2aacdd0e7b
When constructing a SearcherManager with an IndexWriter, as SubIndex does, the returned searcher has immediate access to the unflushed writes, so there is no need to commit after every write to make the writes visible in the running process. This allows us to "decouple durability to hardware/OS crashes from visibility of changes to a new IndexReader" [1]. Since we don't know exactly when changes will get committed to disk, start a thread to periodically refresh the SearcherManagers. The 100ms poll time is a bit of a guess, but [1] says even 50ms should not be a significant load. [1] http://blog.mikemccandless.com/2011/11/near-real-time-readers-with-lucenes.html Change-Id: I1d413e8334057e1b04f9d6414635a2d1c53507f4