Dawid Weiss created LUCENE-5977:
-----------------------------------

             Summary: IW should safeguard against token streams returning 
invalid offsets for multi-valued fields
                 Key: LUCENE-5977
                 URL: https://issues.apache.org/jira/browse/LUCENE-5977
             Project: Lucene - Core
          Issue Type: Improvement
    Affects Versions: 4.10.1, 4.10, 4.9.1, 4.9
            Reporter: Dawid Weiss
            Priority: Minor


We have a custom token stream that emits information about offsets of each 
token. My (wrong) assumption was that for multi-valued fields a token stream's 
offset information is magically shifted, much like this is the case with 
positions. It's not the case -- offsets should be increasing and monotonic 
across all instances of a field, even if it has a custom token streams. So, 
something like this:

{code}
        doc.add(new Field("field-foo", new CannedTokenStream(token("bar", 1, 
150, 160)), ftype));
        doc.add(new Field("field-foo", new CannedTokenStream(token("bar", 1,  
50,  60)), ftype));
{code}

where the token function is defined as:

{code}
token(String image, int positionIncrement, int startOffset, int endOffset)
{code}

will result in either a cryptic assertion thrown from IW:

{code}
Exception in thread "main" java.lang.AssertionError
        at 
org.apache.lucene.index.FreqProxTermsWriterPerField.writeOffsets(FreqProxTermsWriterPerField.java:99)
{code}

or nothing (or a codec error) if run without assertions.

Obviously returning non-shifted offsets from subsequent token streams makes 
little sense but I wonder if it could be made more explicit (or asserted) that 
offsets need to be increasing between multiple-values. The minimum is to add 
some documentation to OffsetAttribute. I don't know if offsets should be 
shifted automatically, as it's the case with positions -- this would change the 
semantics of existing tokenizers and filters which implement such shifting 
internally already.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to