The previous discussion concentrated primarily on the composition from relational data to XML data. In the case of shredded XML, we must first solve the opposite problem: How do we take XML data and represent it in a relational database? Stated another way, How do we encode an XML document as a set of relational entities, relations, and attributes?
The issues involved for shredding are very similar to the issues for update for any composed mapping technique, which is not surprising, given that shredding an XML document is essentially an update operation that adds a new instance of the document to the store. Thus it is also not surprising that the same basic approaches that work for composed mapping techniques work for shredding as well—so long as they preserve the ability to update. In this section we explore a few additional issues that are unique to shredding.
Creation of the Database
First, when XML is being composed from relational data, we normally assume that the relational data already exists. That is, either the tables and schemas for the relational data have already been created for some other application, or else some application designer has expressly created tables for storing XML data—but in either case, the act of designing and installing the tables has been accomplished before we consider the details of how to map those tables to XML. With shredding, the opposite is usually the case. Initially we are presented with XML files, a DTD, or an XML schema, and we have to design and install the tables and columns that will hold the corresponding data.
Thus shredding always begins with a “design” or “registration” step, in which the database structures to hold the XML are actually created. Different vendors have different approaches to this. The process can be completely automated and implicit (analogous to a default mapping, but the other way around), or an explicit technique can be used to guide the creation of the table layout. Ideally, but not always, the same composition technique (e.g., annotated template) can be used to accomplish both shredding and (re)composition. An explicit composition technique may include additional content to provide instructions for things like precise mapping of datatypes or definition of table keys, foreign keys, and indices.
Adding Extra Information to the Composition
One difference between shredding and general composition techniques is that a shredding implementation has the opportunity to store extra “hidden” information that it derives while parsing an XML document. For example, it is possible to assign a unique, internal identifier (a GUID or row-id) to some or all nodes in an XML document and then use these identifiers as keys when generating hierarchy through joins, rather than insisting on the existence of some key in the XML itself. This makes it possible to reliably represent, and update, XML data that may contain duplicates, for example. (Note: It might be tempting to use XML ID attributes for this purpose, if they are present, but that is generally a bad idea, because (1) many applications will not validate that IDs are in fact unique within a document, and (2) IDs only need to be unique within a single document, and most applications need to manage multiple XML documents in the same database.)
In addition to identifiers for nodes, ordinal values (the position of a node within its parent) can be extracted and stored. Preserving order in XML documents can be a complex task otherwise and may even be impossible in some cases (e.g., arbitrarily interleaved lists of different types of elements). Care must be taken with the representation of the ordinal value if update is required (while a simple numeric index would work, the computational cost of inserting a new node into the document would be significant).
Inlining and Consolidation
Shredding involves two basic table layout choices: when to break information across multiple tables and when to consolidate tables for different elements. A simple algorithm for defining the database layout starts at the top of the XML document, with a root element (or set of possible root elements). The element is associated with a table, and then, for each of the possible children of that element, a decision is made whether to put them in the same table (inlining), or start a new table (and either find in the XML data, or create artificially, a field or fields that will be the join condition between the two tables).
The simplest rule for breaking XML information across multiple tables is to create a new table for every element that can occur multiply (has maxOccurs > 1) within its parent element. Elements that occur exactly once within a parent element are placed in the same table as the parent (and any reasonable shredding algorithm should be able to do this through multiple levels of singular elements). But what about optional elements? These could be placed in their own table, or they could be placed in the same table with the parent, using null when the element is not present. Inlining optional elements avoids an extra join at runtime, but can result in larger database sizes, particularly if the optional element is large and does not occur often in the data. Neither choice is optimal in all circumstances.
Consolidation is the process of re-using the same table for data that occurs in multiple locations in the database, most often because a common element type (e.g., something like an address) is re-used multiple times. For consolidation to be possible, a consistent set of fields must be used to join between the consolidated table and various parent tables (either the same fields are used in each case, or at least the entire set of fields is populated consistently). This is an easy condition to satisfy if node-ids are used as join conditions, but can be more difficult otherwise. Consolidation reduces the total number of tables required for a shredded representation.
Support of Full XML
The most important factor that affects shredding applications is that input XML sources can use arbitrary features of XML (and DTDs or XML Schema), including features that do not map well to relational storage. Examples of these features include poor mapping of datatypes (e.g., strings with unbounded length), variability of representation (e.g., xs:choice), ordered polymorphic sequences (particularly sequences containing repeating subsequences), mixed content, and arbitrary extensibility (e.g., xs:any). DTDs suffer from many of the same problems. XML with variable or no schema is even harder to handle. For variable schemas, if the schemas are known in advance, it may be possible to create a shredded representation that encompasses the union of all the possible structures. When the schema is not known in advance, or doesn't exist at all, it may be possible to guess a “shape” of the XML by looking at sample documents, but there is no guarantee that future documents will share that shape. In general, the more amorphous or varying the structure of the input XML, the more likely it is that a LOB representation will be more appropriate than a composed representation.
One possibility for handling the vagaries of full XML is to use a relational representation that is totally independent of the schema of the XML data. An example of such a representation is an edge table, in which a single table contains information about the relationship between parent and child elements (using node identifiers), and another column contains simple element content. Listing 6.14 provides an example of an XML fragment and an edge-table representation of it.
Listing 6.14 An Edge Table for an XML Fragment
<Department id="A12x" name="Physics"> <Employee id="555-23-4567"> <name>Mary Phillips</name> <office>21</office> </Employee> <Employee id="544-12-3456"> ... edge table parentid childid name value _____________________________________________________________________ null 1 Department null 1 2 @id A12x 1 3 @name Physics 1 4 Employee null 4 5 @id 555-23-4567 4 6 name Mary Phillips 4 7 office 21 1 8 Employee null 8 9 @id 544-12-3456
While edge tables or similar structures offer a certain attractive simplicity and can represent almost any XML input, they have not been widely used in layered processing models, since they do not take effective advantage of relational database capabilities: Every single navigation requires an independent join onto the same table or tables, and indexing is not very effective. However, the edge table can be a useful model for a native XML implementation within a relational database, in which the database can add additional crucial indexing information to the table (see Chapter 7 for a discussion).