Storable - persistency for perl data structures


 use Storable;
 store \%table, 'file';
 $hashref = retrieve('file');
 use Storable qw(nstore store_fd nstore_fd freeze thaw dclone);
 # Network order
 nstore \%table, 'file';
 $hashref = retrieve('file');   # There is NO nretrieve()
 # Storing to and retrieving from an already opened file
 store_fd \@array, \*STDOUT;
 nstore_fd \%table, \*STDOUT;
 $aryref = fd_retrieve(\*SOCKET);
 $hashref = fd_retrieve(\*SOCKET);
 # Serializing to memory
 $serialized = freeze \%table;
 %table_clone = %{ thaw($serialized) };
 # Deep (recursive) cloning
 $cloneref = dclone($ref);
 # Advisory locking
 use Storable qw(lock_store lock_nstore lock_retrieve)
 lock_store \%table, 'file';
 lock_nstore \%table, 'file';
 $hashref = lock_retrieve('file');


The Storable package brings persistency to your perl data structures containing SCALAR, ARRAY, HASH or REF objects, i.e. anything that can be convenientely stored to disk and retrieved at a later time.

It can be used in the regular procedural way by calling store with a reference to the object to be stored, along with the file name where the image should be written. The routine returns undef for I/O problems or other internal error, a true value otherwise. Serious errors are propagated as a die exception.

To retrieve data stored to disk, use retrieve with a file name, and the objects stored into that file are recreated into memory for you, a reference to the root object being returned. In case an I/O error occurs while reading, undef is returned instead. Other serious errors are propagated via die.

Since storage is performed recursively, you might want to stuff references to objects that share a lot of common data into a single array or hash table, and then store that object. That way, when you retrieve back the whole thing, the objects will continue to share what they originally shared.

At the cost of a slight header overhead, you may store to an already opened file descriptor using the store_fd routine, and retrieve from a file via fd_retrieve. Those names aren't imported by default, so you will have to do that explicitely if you need those routines. The file descriptor you supply must be already opened, for read if you're going to retrieve and for write if you wish to store.

        store_fd(\%table, *STDOUT) || die "can't store to stdout\n";
        $hashref = fd_retrieve(*STDIN);

You can also store data in network order to allow easy sharing across multiple platforms, or when storing on a socket known to be remotely connected. The routines to call have an initial n prefix for network, as in nstore and nstore_fd. At retrieval time, your data will be correctly restored so you don't have to know whether you're restoring from native or network ordered data. Double values are stored stringified to ensure portability as well, at the slight risk of loosing some precision in the last decimals.

When using fd_retrieve, objects are retrieved in sequence, one object (i.e. one recursive tree) per associated store_fd.

If you're more from the object-oriented camp, you can inherit from Storable and directly store your objects by invoking store as a method. The fact that the root of the to-be-stored tree is a blessed reference (i.e. an object) is special-cased so that the retrieve does not provide a reference to that object but rather the blessed object reference itself. (Otherwise, you'd get a reference to that blessed object).


The Storable engine can also store data into a Perl scalar instead, to later retrieve them. This is mainly used to freeze a complex structure in some safe compact memory place (where it can possibly be sent to another process via some IPC, since freezing the structure also serializes it in effect). Later on, and maybe somewhere else, you can thaw the Perl scalar out and recreate the original complex structure in memory.

Surprisingly, the routines to be called are named freeze and thaw. If you wish to send out the frozen scalar to another machine, use nfreeze instead to get a portable image.

Note that freezing an object structure and immediately thawing it actually achieves a deep cloning of that structure:

    dclone(.) = thaw(freeze(.))

Storable provides you with a dclone interface which does not create that intermediary scalar but instead freezes the structure in some internal memory space and then immediatly thaws it out.


The lock_store and lock_nstore routine are equivalent to store and nstore, only they get an exclusive lock on the file before writing. Likewise, lock_retrieve performs as retrieve, but also gets a shared lock on the file before reading.

Like with any advisory locking scheme, the protection only works if you systematically use lock_store and lock_retrieve. If one side of your application uses store whilst the other uses lock_retrieve, you will get no protection at all.

The internal advisory locking is implemented using Perl's flock() routine. If your system does not support any form of flock(), or if you share your files across NFS, you might wish to use other forms of locking by using modules like LockFile::Simple which lock a file using a filesystem entry, instead of locking the file descriptor.


The heart of Storable is written in C for decent speed. Extra low-level optimization have been made when manipulating perl internals, to sacrifice encapsulation for the benefit of a greater speed.


