=head1 NAME

Tangram::Storage - persistent object database

=head1 SYNOPSIS

   use Tangram;
   
   $storage = Tangram::Storage->connect( $schema,
      $data_source, $username, $password );

   $oid = $storage->insert( $obj );
   @oids = $storage->insert( @objs );

   $storage->update( $obj );
   $storage->update( @objs );

   $obj = $storage->load( $oid );
   @objs = $storage->load( @oids );

   @objs = $storage->select( $class );
   @objs = $storage->select( $remote, $filter );

   $cursor = $storage->cursor( $remote, $filter );

   if ($storage->oid_isa($oid, "ClassName")) {
       # oid $oid is a ClassName
   }

   $storage->disconnect();

=head1 DESCRIPTION

A Tangram::Storage object is a connection to a database configured for
use with Tangram.

=head1 MEMORY MANAGEMENT

Starting with version 1.18, Tangram attempts to use the support for
weak reference that was introduced in Perl 5.6. Whether that support
is found or not has a major impact on how Storage influences object
lifetime.

If weakref support I<is> available, Storage uses weak references to
keep track of objects that have already been loaded. This does I<not>
prevent the objects from being reclaimed by Perl. IOW, the I<client>
code decides how long an object remains in memory.

If weakref support I<is not> available, Storage uses normal, 'strong'
references. Storage will pin in memory all the objects that have been
loaded I<and> inserted through it, until you call L<"disconnect"> or
L<"unload">.

In either case, Tangram will I<not> break circular structures for you.

Note that caching objects between transactions is a great way to ruin
the transactional guarantees that your database (hopefully) provides.

That being said, be sure to check out the C<unload_all()> method.

=head1 INTERNAL CONNECTION

Except in the implementation of cursor(), Tangram uses a single DBI
connection in its operations.  That connection is called the
'internal' connection. Since, in general, database managers do not
allow multiple result sets on the same connection, the internal
connection can be used only to carray a single task at a time.

L<Tangram::Cursor>s returned by cursor() do not suffer from this
limitation because they use a separate DBI connection.

=head1 CLASS METHODS

=head2 connect

   $storage = connect( $schema,
      $data_source, $username, $auth, \%options )

Connects to a storage and return a handle object. Dies in case of failure.

$schema is an L<Tangram::Schema> object consistent with the database.

$data_source, $username and $auth are passed directly to
DBI::connect().

\%options is a reference to a hash that may contain the following
fields:

=over 4

=item * dbh

Pass in an already connected DBI handle

=item * no_tx

Specify explicitly whether or not transactions are possible.  If they
are not, then Tangram can guarantee consistency by serialising
transaction updates - which guarantees poor performance and means that
you can never use C<$storage-E<gt>rollback>.

If you are using MySQL, you should consider using the InnoDB table
type to avoid this problem.  Also note that you will explicitly have
to set this option if you have InnoDB tables configured, as there is
no real way of telling if transactions are available for any given
query without either trying to do a rollback, or querying the table
types for every table.  Which I don't think it's Tangram's duty to do!

=item * no_subselects

Functions that need to perform sub-selects will die immediately or
attempt to emulate the functionality required, rather than relying on
the RDBMS to return a failure.

This is currently ignored, but that's not functionally relevant :-).
It can be read as C<$storage-E<gt>{no_subselects}> however, as the
correct value is automatically detected on connection.

=back

All fields are optional.

C<dbh> can be used to connect a Storage via an existing DBI
handle. $data_source, $username and $auth are still needed because
Tangram may need to open extra connections (see below).

=head1 INSTANCE METHODS

=head2 insert

   $storage->insert( @objs );

Inserts objects in storage. Returns the ID(s) assigned to the
object(s).  This method is valid in both L<scalar and list contexts>.

The inserted objects must be of a class described in the schema
associated to the storage.

Attempting to insert an object that is already persistent in the
storage is an error.

Tangram will automatically insert any object that is refered by $obj
if it is not already present in storage. In the following example:

   my $homer = NaturalPerson->new(
      firstName => 'Homer', name => 'Simpson',
      children => Set::Object->new(
         NaturalPerson->new(
            firstName => 'Bart', name => 'Simpson' ),
         NaturalPerson->new(
            firstName => 'Lisa', name => 'Simpson' ),
         NaturalPerson->new(
            firstName => 'Maggie', name => 'Simpson'
      ) ) );

   $storage->insert( $homer );

...Tangram automatically inserts the kids along with Homer.

