[
https://issues.apache.org/jira/browse/SOLR-17447?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17881420#comment-17881420
]
Siju Varghese commented on SOLR-17447:
--------------------------------------
[~epugh] our latency budget for the "search as you type" use case is < 10ms.
Our index is about 150GB.
We plan to use both timeAllowed and maxHits. The timeAllowed is an upper bound,
the maxHits is a lower bound. Basically the idea is that for the search as you
type use case, just go through a limited number of docs and return what you
find.
> Add support for maxHits
> -----------------------
>
> Key: SOLR-17447
> URL: https://issues.apache.org/jira/browse/SOLR-17447
> Project: Solr
> Issue Type: New Feature
> Security Level: Public(Default Security Level. Issues are Public)
> Components: SearchComponents - other
> Reporter: Siju Varghese
> Priority: Minor
>
> Currently there are 3 mechanisms to control # of hits for a query
> * Use of the _timeAllowed_ query parameter - Though this does not directly
> control the number of hits, it has a similar effect with the collector
> terminating after the specified time budget has exceeded. The primary
> objective of this switch is to control runaway queries.
> * Use of {{{}segmentTerminateEarly{}}}{\{ __ }}parameter - This parameter is
> only applicable for sorted segments where the sort criteria requested matches
> the sort criteria used in the SortingMergePolicy
> * Use of cpuAllowed parameter to put upper bound on cpu time for a query.
>
> I would like to propose a new _maxHits_ parameter. This parameter early
> terminates the query once it has gone past the provided number of hits per
> shard.
> For us the motivation for such a parameter is the following:
> Our search is extremely latency sensitive and the query set is a mix of very
> high frequency tokens where we favor fast recall and typical search queries
> where we favor precision at low latency. The former can be thought of as a
> search as you type use case and we want to ensure that we return the results
> quickly and just go over enough documents we plan to control via the maxHits
> parameter. We can't use a sorted index for our use case because the sort
> criteria is a ranking function which is based off document features and the
> user input.
> With the maxHits parameter, it is quite likely that the results returned
> might not be the most relevant ones, however that is acceptable for us.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
