It works via two Unix file system tricks. 1) Files are not directly bound with with filenames, instead there is a layer of indirection called an 'inode'. So, multiple file and directory names point to the same physical file. The "." and ".." directory entries are implemented this way.
2) Physical files are bound to all open file descriptors even after there are no file names for the files. So, file data exists until all file names are gone AND all open files are gone. Lance -----Original Message----- From: Tracy Flynn [mailto:[EMAIL PROTECTED] Sent: Saturday, December 15, 2007 7:36 AM To: solr-user@lucene.apache.org Subject: Re: Replication hooks - changing the index while the slave is running ... That helps Thanks for the prompt reply On Dec 15, 2007, at 10:15 AM, Yonik Seeley wrote: > On Dec 14, 2007 7:36 PM, Tracy Flynn > <[EMAIL PROTECTED]> wrote: >> 1) The existing index(es) being used by the Solr slave instance are >> physically deleted >> 2) The new index snapshots are renamed/moved from their temporary >> installation location to the default index location >> 3) The slave is sent a 'commit' to force a new IndexReader to start >> to read the new index. >> >> What happens to search requests against the existing/old index during >> step 1) and between steps 1 and 2? > > Search requests will still work on the old searcher/index. > >> Where do they get information if >> they need to go to disk for results that are not cached? Do they a) >> hang b) produce no results c) error in some other way? > > A lucene IndexReader keeps all the files open that aren't loaded into > memory... and external deletion has no effect on the ability to keep > reading these open files (they aren't really deleted yet). > > -Yonik
