NAME

Ref::Store::Guts - Implementation details

INTRODUCTION

Ref::Store (The name might change) aims to implement a dynamically garbage collected object lookup table. Lookup objects, and index by objects.

IMPLEMENTATION

All insertion operations consist of attaching a perl magic structure to the underlying SV object. This magic structure is used for chained garbage collection. The magic structure contains one more of an HR_Action structure which contains specifiers and hints about which type of action.

In effect, this is a second layer of back-referencing which extends the weakref concept (a common design pattern in general). I will call this back-deletion.

Values MUST be references, and keys are internally converted and/or encapsulated into (opaque) references themselves. More about this later.

LOOKUP TABLES

There are three primary lookup tables which are accessed internally:

Scalar Lookup Table

This table maps scalar 'user' strings to internal key objects. These 'strings' can also be object references, but have in common that they represent a user- provided identifier for our lookup. It looks something like this:

    $hash->store("foo", $value);
        # =>
        my $object_key = make_new_object_key("foo");
        $hash->scalar_lookup->{$foo} = $object_key;
Forward Lookup Table

This maps user strings to value objects. There is not much happening here.

Reverse Lookup Table

This maps a value object to its associated keys and attribute objects. The idea being that the value will 'own' its lookup properties. In the case of key lookups, the value acquires exclusive ownership, and in the case of attributes, it has shared ownership.

What this actually means is that the key object's primary reference is stored under the value's reverse lookup. Thus, when a value is deleted along with its reverse lookup table. Thus, it looks something like this;

    $reverse_lookup->{$value_address} = {
        $key1_address => $key1_object,
        $key2_address => $key2_object,
    };

KEY LOOKUPS

When a value is inserted into the database, two entries are created for it; one is a forward entry which maps the key to the value, and the other is an entry which maps the value back to the key. Since a value can have more than one key, the reverse lookup really contains multiple keys. The keys themselves are converted to objects, and what is stored in the value's reverse lookup table are key object references; each key object has a weak (or strong) reference indexed by its reference address.

In order to map the scalar/string key to a key object, yet another lookup table exists - a scalar lookup, which maps strings to their key objects.

Keys are converted to object references in order to aid garbage collection. Key objects contain back-deletes to their scalar lookup table as well as to their forward entries.

When either a key or a value is destroyed, the other always goes along with it. If a value is destroyed, the key's last strong reference (in the reverse table) is destroyed as well. When the last key is destroyed, its entry in the forward table is deleted as well -- and if this is the last key for the value, then the value is deleted as well.

To visualize in perl code, it would look something like

    $scalar_lookup->{"scalar string"} = $weak_reference_to_key;
    $reverse_lookup->{$value_refaddr}->{$key_refaddr} = $strong_reference_to_key;
    $forward_lookup->{"scalar string"} = $weak_or_strong_reference_to_value;

OBJECT-ENCAPSULATING KEYS

This lookup table supports not only simple scalar string keys, but also object keys. Object keys provide more versatility; for example, one can have a value automatically be deleted when its object key is destroyed, or vice versa. It thus allows creation of one-to-many mutual dependency relationships.

Object key implementation is a bit complex. Whereas string keys are always only deleted via the API, object keys can also be deleted whenever their encapsulating object goes out of scope.

As a result, some of the magic attached to a normal key object is transferred to the encapsulated object. Specifically, the encapsulated object will trigger deletion of the key object when the former is destroyed. The key object then also implements a DESTROY/cleanup method which does a subset of value deletion (if necessary).

ATTRIBUTE LOOKUPS

Attributes are different from keys in the sense that a single key can only hold a single value at any given time. Attributes can contain multiple values. A key is a unique identity of an object (though an object can have multiple identities, so long as they are unique), whereas an attribute is a property or state of any number of objects.

Attributes have a similar external API as keys, but are implemented slightly differently. Attributes cannot have an exclusive mapping to objects like keys do, and thus their implementation is slower and more complex.

The lookup mechanism for attributes is a bit more complex. Attribute objects reside inside an attribute lookup table (which is really the same as the scalar lookup table). Unlike keys, however, which store only a weak reference in the scalar lookup, attributes store a strong reference.

Attribute objects contain an internal hash which contains references to the value entries contained therein. References are optionally weak or strong, depending on user configuration, and values contain back-deletions to this internal hash structure.

The internal hash is special because it is tied. Every deletion operation is monitored, and when the key count becomes 0; that is, when no more values are using this attribute, the entry is deleted from the attribute lookup table.

PERFORMANCE AND OPERATIONS

KEY STORE:

new keys:

  1. Convert string to key object

  2. Store key object in the scalar lookup table

  3. Initialize magic on the key object

  4. Add back-delete to scalar lookup

  5. Insert into value's reverse hash

  6. If this is an encapsulating key, add magic to the object

new values:

  1. Initialize magic on value

  2. Initialize value's reverse entry

  3. Insert key in reverse entry. If the key is encapsulating, add back-delete for individual key's entry in reverse entry (This is needed to trigger the removal of magic from the encapsulated object).

  4. Add back-delete for reverse entry

KEY FETCH

  1. Look up user string and map to key object

  2. Lookup key-object ID in forward table, and return result

KEY DELETE

  1. Fetch (see FETCH)

  2. Remove key from scalary lookup

  3. Remove key from value's reverse lookup.

  4. Check if this is the last key for the value, if so then:

    • Delete vaue's reverse entry

    • De-initialize value's magic

VALUE DELETE

  1. Fetch reverse entry

  2. Delete all scalar entries for each value in the private reverse table

  3. Chain key-specific cleanup hooks

LICENSE AND COPYRIGHT

Copyright (C) 2011 M. Nunberg,

This program is free software; you can redistribute it and/or modify it under the terms of either: the GNU General Public License as published by the Free Software Foundation; or the Artistic License.

See http://dev.perl.org/licenses/ for more information.