8. C++ Developers Guide

8.1. Preface

8.1.1. Overview

This guide provides information about how to use the rasdaman database system (in short: rasdaman). The booklet explains usage of raslib, the rasdaman API, through its C++ binding.

Follow the instructions in this guide as you develop your application which makes use of rasdaman services. Explanations detail how, from within a C++ program, to create databases, collections, and instances; how to retrieve from databases; how to manipulate and delete instances within databases; how to influence physical storage parameters; how to do transaction handling and other administrative tasks.

The rasdaman interfaces are available on different operating system platforms. Although there are some differences in the way rasdaman appears in these different versions, the functionality is the same.

8.1.2. Audience

The information in this manual is intended primarily for application developers and for database administrators.

8.1.3. Rasdaman Documentation Set

This manual should be read in conjunction with the complete rasdaman documentation set which this guide is part of. The documentation set in its completeness covers all important infor­mat­ion needed to work with the rasdaman system, such as programming and query access to databases, guidance to utilities such as raswct, release notes, and additional information on the rasdaman wiki.

The rasdaman Documentation Set consists of the following docu­ments:

  • Installation and Administration Guide
  • Query Language Guide
  • C++ Developer’s Guide
  • Java Developer’s Guide
  • raswct Developer’s Guide
  • the rasdaman wiki, accessible at www.rasdaman.org

8.2. Introduction

See the corresponding Introduction Section in the rasdaman Query Language Guide.

8.3. Terminology

See the corresponding Terminology Section in the rasdaman Query Language Guide.

8.4. Application Examples

The following sections contain three examples of using the rasdaman API. Every example contains a code fragment including variable decla­rations and definitions, database open/close and transaction be­gin/commit statements. The numbers at the beginning of the code lines are used as references in the explaining text. Code segments which are in more than one example are explained where they occur first. For more clarity, error handling was omitted.

As raslib heavily makes use of templates, some platform specifics have to be considered when compiling and linking application programs. These are collected in Compilation and Linkage of Client Programs.

For details on the operational semantics of the rasdaman data model the reader is strongly encouraged to study the rasdaman Query Language Guide.

8.4.1. Basic Application Program Structure

Operation sequence

In order to access data in a database, variables have to be first defined and initialized, the database has to be opened and a transaction started. In the end, the transaction has to be committed and the data­base closed. Hence, an application basically consists of the following com­ponents (sample C++ code interspersed as far as rasdaman access is concerned):

  • Declaration and definition of database and transaction variables and other data like images or image sets

    r_Database database;
    r_Transaction transaction{&database};
    
  • Set the server name using the default port 7001.

    database.set_servername( "ServerName" );
    
  • Set user identification.

    database.set_useridentification( "me", "myPassword" );
    
  • Open the database.

    database.open( "DatabaseName" );
    
  • Begin the transaction.

    transaction.begin();
    
  • Work with the database.

  • Commit the transaction.

    transaction.commit();
    
  • Close the database.

    database.close();
    

Synchronous query execution

When a query is sent to the rasdaman server it will be executed in completeness - a running query cannot be aborted [1]. Care should be taken therefore not to start queries requir­ing resources beyond the capability of the server hardware and soft­ware environment, as the rasdaman service may be blocked for an indefinite time period.

8.4.2. Insertion of MDD

The following example creates a new MDD set with the name ULongSet and inserts two images into this set. The first image is initialized with zero, and the second one by way of some assumed initialization function.

  1. The variable declaration part includes one instance of type r_Database to represent the database and one instance of type r_Transaction to serve for the transaction handling. The do­main of type r_Minterval is used for specifying the spatial do­main of the images. In order to hold a persistent image, image has to be declared as an r_Ref pointer on the r_Marray struc­ture. The same applies for image_set which is an r_Ref to the set of images.

    r_Database database;
    r_Transaction transaction{&database};
    r_Minterval domain;
    r_Ref< r_Marray<r_ULong> > image;
    r_Ref< r_Set< r_Ref< r_Marray<r_ULong> > > > image_set;
    
  2. Server name and password are set (see Class r_Database).

    database.set_servername( "MyServer" );
    database.set_useridentification( "me", "myPassword" );
    
  3. An open message with the database name is sent to the data­base object.

    database.open( "MyDatabase" );
    
  4. The transaction is opened using the transaction object.

    transaction.begin();
    
  5. Memory for the image set is allocated using the new operator of class r_Object. As additional arguments, the new operator gets the database object in which it is to be inserted and the type name which was created in the database (see Type Definition Using rasql).

    image_set = new( &database, "ULongSet" )
                r_Set< r_Ref< r_Marray<r_ULong> > >(&transaction);
    
  6. To give a name to the set for later retrieval, a set_object_name message is sent to the database object.

    database.set_object_name(*image_set, "MyCollection");
    
  7. The spatial domain domain of the first is initial­ized with [1:91,1:91]. For doing so, a temporary two-dimensional object of type r_Minterval is filled with r_Sintervals specifying lower and upper bounds per dimension and then gets as­signed to domain.

    domain = r_Minterval(2) << r_Sinterval( 0l, 9l )
                            << r_Sinterval( 0l, 9l );
    
  8. Memory for a persistent object of type r_Marray is allocated us­ing the new operator of r_Ref. Again, the new operator gets the current database and the type name of the MDD object (creation of types is described in Type Definition Using rasql). The constructor of r_Marray gets the value zero which is used for initializing the whole MDD.

    image = new( &database, "ULongImage" )
            r_Marray<r_ULong>( domain, 0ul, &transaction );
    
  9. The image created in (7) is now inserted into the set. From now on, the persistent object is accessible via the collection.

    image_set->insert_element( image );
    
  10. The second image is created with a function pointer as second ar­gument for the r_Marray constructor. The function must be of type r_ULong (*initFunction)(const r_Point& pt). The function is invoked for each cell of the MDD with the current coordinates of the cell passed as the pt argument. The result value of type r_ULong is taken for the initial value of the cell.

    image = new( &database, "ULongImage" )
            r_Marray<r_ULong>( domain, &initWithCoordinates, &transaction );
    
  11. The image created in (9) is inserted into the set.

    image_set->insert_element( image );
    
  12. The transaction is committed. At this time, the image set is cre­ated in the database and the images are inserted. The data is made persistent and becomes visible to other transactions. The transient memory used to store the image on client side is freed and pointers to these objects (image_set and image) become invalid.

    transaction.commit();
    
  13. The last statement closes and disconnects the database.

    database.close();
    

For completeness, the following code segment describes the function used for initializing each cell of an MDD with the coordinates x+256*y:

r_ULong initWithCoordinates( const r_Point& pt )
{
    r_ULong value = pt[0] + pt[1] * 0x100;
    return value;
}

8.4.3. Lookup of an MDD set by its name

This example demonstrates retrieval of a set containing MDD objects as elements and iteration through the retrieved result set using raslib.

