Storable - persistence for Perl data structures |
Storable - persistence 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 persistence to your Perl data structures containing SCALAR, ARRAY, HASH or REF objects, i.e. anything that can be conveniently 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.
The objects stored into that file are recreated into memory for you,
and a reference to the root object is 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 explicitly 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 immediately thaws it out.
The lock_store
and lock_nstore
routine are equivalent to
store
and nstore
, except that they get an exclusive lock on
the file before writing. Likewise, lock_retrieve
does the same
as retrieve
, but also gets a shared lock on the file before reading.
As 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 such as 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 optimizations have been made when manipulating perl internals, to sacrifice encapsulation for the benefit of 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.
This release of Storable can be used on a newer version of Perl to
serialize data which is not supported by earlier Perls. By default,
Storable will attempt to do the right thing, by croak()
ing if it
encounters data that it cannot deserialize. However, the defaults
can be changed as follows:
croak()
.
To change this behaviour so that Storable deserializes utf8 encoded
values as the string of bytes (effectively dropping the is_utf8 flag)
set $Storable::drop_utf8
to some TRUE
value. This is a form of
data loss, because with $drop_utf8
true, it becomes impossible to tell
whether the original data was the Unicode string, or a series of bytes
that happen to be valid utf8.
croak()
instead, set
$Storable::downgrade_restricted
to a FALSE
value. To restore
the default set it back to some TRUE
value.
This version of Storable will defer croaking until it encounters a data type in the file that it does not recognize. This means that it will continue to read files generated by newer Storable modules which are careful in what they write out, making it easier to upgrade Storable modules in a mixed environment.
The old behaviour of immediate croaking can be re-instated by setting
$Storable::accept_future_minor
to some FALSE
value.
All these variables have no effect on a newer Perl which supports the relevant feature.
Storable uses the ``exception'' paradigm, in that it does not try to workaround
failures: if something bad happens, an exception is generated from the
caller's perspective (see the Carp manpage and croak()
). Use eval {} to trap
those exceptions.
When Storable croaks, it tries to report the error via the logcroak()
routine from the Log::Agent
package, if it is available.
Normal errors are reported by having store()
or retrieve()
return undef
.
Such errors are usually I/O errors (or truncated stream errors at retrieval).
Any class may define hooks that will be called during the serialization and deserialization process on objects that are instances of that class. Those hooks can redefine the way serialization is performed (and therefore, how the symmetrical deserialization should be conducted).
Since we said earlier:
dclone(.) = thaw(freeze(.))
everything we say about hooks should also hold for deep cloning. However, hooks get to know whether the operation is a mere serialization, or a cloning.
Therefore, when serializing hooks are involved,
dclone(.) <> thaw(freeze(.))
Well, you could keep them in sync, but there's no guarantee it will always hold on classes somebody else wrote. Besides, there is little to gain in doing so: a serializing hook could keep only one attribute of an object, which is probably not what should happen during a deep cloning of that same object.
Here is the hooking interface:
STORABLE_freeze
obj, cloningArguments: obj is the object to serialize, cloning is a flag indicating
whether we're in a dclone()
or a regular serialization via store()
or freeze().
Returned value: A LIST ($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, ...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 the perlmod manpage
warned you.
It is up to you to use this information to populate obj the way you want.
Returned value: none.
Predicates are not exportable. They must be called by explicitly prefixing them with the Storable package name.
Storable::last_op_in_netorder
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.
Storable::is_storing
Storable::is_retrieving
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 serializing and deserializing things, so why not use it to handle the serialization string?
There are a few things you need to know, however:
freeze()
(for instance) point back to the object we're trying to serialize in
the hook.
Shared references among objects will not stay shared: if we're serializing
the list of object [A, C] where both object A and C refer to the SAME object
B, and if there is a serializing hook in A that says freeze(B), then when
deserializing, we'll get [A', C'] where A' refers to B', but C' refers to D,
a deep clone of B'. The topology was not preserved.
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.
There is a 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.