=head2 update

   $storage->update( @objs );

Save objects to storage.  This method is valid in both L<scalar and
list contexts>.

The objects must be of a class described in the schema associated to
the storage.

Attempting to update an object that is not already present in the
storage is an error.

Tangram will automatically insert any object that is refered by an
inserted object if it is not already present in storage. It will not
automatically update the refered objects that are already stored. In
the following example:

   my $homer = NaturalPerson->new(
      firstName => 'Homer', name => 'Simpson' );
   $storage->insert( $homer );

   my $marge = NaturalPerson->new(
      firstName => 'Marge', name => 'Simpson',
      age => 34 );
   $storage->insert( $marge );

   $marge->{age} = 35;

   $homer->{partner} = $marge;

   $homer->{children} = Set::Object->new(
      NaturalPerson->new(
         firstName => 'Bart', name => 'Simpson' ),
      NaturalPerson->new(
         firstName => 'Lisa', name => 'Simpson' ),
      NaturalPerson->new(
         firstName => 'Maggie', name => 'Simpson' ) );

   $storage->update( $homer );

...Tangram automatically inserts the kids when their father is
updated. OTOH, $marge will not be automatically inserted nor updated;
her age will remain '34' in persistent storage.

Tangram does not perform any deadlock detection on updates.  You have
to rely on your database back-end for that.

=head2 id

   $id = $storage->id( $obj );
   @id = $storage->id( @obj );

Returns the IDs of the given objects. If an object is not persistent
in storage yet, its corresponding ID is undef().

This method is valid in both L<scalar and list contexts>.

=head2 oid_isa

   if ($storage->oid_isa($id, "ClassName")) {
      ...
   }

Checks that the passed Object ID, C<$id>, is a "ClassName" according
to the schema.  This check relies solely on the information in the
schema, not Perl's idea of C<-E<gt>isa> relationships.

=head2 load

   $obj = $storage->load( $id );
   @obj = $storage->load( @id );

Returns a list of objects given their IDs.
Dies if any ID has no corresponding persistent object in storage.

This method is valid in both L<scalar and list contexts>.

=head2 remote

   @remote = $storage->remote( @classes );

Returns a list of C<Tangram::Remote> objects of given classes.  See
L<Tangram::Remote> for a more detailed description.  These objects are
called I<remote> objects in the documentation.

=head2 select

   @objs = $storage->select( $remote );

   @objs = $storage->select( $remote, $filter );

   @objs = $storage->select( $remote,
      opt1 => val1, opt2 => val2, ...);

Valid only in list context. Returns a list containing all the objects
that satisfy C<$filter>.

$remote can be either a I<remote> object of an array of I<remote>
objects. If it is a single I<remote> object, a list of objects is
returned. If it is an array, a list of arrays of objects is returned.

If one argument is passed, return all the objects of the given type.

If two arguments are passed, the second argument must be a
Filter. C<select()> returns the objects that satisfy C<$filter> and
are type-compatible with the corresponding I<remote> object.

If more than two arguments are passed, the arguments after C<$remote>
are treated as key/value pairs.  Currently Tangram recognizes the
following directives:

=over 4

=item * filter

=item * distinct

=item * order

=item * desc

=item * distinct

=item * limit

=item * outer_filter

=item * force_outer

=back

C<filter> specifies a Filter that can be used to restrict the result
set.

Filters are based on simple Perl expressions involving I<remote> objects.
The expression is eventually compiled into its SQL equivalent, becoming
part of a WHERE-CLAUSE.

For example:

    my $remote_person = $storage->remote('Foo::Person');
    my @martians = $storage->select(
        $remote_person,
        filter => ($remote_person->{location} eq 'Mars')
    );

Would retrieve all martians from the database.

Note that the fields are accessed as hash reference keys instead of the
(expected) method calls.