  1. An iteration variable named iterator is defined. It needs the ele­ment type of the collection being iterated as template argu­ment.

    r_Database database;
    r_Transaction transaction{&database};
    r_Ref< r_Set< r_Ref< r_GMarray > > > image_set;
    r_Ref< r_GMarray > image;
    r_Iterator< r_Ref< r_GMarray > > iter;
    
  2. A read-only transaction is started for the retrieval query. Read-only transactions should be used whenever possible, i.e., when no update operations occur within this transaction, in order to have maximal performance.

    database.set_servername( "ServerName" );
    database.set_useridentification( "me", "myPassword" );
    database.open( "DatabaseName" );
    transaction.begin( r_Transaction::read_only );
    
  3. The set is retrieved by sending a lookup_object message with the set name to the database object. At this moment, just a set of object identifiers is sent back to the client.

    image_set = database.lookup_object("CollectionName");
    
  4. The statement creates an iteration variable pointing to the first element of the set.

    iter = image_set->create_iterator();
    
  5. A simple for loop is used for iterating through the collection. An element of the collection, which is an r_Ref pointer to the MDD object, can be accessed by dereferencing the iteration variable iter. The image itself is retrieved from the server when the r_Ref pointer is dereferenced for the first time.

    for (iter.reset(); iter.not_done(); iter++)
    {
        image = (*iter);
        // work with the image
        // for example print its spatial domain
        cout << image->spatial_domain() << endl;
    }
    
  6. The query result is valid only until transaction end.

    transaction.commit();
    database.close();
    

8.4.4. Invocation of RasML statements

This example shows the creation and invocation of RasML queries us­ing the raslib classes:

  1. Two domains, a collection name, and a threshold value are de­fined to use them at creation stage of the RasML query.

    r_Minterval select_domain = r_Minterval("[0:4,0:4]");
    r_Minterval where_domain = r_Minterval("[8:9,8:9]");
    char collection_name[] = "CollectionName";
    r_ULong threshold_value = 10;
    
    r_Database database;
    r_Transaction transaction{&database};
    r_Set< r_Ref< r_GMarray > > image_set;
    r_Ref< r_GMarray > image;
    r_Iterator< r_Ref< r_GMarray > > iter;
    
    database.set_servername( "ServerName" );
    database.set_useridentification( "me", "myPassword" );
    database.open( "DatabaseName" );
    
  2. A read-only transaction is started for the retrieval query. Read-only transactions should be used whenever possible, i.e., when no update operations occur within this transaction, in order to have maximal performance.

    transaction.begin( r_Transaction::read_only );
    
  3. The query object of type r_OQL_Query is created and initialized with the parameterized query string.

    r_OQL_Query query( "select a$1 from $2 as a where some_cells( a$3 > $4 )" );
    
  4. The query parameters are filled using stream operators on the query object. First, the domain of type r_Minterval for the select part is applied, then the collection name, the domain for the where clause, and the threshold value are inserted.

    query << select_domain << collection_name << where_domain << threshold_value;
    

    The resulting query string looks like follows:

    select a[0:4,0:4]
    from CollectionName as a
    where some_cells( a[8:9,8:9] > 10 )
    
  5. Finally, the query is executed using the global function r_oql_execute. The query result is returned in the call-by-refer­ence parameter image_set. As query results are transient, the data of the whole result is sent to the client at this point.

    r_oql_execute( query, image_set, &transaction );
    iter = image_set.create_iterator();
    for( iter.reset(); iter.not_done(); iter++ )
    {
        image = (*iter);
        // work with the image
    }
    transaction.commit();
    database.close();
    

8.5. Raslib Classes

8.5.1. Overview

The raslib classes represent the rasdaman programming interface. It relies on the ODMG standard with some extensions sup­porting a smooth integration of the rasdaman-specific array struc­tures into the conventional C++ programming model.

raslib classes are categorized in

  • Type Classes providing type information for MDD objects,
  • Object Classes for handling persistent MDD objects,
  • System Classes for general tasks such as session maintenance and database querying,
  • Schema Access Classes to get runtime type information,
  • Stor­age Layout Classes for handling the storage structure, and
  • Error Classes for error handling.

8.5.2. Type Classes

_images/image41.png

Figure 8.1 Primitive Types

The types r_Long, r_ULong, r_Short, r_UShort, r_Octet, r_Char, r_Boolean, r_Float, and r_Double are atomic, serving as base types for MDD objects (Figure 8.1). Com­posite types built from atomic (primitive) or other complex (struc­tured) types are built using the record (struct) constructor.

Complex numbers, while by nature equivalent to a record structure {float re,im;}, are provided as a built-in type. Type complex implements complex numbers on single-precision float components while complexd implements double-precision.

Null values, i.e., values of cells which have not been assigned a value yet, always are the numerical zero value of the corresponding type. This extends in the obvious way to composite cells.

Table 8.1 Correspondence between rasql and C++ types
rasql C++ binding Length Description
octet r_Octet 8 bit signed integer
char r_Char 8 bit unsigned integer
short r_Short 16 bit signed integer
unsigned short r_Ushort 16 bit unsigned integer
long r_Long 32 bit signed integer
unsigned long r_Ulong 32 bit unsigned integer
float r_Float 32 bit single precision floating point
double r_Double 64 bit double precision floating point
boolean r_Boolean 1 bit [2] true (nonzero value) false (zero value)
complex r_Complex 64 bit Single precision complex number
complexd r_Complex 128 bit Double precision complex number

8.5.3. Object Classes

Object Classes are used for the data exchange with the database. They consist of classes able to generate and handle persistent arrays, i.e., arrays stored in a database, intervals, multidimensional intervals, multidimensional points, and scalar data which can either be atomic (primitive) or complex (structured). Figure 8.2 shows the object classes provided by rasdaman.

_images/image91.png

Figure 8.2 Object Classes

8.5.3.1. Class r_Point

Class r_Point handles multidimensional points.