Normally Storable stores elements of hashes in the order they are stored internally by Perl, i.e. pseudo-randomly. If you set $Storable::canonical to some TRUE value, Storable will store hashes with the elements sorted by their key. This allows you to compare data structures by comparing their frozen representations (or even the compressed frozen representations), which can be useful for creating lookup tables for complicated queries.

Canonical order does not imply network order, those are two orthogonal settings.


Storable uses the ``exception'' paradigm, in that it does not try to workaround failures: if something bad happens, an exception is generated from the caller's perspective (see Carp and croak()). Use eval {} to trap those exceptions.

When Storable croaks, it tries to report the error via the logcroak() routine from the Log::Agent package, if it is available.

Normal errors are reported by having store() or retrieve() return undef. Such errors are usually I/O errors (or truncated stream errors at retrieval).



Any class may define hooks that will be called during the serialization and deserialization process on objects that are instances of that class. Those hooks can redefine the way serialization is performed (and therefore, how the symetrical deserialization should be conducted).

Since we said earlier:

    dclone(.) = thaw(freeze(.))

everything we say about hooks should also hold for deep cloning. However, hooks get to know whether the operation is a mere serialization, or a cloning.

Therefore, when serializing hooks are involved,

    dclone(.) <> thaw(freeze(.))

Well, you could keep them in sync, but there's no guarantee it will always hold on classes somebody else wrote. Besides, there is little to gain in doing so: a serializing hook could only keep one attribute of an object, which is probably not what should happen during a deep cloning of that same object.

Here is the hooking interface:

STORABLE_freeze obj, cloning
The serializing hook, called on the object during serialization. It can be inherited, or defined in the class itself, like any other method.

Arguments: obj is the object to serialize, cloning is a flag indicating whether we're in a dclone() or a regular serialization via store() or freeze().

Returned value: A LIST ($serialized, $ref1, $ref2, ...) where $serialized is the serialized form to be used, and the optional $ref1, $ref2, etc... are extra references that you wish to let the Storable engine serialize.

At deserialization time, you will be given back the same LIST, but all the extra references will be pointing into the deserialized structure.

The first time the hook is hit in a serialization flow, you may have it return an empty list. That will signal the Storable engine to further discard that hook for this class and to therefore revert to the default serialization of the underlying Perl data. The hook will again be normally processed in the next serialization.

Unless you know better, serializing hook should always say:

    sub STORABLE_freeze {
        my ($self, $cloning) = @_;
        return if $cloning;         # Regular default serialization

in order to keep reasonable dclone() semantics.

STORABLE_thaw obj, cloning, serialized, ...
The deserializing hook called on the object during deserialization. But wait. If we're deserializing, there's no object yet... right?

Wrong: the Storable engine creates an empty one for you. If you know Eiffel, you can view STORABLE_thaw as an alternate creation routine.

This means the hook can be inherited like any other method, and that obj is your blessed reference for this particular instance.

The other arguments should look familiar if you know STORABLE_freeze: cloning is true when we're part of a deep clone operation, serialized is the serialized string you returned to the engine in STORABLE_freeze, and there may be an optional list of references, in the same order you gave them at serialization time, pointing to the deserialized objects (which have been processed courtesy of the Storable engine).

When the Storable engine does not find any STORABLE_thaw hook routine, it tries to load the class by requiring the package dynamically (using the blessed package name), and then re-attempts the lookup. If at that time the hook cannot be located, the engine croaks. Note that this mechanism will fail if you define several classes in the same file, but perlmod(1) warned you.

It is up to you to use these information to populate obj the way you want.

Returned value: none.


Predicates are not exportable. They must be called by explicitely prefixing them with the Storable package name.

The Storable::last_op_in_netorder() predicate will tell you whether network order was used in the last store or retrieve operation. If you don't know how to use this, just forget about it.

Returns true if within a store operation (via STORABLE_freeze hook).

Returns true if within a retrieve operation, (via STORABLE_thaw hook).


With hooks comes the ability to recurse back to the Storable engine. Indeed, hooks are regular Perl code, and Storable is convenient when it comes to serialize and deserialize things, so why not use it to handle the serialization string?

There are a few things you need to know however:

That's why STORABLE_freeze lets you provide a list of references to serialize. The engine guarantees that those will be serialized in the same context as the other objects, and therefore that shared objects will stay shared.

In the above [A, C] example, the STORABLE_freeze hook could return:

        ("something", $self->{B})

and the B part would be serialized by the engine. In STORABLE_thaw, you would get back the reference to the B' object, deserialized for you.

Therefore, recursion should normally be avoided, but is nonetheless supported.