In the previous example, C<-E<gt>{location}> is seen as a scalar from Perl
and as some derivative of a VARCHAR/TEXT field on the database side.
But filters can operate on many other types, including references to other
persistent objects. For instance:

    # instantiate the obj and add it to the DB
    my $mars = Foo::Location->new( name => 'Mars');
    $storage->insert($mars);

    my $remote_person = $storage->remote('Foo::Person');
    my @martians = $storage->select(
        $remote_person,
        filter => ($remote_person->{location} == $mars)
    );

In this case, having a reference to the persistent object C<$mars>
handy allows us to look for all objects that reference it. Keep in
mind that these are introductory examples - the relationship between
two classes of objects and how they behave depends on defined
relationships between them - whether it's a C<ref>, an C<array>, etc
-- see L<Tangram::Schema> and L<Tangram::Type> for more information on
relationship types.

Filters can also be joined together with boolean expressions:

    my $r_user = $storage->remote('My::Users');
    my @active_premium_users = $storage->select( $r_user,
        filter => (# "&" is not a typo - see below
                   ($r_user->{is_logged_in} eq 'Y') &
                   ($r_user->{is_premium} eq 'Y' )
                  )
    );

This select retrieves all the users currently logged in who also have
a premium account. Note the use of C<&> instead of C<&&> (or C<and>) -
this is due to a problem in the way Perl handles operator overloading
(C<&&> may not be overloaded).  For the basic boolean operators, use
C<&> as AND, C<|> as OR and C<!> as NOT.

Other overloaded bits that work as expected are:

    + - * / == eq != ne < lt <= le > gt >= ge cos sin acos

...which are translated to their SQL counterparts as closely as possible.

Tip: Filters can also be created beforehand by using this simple syntax:

    my $new_filter = ($r_user->{is_logged_in} eq 'Y');

Then you can add expressions to it by doing (for example):

    $new_filter &= (r_user->{is_premium} eq 'Y');

and use it in the expression like so:

    my @active_premium_users = $storage->select
        ( $r_user,
          filter => $new_filter
        );

As of Tangram 2.08_02, The scalar value C<1> may be used as an
"identity" filter.

See also C<Tangram::Expr>.

C<distinct> specifies that each object in the result set must be
unique (Tangram generates a SELECT DISTINCT).

C<order> specifies attributes in terms of one or more of the remote
objects - any that are being selected, or any that appear in the
filter.

As of Tangram 2.09, you can also directly use SQL expressions in
C<order> expressions, though you should consider how portable this
may or may not be.

C<desc> specifies that the order should be descending. For example:

    $storage->select( $object, filter => (...),
                      order => [ $remote_foo->{field1} ],
                      desc => 1  );

would order DESC (descending, high to low) all the fields listed in
the C<order> clause.

Passing:

                      desc => 0

would order all the fields ASC (ascending, low to high).

To specify which fields should be ordered DESC and which ones should be
ordered ASC, pass an array ref to C<desc>, like this:

    $storage->select( $object, filter => (...),
                      order => [
                                $remote_foo->{field1},
                                $remote_foo->{field2},
                                $remote_foo->{field3},
                               ],
                      desc => [ 1, 0, 1 ]  );

This will order C<field1> and C<field3> descending, and C<field2>
ascending.

C<distinct> is a boolean; a true value specifies that the same object
should ocur only once in the result set.  In general, this is a good
idea; 

C<limit> is a maximum number of rows to retrieve; in fact, with some
databases you can give two numbers to this to get the rows between N
and M of a select.  See your RDBMS manual for more.  If you want to
specify more than one number, you may use the following syntax:

   $storage->select( $object, filter => (...),
                     limit => [ 5, 10 ] );

The above example would return rows 6 through 15 on a MySQL database.

The select method is valid only in list context.

C<outer_filter> and C<force_outer> are EXPERIMENTAL API features.

If you pass any filter conditions into C<outer_filter> instead of
C<filter>, then any mentioned tables are connected by an outer join.
What this means is that the object does not necessarily have to be
present for the select to return a row; it may also be C<undef>.

The C<force_outer> option expects an array ref of L<Tangram::Remote>
objects.  These tables are joined with an outer join clause.

The outer join related code is extremely hairy, and you are advised to
ensure that you test each outer join query that you are going to use
with new versions of Tangram.