Example

r_Point pointname( 5, 4 );

8.5.3.2. Class r_Sinterval

Class r_Sinterval represents a one-dimensional interval with lower and upper bound. Both bounds can either be fixed or variable (indi­cated by an asterisk ‘*’). Operations on intervals are defined following conventional interval arithmetics.

Example

r_Sinterval(100L, 200L)

specifies the interval [100:200].

8.5.3.3. Class r_Minterval

The spatial domain of an MDD is represented by an object of class r_Minterval (“multidimensional interval”). It specifies lower and upper bound of the point set for each dimension of an MDD. Internally, the class is implemented through an array of intervals of type r_Sinterval.

Example

r_Minterval intervalname("[0:100, 0:300]");

The object generated by the above expression yields the following output:

intervalname.dimension() = 2
intervalname[0].low()    = 0
intervalname[0].high()   = 100

8.5.3.4. Class r_OId

This handles object identifiers. Every array has a unique object identi­fier it can be addressed with.

8.5.3.5. Class r_Object

r_Object is an abstract class. Instances can only be generated from the non abstract classes inheriting from this class, that is r_Set, r_GMarray and r_Marray<T>. All these subclasses are capable of hav­ing persistent as well as transient instances and therefore are called persistent capable classes.

Objects of these classes can be generated using the overloaded new operator:

void* operator new( size_t size )                          // (1)
void* operator new( size_t size, r_Database* database,
                    const char* type_name = 0 )            // (2)
void* operator new( size_t size, const char* type_name )   // (3)
  1. is used to create transient objects. The only argument is the size of the new object.
  2. To generate persistent instances one also has to specify the data­base the object is to be inserted in.
  3. is the new operator for transient objects carrying type information.

Calling the delete operator

void operator delete( void* obj_ptr )

removes the object from memory and, in case it is a persistent object, from the database.

8.5.3.6. Classes r_Marray<T> and r_GMarray

The template class r_Marray<T> represents an MDD object over cell type T. Class r_GMarray is more generic in that it is able to represent MDD objects of any base type. This is necessary, firstly, for having a generic class for query results where the base type is not known at compile time and, secondly, for composite (multi-band) types.

The template class r_Marray<T> for specific base types inherits from r_GMarray; the constructor r_Marray<T>( r_GMarray& ) is provided for easy transformation to cell type safe m-arrays where the base type is known at compile time. Operations for accessing single cells are only available for r_Marray<T>.

8.5.3.7. Class r_Collection

r_Collection is an abstract class. It is the basic class of a collection. Possible subclasses are r_Set , r_Bag and r_List. The protected members isOrdered and allowsDuplicates are not initialized here, they have to be initialized in the respective subclasses. The method

virtual void insert_element ( const T& element, int no_modification = 0 )

inserts an element into the collection. If no_modification is set, the mark_modified() method of r_Object is not invoked and, therefore, a modification will not be recognized at the transaction commit point.

8.5.3.8. Class r_Set

The class implements a set container. It inherits most of the function­ality from r_Collection. The set can not have any duplicates and it is not ordered. The method

virtual void insert_element ( const T& element, int no_modification = 0 )

inserts an element into the collection. If no_modification is set, the mark_modified() method of r_Object is not invoked and, therefore, a modification will not be recognized at the commit point.

8.5.3.9. Classes r_Scalar, r_Primitive and r_Structure

The subclasses of r_Scalar are used to represent query results of the primitive types r_Boolean, r_Char, r_Octet, r_Short, r_UShort, r_Long, r_ULong, r_Float, r_Double and types composed of the primitive ones.

Class r_Primitive supports type-safe value access methods. r_Structure allows to access its elements by the subscript operator [].

Examples

The following line shows access to an unsigned short value:

r_Primitive primitive;
// ...
r_UShort value = primitive.get_ushort();

A structured value consisting of three long values can be accessed as follows:

r_Structure structuredValue;
// ...
for( int i=0; i<structuredValue.count_elements(); i++ )
{
    value = ((r_Primitive&)structuredValue[i]).get_long();
    //...
}

8.5.4. System Classes

_images/image51.png

Figure 8.3 System Classes

8.5.4.1. Class r_Database

Class r_Database allows to open and close connections to a specific database. The database name and the address of a running server manager must be indicated. Further optional parameters are

  • port number (default: 7001),
  • access mode (read/write or read-only; by default: read-only),
  • login (default: "rasguest")
  • password (default: "rasguest").

A database object must be instantiated and opened before starting any transaction on the database, and closed after end­ing these transactions (with a commit or abort).

Which Server to Contact?

Note that the server/port to be indicated must address the rasdaman server manager (not a particular rasdaman server); if in doubt, consult your system administrator.

Example

r_Database database;
database.set_servername( "Server Name" );
database.set_useridentification( "login name", "passwd" );
database.open( "Database Name" );
// ...
database.close( );

Storage Format

The r_Database class also allows to set the storage format, both for storage in MDD objects in the server and for their transfer between client and server. See Class r_Convertor and Subclasses for details.

8.5.4.2. Class r_Transaction

To use a transaction, an object of type r_Transaction has to be in­stantiated with an optional r_Database object as an argument (not thread-safe if the database parameter is not specified). Transactions can be started either in read/write or read-only mode, committed, aborted, and checkpointed. It is important to note that all access, creation, modification, and deletion of persistent objects must be done within a transaction. In order to achieve maximal performance, read-only transactions should be used when­ever possible, i.e., when no update operations occur within this trans­action. Right now checkpointing is not supported.

r_Transaction transaction{&database};
transaction.begin( );
// ...
transaction.commit( );

8.5.4.3. Classes r_Ref<T> and r_Ref_Any

An instance of template class r_Ref<T> is a reference to an instance of type T and is used to reference persistent sets (r_Set<T>) and MDD objects (r_GMarray and r_Marray<T>). It behaves like a normal C++ pointer but is capable of managing persistent data of type T within a transaction. In case the r_Ref<T> pointer is dereferenced (using the operator ->) and the object it is pointing to is not in the client memory yet, it is retrieved from the server.

The class r_Ref_Any is defined to support a reference to any type. Its primary purpose is to handle generic references and allow conversions of r_Ref<T> in the type hierarchy. A r_Ref_Any object can be used as an intermediary between any two types r_Ref<X> and r_Ref<Y> where X and Y are different types. A r_Ref<T> can always be converted to a r_Ref_Any; there is a function to perform the conversion in the r_Ref<T> template. Each r_Ref<T> class has a constructor and assignment operator that takes a reference to a r_Ref_Any.

