all potential distributions quite well. If somebody needs something very specific, use FSTCompletionBuilder directly (providing buckets), construct the automaton and then load it with FSTCompletionLookup.

Issue Links

• incorporates: SOLR-2887 FSTSuggester shouldn't OOM on large inputs
• incorporates: SOLR-2761 FSTLookup should use long-tail like discretization instead of proportional (linear)