Do not try to combine inheritance and outer joins if you want to run
your application on toy databases, currently this means SQLite and
MySQL.  SQLite does not parse SQL nested join syntax and MySQL just
gets the join all wrong.  At least, on my testbed system.  YMMV.

=head2 sum( $expr, [$filter] )

Returns the total of the remote expression ($expr) for all rows that
match $filter, as summed by the RDBMS.  $filter is optional, and if
not passed the implication is to sum the value for ALL objects of that
type.

   my $r_thing = $storage->remote("Thing");
   $sum = $storage->sum( $r_thing->{field},
                         ($r_thing->{foo} eq "bar") );

It is also possible to pass a list of fields to sum, as an array ref:

   ($sum_expr1, $sum_expr2)
       = $storage->sum( [ $expr1, $expr2 ], $filter );

=head2 count( $expr, [$filter] )

Works as B<sum()>, but returns the count of the given objects or
columns instead of the sum.

This function does I<not> support counting multiple columns by passing
an array ref.  However, this can be achieved using the
C<-E<gt>count()> remote expression function (see L<Tangram::Expr>).

=head2 cursor

   $cursor = $storage->cursor( $remote );
   $cursor = $storage->cursor( $remote, $filter );
   $cursor = cursor( $remote,
      opt1 => val1, op2 => val2, ...);

Valid only in scalar context.

Returns a Cursor on the objects that are type-compatible with $remote.

If one argument is passed, the cursor returns all the objects of the given type.

If two arguments are passed, the second argument must be a
Filter. The cursor returns the objects that satisfy $filter and are
type-compatible with the corresponding Remote.

If more than two arguments are passed, the arguments after $remote
are treated as key/value pairs. Currently Tangram recognizes the following
directives:

=over 4

=item * filter

=item * order

=item * desc

=item * distinct

=item * retrieve

=back

For options C<filter>, C<order>, C<desc> and C<distinct>, see C<select>.

Option C<retrieve> is an array of Expr, to be retrieved in addition to
the object itself.

=head2 prefetch

   $storage->prefetch("Class", "collection", $filter);

This method fetches all the "collection" collections from "Class",
where $filter.

You need to be very careful with your filter - it is quite easy to end
up with a filter that will include a single table twice with no join.

You B<should not> include an expression in the filter that matches the
type of object that you are prefetching, unless that is a *different*
object to the one you want to load.

You B<should> replace the text "Class" with a Tangram::Remote object
from your C<$filter> if it appears in the expression.

This code is OK:

   my $r_parent = $storage->remote( "NaturalPerson" );
   my $filter = ($r_parent->{age} > 40);

   my @parent = $storage->select($r_parent, $filter);
   $storage->prefetch($r_parent, "children" $filter);

But this code has the problem:

   my $r_parent = $storage->remote( "NaturalPerson" );
   my $r_child  = $storage->remote( "NaturalPerson" );

   my $filter = (
                 ($r_parent->{age} > 40) &;
                  $r_parent->{children}->includes($r_child)
                );

   my @parent = $storage->select($r_parent, $filter);
   my @children = $storage->select($r_child, $filter);

   $storage->prefetch($r_parent, "children", $filter);

Because C<$filter> contains an extra `unnecessary' relationship with
C<$r_child>, the filter that Tangram builds internally ends up looking
like:

    (
     ($r_parent->{age} > 40) &
     $r_parent->{children}->includes($r_child) &
     $r_parent->{children}->includes($r_child2) &
    );

So, you end up including extra tables without joining them.  This
situation does not make any sense, but unfortunately because of the
definition of how RDBMS' work, it is required behaviour for it to give
you a permutation of all of the unjoined tables.  <sigh>

=head2 erase

   $storage->erase( @obj );

Removes objects from persistent storage. The objects remain present in transient storage.

=head2 tx_start

   $storage->tx_start();

Starts a new Tangram transaction.  Tangram transactions can be nested,
but currently this does not actually make SQL C<SAVEPOINT>'s (for
partial transaction rollback).

Instead, tangram maintains a transaction nesting count for each
storage object and commits the operations only when that count reaches
zero. This scheme makes it easy for a function to collaborate with its
caller in the management of the L<internal connection>.

Example:

   sub f
   {
      $storage->tx_start();
      $storage->update( $homer );
      $storage->tx_commit(); # or perhaps rollback()
   }

   sub g
   {
      $storage->tx_start();
      f();
      $storage->update( $marge );
      $storage->tx_commit(); # or perhaps rollback()
   }

   f(); # 1
   g(); # 2


In (1), f() commits the changes to $homer directly to the database.

In (2), f() transparently reuses the transaction opened by g().
Changes to both $homer and $marge are commited to the database when
g() calls tx_commit().

By default with ACID compliant database back-ends (such as Pg,
MySQL/InnoDB, Oracle and pretty much any commercial RDBMS), the first
time you open a database connection, you are beginning a transaction.
However, this is not the case with the L<Tangram::SQLite> or
L<Tangram::mysql> back-ends, both of which do not implement
transaction isolation; therefore it is not good to assume that the
database can handle concurrent writing efficiently.

To be run safely on these non-compliant back-ends, you should
explicitly C<tx_start()> at the beginning of transaction blocks rather
than relying on the default behaviour.

=head2 tx_commit

   $storage->tx_commit();

Commits the current Tangram transaction for this storage.  If the
transaction being commited is the outermost transaction for this
storage, the DBI transaction is also commited.

When using the L<SQLite> back-end, when the DBI transaction is
committed, the connection is also marked read-only (ie, AutoCommit is
enabled).

=head2 tx_rollback

   $storage->tx_rollback();

Rolls back the current Tangram transaction for this storage.
If the transaction being rolled back is the outermost transaction for this storage, the DBI
transaction is also rolled back.

=head2 tx_do

   $storage->tx_do( sub { ... } );

Executes CODEREF under the protection of a Tangram transaction and
pass it @args in the argument list.

Rolls back the transaction if CODEREF dies; in which case the
exception is re-thrown.

Returns the results of CODEREF, either as a scalar or as a list
depending on the context in which tx_do was called.

Example:

   $storage->tx_do(
      sub
      {
         $storage->update( $homer );
         # do things, die perhaps
         $storage->update( $marge );
      } );

Both $homer and $marge will be updated, or none will, depending on
whether the anonymous subroutine passed to tx_do() dies.

=head2 unload

   $storage->unload( @obj );

Drops references to persistent objects present in memory. @objs may
contain both objects and object ids. If @objs is empty, unloads all
the objects loaded by this storage.

Storage keeps track of all the persistent objects that are present in
memory, in order to make sure that loading the same object twice
results in a single copy of the object.

As a consequence, these objects will not be reclaimed by Perl's
automatic memory management mechanism until either disconnect() or
unload() is called.

unload() should be called only when no other references exist to
persistent objects, otherwise the same object (in the database) may
end up having two copies in transient storage, or vice versa!

In most cases, you never want to use this function - letting objects
pass out of scope and be cleaned up is a much more natural way to let
the object cache take care of itself.

=head2 unload_all( [ $notify_method ])

Drops references to all objects in the object cache.  If you pass a
notify method, then this will be passed to all objects as they are
dumped (so long as they C<-E<gt>can()> handle it).  This can be used,
for instance, with Class::Tangram objects to make sure all circular
references in cached objects are cleared, if you pass C<clear_refs> as
the C<$notify_method>.

Similar warnings apply to this function as C<$storage-E<gt>unload()>.

This function is particularly useful in OLTP (online transaction
processing) servers.  In those, it should be called before the first
C<$storage-E<gt>tx_start()>, so that all objects are known to be
"fresh" in the current transaction.  Due to ACID guarantees of
consistent reads etc (not on MySQL/MyISAM!), you should then not have
the classic "dirty read" problem - so long as you wrap the entire
transaction in a function that catches a failure on
C<-e<gt>tx_commit()> and attempts a retry (make sure to clear the
cache again before a retry!).

You might also want to see your RDBMS manual under the topic of
"transaction isolation", in particular the SQL command C<SET
TRANSACTION ISOLATION LEVEL>.

=head2 disconnect

   $storage->disconnect();

Disconnects from the database. Drops references to persistent
objects present in memory (see C<unload>).

=cut