8.5.4.4. Class r_Iterator<T>

The template class r_Iterator<T> defines the generic behavior for iteration. An object of this class can be used within a for loop for iterating through a collection of MDD objects. All iterators use a consistent protocol for sequentially returning each element from the collection over which the iteration is defined. When an iterator is constructed, it is either initialized with another iterator or is set to null. When an iterator is constructed via the method r_Collection<T>::create_iterator(), the iterator is initialized to point to the first element, if there is one.

8.5.4.5. Class r_OQL_Query and the freestanding function r_oql_execute()

A query statement is represented through an object of class r_OQL_Query (see Invocation of RasML statements). The r_OQL_Query constructor gets a query string which optionally can be parametrized. In this case, $i indicates the i-th parameter. The $i do not have to appear in a strict order - for example, $3 may appear before $2 in the statement.

The overloaded stream operator inserts the corresponding parameter values into the query, at the same time preserving their respective types. The query object expects parameters in the sequence of $1, $2, and so on. If any of the $i is not followed by a parameter at the point r_oql_execute() is called, an r_Error exception object of kind r_Error_QueryParameterCountInvalid will be thrown.

A query is executed against an open database through invocation of the freestanding function r_oql_execute(). This overloaded function exists in four variants:

void r_oql_execute( r_OQL_Query & query,
                    r_Transaction* transaction = nullptr )

void r_oql_execute( r_OQL_Query & query, r_Set<r_Ref_Any>& result, int dummy,
                    r_Transaction* transaction = nullptr );

void r_oql_execute( r_OQL_Query & query, r_Set<r_Ref<r_GMarray>> & result_set,
                    r_Transaction* transaction = nullptr )

void r_oql_execute( r_OQL_Query & query, r_Set<r_Ref<r_Any>> & result_set,
                    r_Transaction* transaction = nullptr )

The first version is used for insert (until v9.1), update, and delete statements where no result is passed back. The second version is used for insert queries, where the result contains the unique OID of the inserted object; the third parameter has no function and is there to distinguish this from the next two versions. The third version is for executing select statements where an MDD is returned; in this case, the second parameter receives the query result. The final case is for general query results which may also contain non-MDD return values, e.g., resulting from select oid(...) or select sdom(...) statements. This version will also be used when the result type of a query is not known in advance (i.e., at compile time). In this case, an r_Ref_Any object is returned, and the application is responsible for decoding the proper type. In support of this, r_Ref_Any objects contain their type information (see Example: Dynamic Type Information of a Query Result).

In all cases, the result_set parameter does not have to be initialised, and any previous contents is discarded by r_oql_execute().

Note

The transaction parameter is optional. If not specified, there is no guarantee on thread-safety (in fact queries will likely fail). The same holds for all other public API: if there is an r_Database or r_Transaction parameter, it is best to specify it to ensure correct usage in concurrent code.

Once a query has been executed via r_oql_execute(), the arguments associated with the $i parameters are cleared and new arguments must be supplied.

Example

The following code fragment creates a query string with two parameters $1 and $2.

r_OQL_Query query1( "select a$1 from $2 as a" );

Now two query parameters are generated:

r_Minterval select_domain = r_Minterval( 2 )
                         << r_Sinterval( 100L, 199L )
                         << r_Sinterval( 0L, 149L );
char collection_name[] = "mr";

Next, the parameters are attached to the query using the stream operator:

query1 << select_domain << collection_name;

The resulting query string is

select a[ 100:199, 0:149 ] from mr as a

Example

The following code shows how to attach an MDD object to an insert query:

r_Marray<r_Char> mddObject(...);                 // (1)
r_OQL_Query query("insert into mr1 values $1");  // (2)
query << mddObject;                              // (3)

Explanation:

  1. A transient MDD named mdd is created.
  2. The query object of type r_OQL_Query is initialized with an insert query statement including a placeholder $1.
  3. The MDD object is attached to the parameter $1 of the query.

8.5.5. Schema Access Classes

The rasdaman Schema Access Classes (cf. Figure 8.4) enable the user to determine the type of a query result at runtime.

_images/image101.png

Figure 8.4 Schema Access Classes

8.5.5.1. Class r_Meta_Object

Instances of class r_Meta_Object are used to describe elements of type information. The class holds a name standing for the type name of its instances.

8.5.5.2. Class r_Type

r_Type is an abstract base class for all type descriptions. It provides runtime type information through the method type_id() which returns a value of type r_Type_Id. It is an identifier of the following list:

BOOL, OCTET, CHAR, SHORT, USHORT, LONG, ULONG, FLOAT, DOUBLE,
STRUCTURETYPE, MARRAYTYPE, COLLECTIONTYPE, SINTERVALTYPE, MINTERVALTYPE,
POINTTYPE, OIDTYPE, COMPLEXTYPE1, COMPLEXTYPE2

8.5.5.3. Class r_Collection_Type

The class represents the type of a collection object. The type of the collection elements can be determined using method element_type().

8.5.5.4. Class r_Base_Type

r_Base_Type is an abstract base class for all type descriptions allowed as MDD base types which can either be primitive or structured types. The method size() delivers the size of a type instance in bytes.

8.5.5.5. Class r_Primitive_Type

This class represents all primitive types in the ODMG-conformant representation of the rasdaman type system.

8.5.5.6. Class r_Structure_Type

This class represents all user defined structured types in the ODMG-conformant representation of the rasdaman type system. They are returned using the method print_status(). Members are described by r_Attribute instances and represent the state or the structure. They can be accessed using an iterator of type attribute_iterator. Structures do not have object identity.

8.5.5.7. Class r_Property

This class is an abstract base class for all elements describing the state of an application-defined type. Right now, the only subclass is r_Attribute.

8.5.5.8. Class r_Attribute

An instance of r_Attribute describes an object or a literal. An attribute has a name and a type. The name is returned by the inherited method r_Meta_Object::name(). The type description of an attribute can be obtained using the inherited method r_Property::type_of(). The method offset() gives back the byte offset of the corresponding data area within a structure. If the attribute is not defined within a structure, the offset is zero.

Example

The structure

struct
{
    char red;
    char green;
    char blue;
};

has three attributes. The name of the third one, for example, is blue, its type is char and its offset 2.

8.5.5.9. Class r_Minterval_Type

The class represents the type of an r_Minterval object.

8.5.5.10. Class r_Sinterval_Type

The class represents the type of an r_Sinterval object.

8.5.5.11. Class r_Point_Type

The class represents the type of an r_Point object.

