KinoSearch::Docs::FileLocking - Manage indexes on shared volumes.
use Sys::Hostname qw( hostname ); my $hostname = hostname() or die "Can't get unique hostname"; my $manager = KinoSearch::Index::IndexManager->new( host => $hostname ); # Index time: my $indexer = KinoSearch::Indexer->new( index => '/path/to/index', manager => $manager, ); # Search time: my $reader = KinoSearch::Index::IndexReader->open( index => '/path/to/index', manager => $manager, ); my $searcher = KinoSearch::Searcher->new( index => $reader );
Manage indexes on shared volumes.
Normally, index locking is an invisible process. Exclusive write access is controlled via lockfiles within the index directory and problems only arise if multiple processes attempt to acquire the write lock simultaneously; search-time processes do not ordinarily require locking at all.
On shared volumes, however, the default locking mechanism fails, and manual intervention becomes necessary.
Both read and write applications accessing an index on a shared volume need to identify themselves with a unique
host id, e.g. hostname or ip address. Knowing the host id makes it possible to tell which lockfiles belong to other machines and therefore must not be removed when the lockfile's pid number appears not to correspond to an active process.
At index-time, the danger is that multiple indexing processes from different machines which fail to specify a unique
host id can delete each others' lockfiles and then attempt to modify the index at the same time, causing index corruption. The search-time problem is more complex.
Once an index file is no longer listed in the most recent snapshot, Indexer attempts to delete it as part of a post-commit() cleanup routine. It is possible that at the moment an Indexer is deleting files which it believes no longer needed, a Searcher referencing an earlier snapshot is in fact using them. The more often that an index is either updated or searched, the more likely it is that this conflict will arise from time to time.
Ordinarily, the deletion attempts are not a problem. On a typical Unix volume, the files will be deleted in name only: any process which holds an open filehandle against a given file will continue to have access, and the file won't actually get vaporized until the last filehandle is cleared. Thanks to "delete on last close semantics", an Indexer can't truly delete the file out from underneath an active Searcher. On Windows, where file deletion fails whenever any process holds an open handle, the situation is different but still workable: Indexer just keeps retrying after each commit until deletion finally succeeds.
On NFS, however, the system breaks, because NFS allows files to be deleted out from underneath active processes. Should this happen, the unlucky read process will crash with a "Stale NFS filehandle" exception.
Under normal circumstances, it is neither necessary nor desirable for IndexReaders to secure read locks against an index, but for NFS we have to make an exception. LockFactory's make_shared_lock() method exists for this reason; supplying an IndexManager instance to IndexReader's constructor activates an internal locking mechanism using make_shared_lock() which prevents concurrent indexing processes from deleting files that are needed by active readers.
Since shared locks are implemented using lockfiles located in the index directory (as are exclusive locks), reader applications must have write access for read locking to work. Stale lock files from crashed processes are ordinarily cleared away the next time the same machine -- as identified by the
host parameter -- opens another IndexReader. (The classic technique of timing out lock files is not feasible because search processes may lie dormant indefinitely.) However, please be aware that if the last thing a given machine does is crash, lock files belonging to it may persist, preventing deletion of obsolete index data.
Copyright 2005-2010 Marvin Humphrey
This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself.