sub SIG_FREE () { 'F' }
sub SIG_SIZE () { 1 }
+=head1 NAME
+
+DBM::Deep::Engine
+
+=head1 PURPOSE
+
+This is an internal-use-only object for L<DBM::Deep/>. It mediates the low-level
+mapping between the L<DBM::Deep/> objects and the storage medium.
+
+The purpose of this documentation is to provide low-level documentation for
+developers. It is B<not> intended to be used by the general public. This
+documentation and what it documents can and will change without notice.
+
+=head1 OVERVIEW
+
+The engine exposes an API to the DBM::Deep objects (DBM::Deep, DBM::Deep::Array,
+and DBM::Deep::Hash) for their use to access the actual stored values. This API
+is the following:
+
+=over 4
+
+=item * new
+
+=item * read_value
+
+=item * get_classname
+
+=item * make_reference
+
+=item * key_exists
+
+=item * delete_key
+
+=item * write_value
+
+=item * get_next_key
+
+=item * setup_fh
+
+=item * begin_work
+
+=item * commit
+
+=item * rollback
+
+=item * lock_exclusive
+
+=item * lock_shared
+
+=item * unlock
+
+=back
+
+They are explained in their own sections below. These methods, in turn, may
+provide some bounds-checking, but primarily act to instantiate objects in the
+Engine::Sector::* hierarchy and dispatch to them.
+
+=head1 TRANSACTIONS
+
+Transactions in DBM::Deep are implemented using a variant of MVCC. This attempts
+to keep the amount of actual work done against the file low while stil providing
+Atomicity, Consistency, and Isolation. Durability, unfortunately, cannot be done
+with only one file.
+
+=head2 STALENESS
+
+If another process uses a transaction slot and writes stuff to it, then
+terminates, the data that process wrote it still within the file. In order to
+address this, there is also a transaction staleness counter associated within
+every write. Each time a transaction is started, that process increments that
+transaction's staleness counter. If, when it reads a value, the staleness
+counters aren't identical, DBM::Deep will consider the value on disk to be stale
+and discard it.
+
+=head2 DURABILITY
+
+The fourth leg of ACID is Durability, the guarantee that when a commit returns,
+the data will be there the next time you read from it. This should be regardless
+of any crashes or powerdowns in between the commit and subsequent read.
+DBM::Deep does provide that guarantee; once the commit returns, all of the data
+has been transferred from the transaction shadow to the HEAD. The issue arises
+with partial commits - a commit that is interrupted in some fashion. In keeping
+with DBM::Deep's "tradition" of very light error-checking and non-existent
+error-handling, there is no way to recover from a partial commit. (This is
+probably a failure in Consistency as well as Durability.)
+
+Other DBMSes use transaction logs (a separate file, generally) to achieve
+Durability. As DBM::Deep is a single-file, we would have to do something
+similar to what SQLite and BDB do in terms of committing using synchonized
+writes. To do this, we would have to use a much higher RAM footprint and some
+serious programming that make my head hurts just to think about it.
+
+=cut
+
+
+
=head2 get_next_key( $obj, $prev_key )
This takes an object that provides _base_offset() and an optional string
projects / dbsrgits/DBM-Deep.git / blobdiff