8.5.5.12. Class r_Marray_Type

The class represents the type of an r_Marray object. The base type of the MDD object can be determined using the method base_type().

8.5.5.13. Class r_Oid_Type

The class represents the type of an r_Oid object. The only meaningful comparison operations are equality and inequality of two OIDs.

8.5.5.14. Entry Points of the Type Schema

The type information can be accessed using one of the following methods:

const r_Type* r_Object::get_type_schema()

const r_Base_Type* r_GMarray::get_base_type_schema()

const r_Type* r_Collection::get_element_type_schema()

8.5.5.15. Example: Dynamic Type Information of a Query Result

In a query, new structures can be created which are not already defined in the database schema. For example, the following query forces the server to introduce an array type based on a 2-component cell structure:

select { img.red, img.green }
from rgb as img

Regardless of a result object’s type being a database type or created on the fly, the type information can be accessed using the previously introduced type functions. The following - incomplete - code piece prints out the type information associated with the MDD objects of a query result.

r_Bag< r_Ref_Any > result_set;
// ...query preparation...
r_oql_execute( query_object, result_set );
r_Iterator< r_Ref_Any > iter = result_set.create_iterator();
for( iter.reset(); iter.not_done(); iter++, i++ )
{
    switch( result_set.get_element_type_schema()->type_id() )
    {
    case r_Type::MARRAYTYPE:
        r_Ref<r_GMarray>(*iter)->print_status( cout );
        break;
    case r_Type::POINTTYPE:
        r_Ref<r_Point>(*iter)->print_status( cout );
        break;
    // etc.
    }
}

Note

A result set may contain structures other than MDD, e.g., when a spatial domain or some aggregate scalar is specified in the select clause. E.g. the query

select sdom( a ) [0].lo
from mr as a

returns a set of integer values.

8.5.6. Storage Layout Classes

A specialized storage structure for MDD objects is used in secondary storage, which is designed to provide fast access to persistent MDD objects for the most typical operations on such objects. This storage structure is configurable so that it is possible to set the different parameters (storage options) that define it. The storage options for an MDD object should be set depending on the access characteristics expected for that object. The current version allows to configure tiling (i.e., the subdivision algorithm used for the MDD objects) and storage format (i.e., the way how MDD tiles are encoded and compressed in the database and how MDD objects are compressed for client/server transfer).

Tiling is the subdivision of the MDD object into multidimensional blocks (tiles) of the same dimensionality as the MDD object. A tile is a multidimensional subarray of an MDD object. Tiling enables fast access to parts of an MDD, since only the tiles intersected by an access are retrieved by rasdaman. Tiling may be done in different ways, resulting in tiles with different formats and sizes. For example, tiles in a two dimensional image may be squares or rectangles with different sizes (Figure 8.5).

_images/image11.jpg

Figure 8.5 Tiling of a 2-D image.

In rasdaman, tiling is done according to a tiling scheme. Different tiling schemes allow the user to specify the subdivision of the domain in different ways. The choice of the tiling scheme and tiling parameters for an MDD object should be based on the most common type of access to the MDD object. The following tiling schemes are provided: aligned, default, directional, areas of interest and statistical tiling. All tiling schemes take into account the tile size parameter, which defines the maximum size in characters for individual tiles of the MDD object.

Aligned tiling divides the object into blocks which are aligned and have the same specified format. Default tiling is the tiling scheme used in case no specific tiling scheme is specified for an MDD object. It is a multidimensional block with sizes of equal lengths along all the directions of the domain. In directional tiling, the MDD object is divided into blocks defined by a partition of the domain of the MDD along different directions of the domain. This subdivision is appropriate for objects which are accessed through selection of linear ranges along only part of the directions of the domain.

The storage format indicates how tiles of an object are stored in the database. This addresses both encoding and compression. Some en­coding always has to be chosen; for compression, various alter­nat­ives are available, ranging from uncompressed storage over losslessly compressed to lossy compressed data.

An overview of the storage layout classes is given on Figure 8.6.

_images/image8.png

Figure 8.6 Storage Layout Classes

8.5.6.1. Class r_Storage_Layout

The classes of the r_Storage_Layout hierarchy are used to express the storage options for r_Marray objects. If an r_Storage_Layout object is passed to the r_Marray constructor, the options specified in it determine the structure of the object in persistent storage, otherwise, the default storage layout is used. It is important to note, however, that the notiling option of the client, activated by an environment variable, overrides the storage layout tiling options specified through r_Storage_Layout. If the rasdaman client is running with the option notiling, no tiling is done, independently of the storage layout chosen.

8.5.6.2. Class r_Tiling

Storage layout classes allow setting of the tiling option through instances of r_Tiling classes. When an r_Marray object is made persistent, in the rasdaman client the object is divided into blocks according to the tiling chosen for the object. These tiles are sent to the server and stored to constitute the MDD object. An index is built to access the tiles belonging to the MDD object.

Each derived class of r_Tiling implements a different decomposition method or tiling scheme. The following tiling classes are provided:

r_Aligned_Tiling

r_Dir_Tiling

r_Interest_Tiling

r_Stat_Tiling.

All these tiling schemes evaluate the tile size parameter tile_size which is the size of a tile in bytes. The default tile size is that specified for the rasdaman client.

Next, these tiling subclasses will be explained.

8.5.6.3. Class r_Aligned_Tiling

Aligned tiling is the regular tiling of an MDD object. Parameters provided are the tile format and tile size. The tile format specifies the sizes of a block along the different directions of the domain. These are interpreted as relative sizes. For example, if a [0:0,0:1] tile format is specified and a tile with exactly that format would have a size much smaller than the given tile size, that tile is stretched proportionally along all directions, so that the final tiles are twice as long in the second direction as in the first and have a size as close as possible to the tile size. An open interval (indicated by an asterisk “*”, see documentation for r_Sinterval and r_Minterval) along one of the directions specifies a direction of preferential access. Tiles will be made as long as possible in that direction.

8.5.6.4. Class r_Dir_Tiling

r_Dir_Tiling implements non-regular decomposition along specific directions of an MDD object. This tiling scheme allows a non-regular subdivision of the space. The user has to give the number of dimensions of the space and the decomposition wanted for each dimension.

8.5.6.5. Class r_Dir_Decompose

The r_Dir_Decompose class is used to specify a decomposition along one direction, i.e., dimension. The resulting tiling structure consists of a non-uniform grid where each grid line goes completely through the MDD and the distance between parallel gridlines is arbitrary.

An array of r_Dir_Decompose objects, with one element for each direction, must be provided.

Example

To specify tiling restrictions on the first two dimensions of a three-dimensional MDD object, the following code would apply:

r_Dir_Decompose decomp[3];
decomp[0] << 0 << 20 << 40 << 50;
decomp[1] << 0 << 15 << 20 << 50 << 60;
r_Dir_Tiling Tiling3DMDD( 3, decomp, ts );

ts in the last line specifies the tile size. The first and last elements put into the r_Dir_Decompose object must be the origin and limit of that dimension or a cross-section of the domain will occur (as if the elements outside the specification wouldn’t mind). In this code example the first dimension is going from 0 to 50 and the second one from 0 to 60.

8.5.6.6. Class r_Interest_Tiling

The class r_Interest_Tiling implements the areas of interest tiling algorithm. The user specifies which areas are of interest (areas which are accessed very often) and tiling is performed accordingly, in order to optimize access to those areas.

_images/image121.png

Figure 8.7 2-D MDD object with two areas of interest

Example:

If the areas [, 50:60] and [, 65:70] are of interest in the [0:1000,0:1000] domain, the following code does specification:

{
    // ...
    r_Minterval domain( "[0:1000,0:1000] ");
    r_Minterval interest1( "10:20,50:60] ");
    r_Minterval interest2( "[18:50,65:70] ");
    std::vector< r_Minterval > interest_areas;
    interest_areas.insert_element( interest1 );
    interest_areas.insert_element( interest2 );
    r_Interest_Tiling( interest_areas );
    // ...
}

In addition to the list of areas of interest, two further parameters can be passed to the constructor, which are default arguments of the constructor :

r_Interest_Tiling( r_Dimension dim,
                   const std::vector<r_Minterval>& interest_areas,
                   r_Bytes ts = RMInit::clientTileSize,
                   Tilesize_Limit strat = SUB_TILING )

ts specifies the tile size to be used, whereas strat is the tile size limitation strategy. The areas of interest algorithm splits the multi­dimensional array into tiles aligned with the areas of interest so that future accesses to those areas result in no cells outside the area being loaded from disk. In order to perform this, the algorithm first calculates a maximum partition of the space using the directional tiling algorithm. Since this is suboptimal and the resulting tiles might have sizes greater than clientTileSize it then performs further merges or subtiling, depending on the tile size limitation strategy. The supported options for it are the following:

  • NO_LIMIT: The blocks generated can have any size.
  • REGROUP: Only when performing grouping/merging of tiles, the size of the resulting tile of two merges is checked against clientTileSize. If it is larger, they are not merged. Tiles larger than clientTileSize may exist (for instance, if the user specifies an interest area with a size larger than clientTileSize).
  • SUB_TILING: In this strategy, regrouping is done regardless of the size of the generated tiles. After all the blocks are created, sub-tiling is performed on those whose size is larger than the tile size.
  • REGROUP_AND_SUBTILING: This combines the last two strategies. When merging blocks, tiles larger than clientTileSize are never created and, when the final tiles are all created, sub-tiling is performed on those whose size is larger then clientTileSize.

8.5.6.7. Class r_Stat_Tiling and r_Access

These classes support statistic tiling and specification of access patterns, respectively. Statistic tiling splits MDD objects based on the access patterns passed to it as a parameter. It actually detects areas of interest out of a set of accesses and then performs tiling by using the areas of interest tiling algorithm. In order to determine the areas of interest, the algorithm performs a check of overlapping accesses to reduce accesses which correspond to the same area of interest to one single area of interest. In this step, the criteria used to reduce a set of accesses to a single area of interest is that if a group of accesses are near up to a given threshold, then they correspond to a single area of interest which is the minimum interval covering the accesses.

The statistic tiling algorithm then eliminates some of the areas of interest. It performs a check of the number of times each of the detected areas was accessed. Those which were accessed less than a given threshold are eliminated (they are accessed too few times to be considered areas of interest).

Five parameters are passed in the constructor of the r_Stat_Tiling class:

r_Stat_Tiling( r_Dimension dim,
               const std::vector<r_Access>& stat_info,
               r_Bytes ts = RMInit::clientTileSize,
               r_Area border_threshold = DEF_BORDER_THR,
               r_Double interesting_threshold = DEF_INTERESTING_THR )

border_th is the border threshold for considering two access patterns to be the same, interesting_th is the interesting threshold, i.e., the percentage of accesses that must take place so that an area is considered being of interest when performing tiling and also ts, the tile size.

A call to merge() should be made prior to performing tiling, so that the statistic information about the accesses to the object can be updated and the tiling operation prepared.

r_Access merge(const std::vector<r_Access>& patterns)

This method inputs the statistic information into the class and calculates the new interest areas that will be used to perform tiling on the object. r_Stat_Tiling contains a list with the statistical information. This list is updated by the method. At the end, the list will contain the filtered and updated accesses count. This information can be used again as input to the method, or it can be stored for later usage.

The class r_Access represents an access pattern to a certain object. r_Stat_Tiling receives a list of these objects so that an appropriate tiling can be defined. The r_Access constructor

r_Access( const r_Minterval& region,
          unsigned long accesses = 1 )

takes as parameter the interval and the number of times the MDD subarray with domain region was accessed.

8.5.6.8. Class r_Convertor and Subclasses

The storage format indicator specifies the compression method used to compress / decompress tiles written to / retrieved from the data­base.

The transfer format indicator specifies the compression method used to compress / decompress tiles when transferred between client and server.

By default storage and transfer format is r_Array which means encoding in the server’s main memory format, without any comp­ress­ion. The r_Database function set_transfer_format() allows to change transfer format and compression, for both directions uni­formly:

void set_transfer_format( r_Data_Format format,
                          const char *formatParams=NULL )

The storage format in the server for MDD objects newly created by the client and its currently open transaction is set through set_storage_format():

void set_storage_format( r_Data_Format format,
                         const char *formatParams=NULL)

Both functions understand these parameters, defined in the enumeration type enum r_Data_Format in module raslib, see Table 8.2.

Table 8.2 Storage and transfer formats and their parameters
Compression type Constant Description
“direct” storage r_Array no compression, row-major memory representation
Data exchange format r_TIFF TIFF format (2-D images, non-compressing)
r_JPEG JPEG format (2-D, lossy compression; not recommended!)
r_HDF HDF format (n-D, non-compressing)
r_PNG PNG format (2-D images, lossless compression)
r_BMP BMP format (2-D images, non-compressing)
r_VFF VFF format (3-D data, non-compressing)
r_PPM PPM format (2-D binary/gray/colour images, lossless)
R_TOR TOR format (used for 2-D geo laser scan images, non-compressing)
R_DEM ASCII format for 2-D digital elevation data (non-compressing)
Dedicated compressions (lossy if not indicated otherwise) R_Auto automatic compression (lossless)
R_Zlib ZLIB compression (lossless)
R_RLE RLE compression (lossless)
R_Wavelet _Haar Haar Wavelet compression
r_Wavelet _Daubechies Daubechies 4-tap Wavelet compression
r_Sep_Zlib ZLIB compression, compress base types separately (lossless)
r_Sep_RLE RLE compression, compress base types separately (lossless)
r_Wavelet _Daub Daubechies n-tap Wavelet compression, n=6, 8, …, 18, 20
r_Wavelet _Least Least asymmetric n-tap Wavelet comp., n=8, 10, …, 18, 20
r_Wavelet _Coiflet Coiflet n-tap Wavelet compression, n=6, 12, 18, 24, 30
r_Wavelet _Qhaar Lossy Haar Wavelet compression

Recommendations

  • If space is not an issue, use r_Array storage for optimal performance.
  • If compression is desired, use r_RLE for relatively homogeneous data, r_Zlib in general. R_Sep_Zlib and r_sep_RLE give an advantage in the compression rate whenever the cell type has a larger number (say, 3 and above) of cell components. All these compress lossless, i.e. a compressed object inserted into the database will look the same after extraction.
  • Use lossy compression only if you are sure that database users can live with information being filtered out of the original data.
  • Almost all of the above formats have further parameters which allow fine tuning. They are passed in a string as comma-separated "name=value" pairs. See the r_Convertor class HTML documentat­ion for the admissible names and values.
  • Moreover, a white paper is available from rasdaman GmbH if you really want to go into the gory details.

Warning

From the “dedicated compression formats” listed above, only the RLE, SepRLE, Zlib, and SepZlib algorithms are fully released. The wavelet algorithms are provided as beta versions only, using them for non-experimental purposes is not recommended in the current version.

8.5.7. Error Classes

Figure 8.8 gives an overview on the rasdaman classes used to report on error situations:

_images/image131.png

Figure 8.8 rasdaman Error Classes

8.5.7.1. Class r_Error

This class implements the relevant part of the ODMG C++ binding’s r_Error class. It extends exception handling through deriving special classes for MDD specific errors. An error object consists of

  • an error number which serves to uniquely identify the error,
  • an error kind,
  • an error text which verbally describes the error.

The error number, hence, serves as an index to a generic textual description of the error.

Error texts are loaded from the text file errtxts located in $RMANHOME/bin using the initialisation function initTextTable(). This mechanism allows the system administrator to translate error messages into target languages other than English.

If no error number is specified, the error kind will be used as error text.

The error description is received calling the member function what().

Further information on error messages can be found in Error Messages.

Example

The following code fragment shows a typical try-catch block printing any potential error reported by rasdaman.

try
{
    // rasdaman access
}
catch( r_Error& errorObj )
{
    cerr << errorObj.what() << endl;
}

8.5.7.2. Class r_Eno_interval

This class represents an error object saying that the result is not an interval.

8.5.7.3. Class r_Eindex_violation

r_Eindex_violation represents an error object saying that the specified index is not within the bounds of the MDD object. In case the spatial domain of object a is [0:199] and the user asks for a[300] an error message of this class is raised.

8.5.7.4. Class r_Edim_mismatch

This class represents an error object saying that the dimensionalities of two objects do not match.

8.5.7.5. Class r_Eno_cell

r_Eno_cell represents an error object saying that the result is no cell. This happens f.e. if the cast operator for casting to the base type of class r_Marray is invoked on an object which is not ‘zero-dimensional’.

8.5.7.6. Class r_Einit_overflow

This class represents an error object saying that an initialization overflow occured. This happens, e.g., if the stream operator is invoked more often than the object has dimensions.

8.5.7.7. Class r_Equery_execution_failed

The class is used for errors occuring through query execution. In most cases, the position which caused the error can be fixed. This position is specified by line number, column number, and the token which is involved. Additionally, the class is generic concerning the error type. Different error types can be specified by stating the error number.

Example

The following code segment shows possible error handling after query execution:

try
{
    // execute a rasdaman query
}
catch( r_Equery_execution_failed& errorObj )
{
    cerr << errorObj.what() << endl;
    cerr << "Line No " << errorObj.get_lineno();
    cerr << "Column No " << errorObj.get_columnno();
    cerr << "Token " << errorObj.get_token();
}

8.6. Linking MDD with Other Data

8.6.1. Sessions

Applications always maintain raster data and descriptive alphanumeric data. The latter often are called metadata - a term we adopt for the purpose of this discussion. Actually, all over the world a lot of effort already has been put into metadata modelling, and many database structures and metadata applications have been developed success­fully. rasdaman does not reinvent the wheel: metadata remain un­changed in their (relational or object-oriented) database; they are not touched by rasdaman, but remain under the sole control of the underlying conventional DBMS (in the rasdaman documentation also referred to as “base DBMS”).

Therefore, to work simultaneously with rasdaman and metadata, an application has to open both a rasdaman database and the database containing the metadata, and it must begin separate transactions in both databases.

Opening of database in rasdaman and the metadata DBMS are completely independent from each other, likewise are transactions in both systems. They can be nested or interleaved in any way.

In order to embed MDD objects and MDD collections in underlying databases, object identifiers and collection names may be used. These constitute references to rasdaman objects (which are stored in the base DBMS).

8.6.2. Collection Names

MDD collections in rasdaman must be named. This name can then be used by an application as a reference to the MDD collection. The most typical usage of these collection names is their storage in a base DBMS object or tuple in order to reference an MDD collection which is related to the object or tuple.

This is illustrated in the following example:

class Patient
{
    // ...
    private:
    d_String name;
    d_Date birthday;
    Address residence;
    SocialSecurityNumber ssn;
    //reference to rasdaman MDD collection:
    d_String XrayCollectionName;
    // ...
};

8.6.3. Object Identifiers

Each MDD object is uniquely identified in rasdaman by an object identifier. Object identifiers are implemented by the r_OId class. A globally unique object identifier has three components describing

  • the system where it was created (system name),
  • the database (base name) and
  • the local object ID within the database.

The object identifier of a rasdaman object is returned by:

r_OId& r_Object::get_oid()

The object identifier may be used as a reference in an underlying data­base.

To be used as a reference in the underlying database the object identifier of a rasdaman object is stored as a member in an object of the underlying database. This is illustrated by the following example:

class SatelliteImage
{
    private:
    Date acquisitionDate;
    Location acquisitionLoc;
    // local reference to rasdaman MDD object:
    double imageRasOid;
    // ...
}

The member variable imageRasOid has to be translated into a rasdaman object identifier. This translation is done by the r_OId constructor:

r_OId::r_OId( const char* )

The string representation for a specific object identifier is returned by:

const char* r_OId::get_string_representation( )

Of course, alternatively the object identifier could be stored in its string representation.

8.7. Compilation and Linkage of Client Programs

8.7.1. Compilation

C++ applications using rasdaman have to include the header file rasdaman.hh which resides in $RMANHOME/include. Technically, rasdaman.hh includes further header files taken from the sub­direct­ories of $RMANHOME/include.

The class library makes intensive use of templates. As templates are handled differently by the various compilers, individual measures have to be taken. To this end, the header files are instrumented to recognise the variable OSTYPE indicating the system platform. For example, setting OSTYPE to linux-gnu (case-sensitive!) indicates a Linux/Gnu environ­ment, whereas the value solaris indicates a SUN/ Solaris platform. You should contact your dealer to find out which plat­forms are supported.

While in the deliverable sources (including the Makefiles provided) platform issues are dealt with, it nevertheless is important to under­stand the particularities. Therefore, some considerations follow next. If in doubt, you may want to contact the hotline.

Gnu

With the Gnu C++ compiler, the good way to handle templates is by early template instantiation using the compile flag -DEARLY_TEMPLATE. A template instantiation source file, template_inst.hh, is provided in the $RMANHOME/include/raslib directory; if the OSTYPE variable is set to linux-gnu, then this instantiation file will be included automatically.

Microsoft

With the Microsoft Visual C++ compiler, situation is similar as with Gnu above: it also needs early template instantiation.

Solaris

With the SUN-provided C++ compiler under Solaris, template instantiation at compile time is done by looking at the .cc files in the $RMANHOME/include subdirectories.

8.7.2. Linkage

For the linkage of an executable several libraries are needed. Those delivered with rasdaman are located in the $RMANHOME/lib directory.

One common problem are the dynamic libraries needed, such as libXmu.so. Usually there are different versions around. The version needed by a rasdaman application can be found out with the Unix ldd command which, for example, states:

libtiff.so.3 => /usr/lib/libtiff.so.3 (0x4001b000)
libstdc++-libc6.1-2.so.3 => not found
libXmu.so.6 => /usr/X11R6/lib/libXmu.so.6 (0x4005e000)
libXt.so.6 => /usr/X11R6/lib/libXt.so.6 (0x40071000)
libX11.so.6 => /usr/X11R6/lib/libX11.so.6 (0x400bc000)
libz.so.1 => /usr/lib/libz.so.1 (0x40160000)
libm.so.6 => /lib/libm.so.6 (0x4016f000)
libc.so.6 => /lib/libc.so.6 (0x4018c000)
libjpeg.so.62 => /usr/lib/libjpeg.so.62 (0x40281000)
libSM.so.6 => /usr/X11R6/lib/libSM.so.6 (0x402a0000)
libICE.so.6 => /usr/X11R6/lib/libICE.so.6 (0x402ab000)
libXext.so.6 => /usr/X11R6/lib/libXext.so.6 (0x402c2000)
/lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000)

As can be seen in the second line, libstdc++-libc6.1-2.so.3 cannot be found where­as all other references to dynamic libraries can be resolved. Sometimes a straightforward link to an older version helps, such as

ln -s libstdc++-libc6.1-2.so.2 $RMANHOME/lib/libstdc++-libc6.1-2.so.3

Obviously very much care should be taken when fooling the system like this, and it is certainly not the recommended way.

Another common problem is to put the libraries into the right order in the link command, and which of them have to be linked twice to resolve all referenced symbols.

Some working examples can be found in the Makefiles of the delivered examples.

8.7.3. Client Environment Parameters

To allow for easier application steering, raslib evaluates the environ­ment parameter RMANCLIENTOPT at program start-up. This variable can contain options similar to command line option syntax.

If contradicting options are set (e.g., -tiling and -notiling), then the last occurrence wins.

Options Known

Note: deprecated with the default rasnet server/client protocol.

-timeout set server communication timeout seconds (default: 3600)
-notimeout disable timeout, wait forever if necessary
-tilesize set tile size bytes (default: 100000)
-notiling disable client-side tiling
-l logfile set log stream to logfile (default: ./client.log)

Example

The following shell dialog shows how an environment is set before invoking a rasdaman client. Settings done are: use timeout of 5 seconds, write log output to /dev/null.

$ export RMANCLIENTOPT="-timeout 5 -l /dev/null"
$ rview

8.7.4. The Example Programs

An example program is delivered in $RMANHOME/examples/c++. This query program sends a rasql query to the rasdaman server and prints the result retrieved.

The code is documented and produces ample screen output, so it should be self explanatory. The programs are built by invoking make in the corresponding subdirectory.

Note

Before the test programs can be used, the rasdaman database has to be initialized.

8.8. HTML Documentation

All classes are described extensively in a set of HTML files shipped with the software. Starting point into the whole documentation is $RMANHOME/doc/index.html. Follow the “raslib” link to enter the description of the C++ interface.

The documentation can be viewed with any Web browser. Only graphical traversal between classes requires Java enabled; however, all links are available in textual form, too.

Top-level entry to the documentation shows the alphabetical listing of definitions, classes and functions; alternatively the class hierarchy display can be selected. Every class name is linked to the related class documentation. The subclass / superclass relations are indicated as indentation levels in the class list. Clicking a class name expands into the full class documentation consisting of three components.

First, there is the class inheritance hierarchy, including links to the direct subclasses and superclasses. The second part gives a short description of all class components, some of which have additional links to a more detailed documentation in the third part of the page. In this third part there is a detailed description of what the class does. Every time a class is used inside method declarations as either a para­meter or return value, a link to the documentation of this class is provided.

[1]This has nothing to do with transactions - after each completion of a query, the embracing transaction can be aborted indeed.
[2]memory usage is one byte per pixel