Deep Cloning

There is a new Clone module available on CPAN which implements deep cloning natively, i.e. without freezing to memory and thawing the result. It is aimed to replace Storable's dclone() some day. However, it does not currently support Storable hooks to redefine the way deep cloning is performed.


Here are some code samples showing a possible usage of Storable:

        use Storable qw(store retrieve freeze thaw dclone);
        %color = ('Blue' => 0.1, 'Red' => 0.8, 'Black' => 0, 'White' => 1);
        store(\%color, '/tmp/colors') or die "Can't store %a in /tmp/colors!\n";
        $colref = retrieve('/tmp/colors');
        die "Unable to retrieve from /tmp/colors!\n" unless defined $colref;
        printf "Blue is still %lf\n", $colref->{'Blue'};
        $colref2 = dclone(\%color);
        $str = freeze(\%color);
        printf "Serialization of %%color is %d bytes long.\n", length($str);
        $colref3 = thaw($str);

which prints (on my machine):

        Blue is still 0.100000
        Serialization of %color is 102 bytes long.


If you're using references as keys within your hash tables, you're bound to disapointment when retrieving your data. Indeed, Perl stringifies references used as hash table keys. If you later wish to access the items via another reference stringification (i.e. using the same reference that was used for the key originally to record the value into the hash table), it will work because both references stringify to the same string.

It won't work across a store and retrieve operations however, because the addresses in the retrieved objects, which are part of the stringified references, will probably differ from the original addresses. The topology of your structure is preserved, but not hidden semantics like those.

On platforms where it matters, be sure to call binmode() on the descriptors that you pass to Storable functions.

Storing data canonically that contains large hashes can be significantly slower than storing the same data normally, as temprorary arrays to hold the keys for each hash have to be allocated, populated, sorted and freed. Some tests have shown a halving of the speed of storing -- the exact penalty will depend on the complexity of your data. There is no slowdown on retrieval.


You can't store GLOB, CODE, FORMLINE, etc... If you can define semantics for those operations, feel free to enhance Storable so that it can deal with them.

The store functions will croak if they run into such references unless you set $Storable::forgive_me to some TRUE value. In that case, the fatal message is turned in a warning and some meaningless string is stored instead.

Setting $Storable::canonical may not yield frozen strings that compare equal due to possible stringification of numbers. When the string version of a scalar exists, it is the form stored, therefore if you happen to use your numbers as strings between two freezing operations on the same data structures, you will get different results.

When storing doubles in network order, their value is stored as text. However, you should also not expect non-numeric floating-point values such as infinity and ``not a number'' to pass successfully through a nstore()/retrieve() pair.

As Storable neither knows nor cares about character sets (although it does know that characters may be more than eight bits wide), any difference in the interpretation of character codes between a host and a target system is your problem. In particular, if host and target use different code points to represent the characters used in the text representation of floating-point numbers, you will not be able be able to exchange floating-point data, even with nstore().


Thank you to (in chronological order):

        Jarkko Hietaniemi <jhi@iki.fi>
        Ulrich Pfeifer <pfeifer@charly.informatik.uni-dortmund.de>
        Benjamin A. Holzman <bah@ecnvantage.com>
        Andrew Ford <A.Ford@ford-mason.co.uk>
        Gisle Aas <gisle@aas.no>
        Jeff Gresham <gresham_jeffrey@jpmorgan.com>
        Murray Nesbitt <murray@activestate.com>
        Marc Lehmann <pcg@opengroup.org>
        Justin Banks <justinb@wamnet.com>
        Jarkko Hietaniemi <jhi@iki.fi> (AGAIN, as perl 5.7.0 Pumpkin!)
        Salvador Ortiz Garcia <sog@msg.com.mx>
        Dominic Dunlop <domo@computer.org>
        Erik Haugan <erik@solbors.no>

for their bug reports, suggestions and contributions.

Benjamin Holzman contributed the tied variable support, Andrew Ford contributed the canonical order for hashes, and Gisle Aas fixed a few misunderstandings of mine regarding the Perl internals, and optimized the emission of ``tags'' in the output streams by simply counting the objects instead of tagging them (leading to a binary incompatibility for the Storable image starting at version 0.6--older images are of course still properly understood). Murray Nesbitt made Storable thread-safe. Marc Lehmann added overloading and reference to tied items support.


There is a Japanese translation of this man page available at http://member.nifty.ne.jp/hippo2000/perltips/storable.htm , courtesy of Kawai, Takanori <kawai@nippon-rad.co.jp>.


Raphael Manfredi <Raphael_Manfredi@pobox.com>