Yes, there's a lot of that :-) But more precisely, in UNIX systems
there's a utility called file
, which recognizes data files based on
their contents (usually their first few bytes). For this to work,
a certain file called magic needs to taught about the signature
of the data. Where that configuration file lives depends on the UNIX
flavour; often it's something like /usr/share/misc/magic or
/etc/magic. Your system administrator needs to do the updating of
the magic file. The necessary signature information is output to
STDOUT by invoking Storable::show_file_magic(). Note that the GNU
implementation of the file
utility, version 3.38 or later,
is expected to contain support for recognising Storable files
out-of-the-box, in addition to other kinds of Perl files.
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 be disappointed 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 sequence of 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 temporary 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().
Storable::drop_utf8
is a blunt tool. There is no facility either to
return all strings as utf8 sequences, or to attempt to convert utf8
data back to 8 bit and croak()
if the conversion fails.
Prior to Storable 2.01, no distinction was made between signed and unsigned integers on storing. By default Storable prefers to store a scalars string representation (if it has one) so this would only cause problems when storing large unsigned integers that had never been coverted to string or floating point. In other words values that had been generated by integer operations such as logic ops and then not used in any string or arithmetic context before storing.
This section only applies to you if you have existing data written out by Storable 2.02 or earlier on perl 5.6.0 or 5.6.1 on Unix or Linux which has been configured with 64 bit integer support (not the default) If you got a precompiled perl, rather than running Configure to build your own perl from source, then it almost certainly does not affect you, and you can stop reading now (unless you're curious). If you're using perl on Windows it does not affect you.
Storable writes a file header which contains the sizes of various C language types for the C compiler that built Storable (when not writing in network order), and will refuse to load files written by a Storable not on the same (or compatible) architecture. This check and a check on machine byteorder is needed because the size of various fields in the file are given by the sizes of the C language types, and so files written on different architectures are incompatible. This is done for increased speed. (When writing in network order, all fields are written out as standard lengths, which allows full interworking, but takes longer to read and write)
Perl 5.6.x introduced the ability to optional configure the perl interpreter
to use C's long long
type to allow scalars to store 64 bit integers on 32
bit systems. However, due to the way the Perl configuration system
generated the C configuration files on non-Windows platforms, and the way
Storable generates its header, nothing in the Storable file header reflected
whether the perl writing was using 32 or 64 bit integers, despite the fact
that Storable was storing some data differently in the file. Hence Storable
running on perl with 64 bit integers will read the header from a file
written by a 32 bit perl, not realise that the data is actually in a subtly
incompatible format, and then go horribly wrong (possibly crashing) if it
encountered a stored integer. This is a design failure.
Storable has now been changed to write out and read in a file header with information about the size of integers. It's impossible to detect whether an old file being read in was written with 32 or 64 bit integers (they have the same header) so it's impossible to automatically switch to a correct backwards compatibility mode. Hence this Storable defaults to the new, correct behaviour.
What this means is that if you have data written by Storable 1.x running
on perl 5.6.0 or 5.6.1 configured with 64 bit integers on Unix or Linux
then by default this Storable will refuse to read it, giving the error
Byte order is not compatible. If you have such data then you you
should set $Storable::interwork_56_64bit
to a true value to make this
Storable read and write files with the old header. You should also
migrate your data, or any older perl you are communicating with, to this
current version of Storable.
If you don't have data written with specific configuration of perl described above, then you do not and should not do anything. Don't set the flag - not only will Storable on an identically configured perl refuse to load them, but Storable a differently configured perl will load them believing them to be correct for it, and then may well fail or crash part way through reading them.
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 references to tied items support.
Storable was written by Raphael Manfredi <Raphael_Manfredi@pobox.com> Maintenance is now done by the perl5-porters <perl5-porters@perl.org>
Please e-mail us with problems, bug fixes, comments and complaints, although if you have complements you should send them to Raphael. Please don't e-mail Raphael with problems, as he no longer works on Storable, and your message will be delayed while he forwards it to us.
Clone.
Storable - persistence for Perl data structures |