This commit introduces custom tiebreakers which allows users to
specify custom tiebreakers when ordering hits to return. A
default tiebreaker is introduced for tie breaking on shard index
first and then docID.
group.query after execution forms QueryCommandResult. In case of
group.main=true or group.format=simple, QueryCommandResult was not
consumed in EndResultTransformer. Also, MainEndResultTransformer assumed
that always group.field would be specified. When group.field not specified
it failed with AIOOBE. After adding suppport for QueryCommandResult in
EndResultTransformers and handling AIOOBE, group.query started giving results
Working on tests exposed few other issues. Results differed b/w standalone
& distributed mode.
* One of the reason is that TopGroupShardResponseProcessor doesn't consider correct
limit and offset when group format is simple. In case of simple, start and rows should be used
as limit and offset instead of group.limit and group.offset.
* Secondly, In distributed second phase grouping, computing docsToCollect didn't consider
group response format. This issue is again similar to above issue
* offset(group.offset or start) not being considered during TopDocs#merge caused
different results. The fix was to use to offset in merge process
* group.offset doesn't support negative values but there is no checks on the value.
In case of negative values AIOOBE. Now, checks are added for negative values and
returns proper error message(this change is for both standalone and distrbuted).
Validation is done only in case of group.format=grouped as that is only case when
group.offset is consumed.
Fixing above issues resolved the differences b/w standalone and distributed mode.
* When ramPerThreadHardLimitMB is not specified, then Lucene's
default value 1945 is used. The specified value should be
greater than 0 and less than 2048MB
* Improve error message when collapsing is not supported on given
fieldtype
* Return 400 error code when unsupported value are passed for max,min
or in case of syntax error
This reverts commit 6d6f14d39123512b8734d63c584bceb9d7bd832f.
Reason for revert: after doing more testing I realized there are tests I overlooked which can (with randomized SSL usage) trigger NullPointerException
(or other related failures) in After/AfterClass due assumptions about cleanup that isn't actaully neccessary due to the AssumptionFailure
that may occur during Before/BeforeClass
* Inplace update supports set and inc operation but when null or
empty list is specified with set op, then it should always be treated
as atomic update since this case is equivalent to removing field
from the document
The UH now detects that parts of the query are not understood by it.
When found, it highlights more safely/reliably.
Fixes compatibility with complex and surround query parsers.
