[LUCENE-8916] GraphTokenStreamFiniteStrings.FiniteStringsTokenStream does not play well with subsequent TokenFilters

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 8.3
• Lucene Fields: New
• Assignee: Alan Woodward
• Reporter: Alan Woodward
• Created: 15/Jul/19 09:33
• Updated: 28/Aug/22 15:48
• Resolved: 19/Jul/19 10:47

Description

GraphTokenStreamFiniteStrings provides a view over multiple paths through a Token graph, which is useful when building queries over multiple length synonyms. This view is exposed as an iterator over simple TokenStreams. However, these TokenStreams do not work correctly when further wrapped in token filters, because they do not use a CharTermAttribute.

For an example of issues this can cause, see https://github.com/elastic/elasticsearch/issues/43976, where elasticsearch uses a special shingle field to speed up phrase searches. Queries are converted to shingles if they have multiple terms. However, if the query resolves into a graph due to synonyms, then this conversion breaks because the FixedShingleFilter is given a token stream built by GTSFS; terms are set using BytesTermAttribute, but then read using CharTermAttribute, and as these have different backing implementations, FSF ends up emitting null tokens.

Issue Links

• links to: GitHub Pull Request #786