Wt::Dbo is a C++ ORM (Object-Relational-Mapping) library.
The library is distributed as part of Wt
for building database-driven web applications, but may be equally well
used independently from it.
The library provides a class-based view on database tables which keeps
an object-hiearchy of database object automatically synchronized with
a database by inserting, updating and deleting database
records. C++ classes map to database tables, class fields to
table columns, and pointers and collections of pointers to database
relations. An object from a mapped class is called a database object
(dbo). Query results may be defined in terms of database objects,
primitives, or tuples of these.
A modern C++ approach is used to solve the mapping problem. Rather
than resorting to XML-based descriptions of how C++ classes and
fields should map onto tables and columns, or using obscure macros,
the mapping is defined entirely in C++ code.
In this tutorial, we will work our way through a blogging example,
similar to the one that is distributed with the library.
Tip
|
The complete source code for the examples used in this tutorial are
available as ready-to-run programs in the examples/feature/dbo/ folder of
Wt.
|
We will start off with using Wt::Dbo for mapping a single class User
to a corresponding table user.
Warning
|
In this tutorial and the examples, we alias the namespace Wt::Dbo to
dbo, and in our explanation we will refer to types and methods
available in that namespace directly.
|
To build the following example, you need to link against the wtdbo
and wtdbosqlite3 libraries.
Mapping a single class (tutorial1.C)
#include <Wt/Dbo/Dbo>
#include <string>
namespace dbo = Wt::Dbo;
class User {
public:
enum Role {
Visitor = 0,
Admin = 1,
Alien = 42
};
std::string name;
std::string password;
Role role;
int karma;
template<class Action>
void persist(Action& a)
{
dbo::field(a, name, "name");
dbo::field(a, password, "password");
dbo::field(a, role, "role");
dbo::field(a, karma, "karma");
}
};
This example shows how persistence support is defined for a C++
class. A template member method persist() is defined which serves as
a persistence definition for the class. For each member in the class,
a call to
Wt::Dbo::field()
is used to map the field to a table column name.
As you may see, standard C++ types such as int, std::string and
enum types are readily supported by the library (a full list of supported
types can be found in the documentation of Wt::Dbo::sql_value_traits<T>). Support for other
types can be added by specializing
Wt::Dbo::sql_value_traits<T>.
There is also support for built-in Wt types such as WDate, WDateTime, WTime and
WString which can be enabled by including <Wt/Dbo/WtSqlTraits>.
The library defines a number of actions which will be applied to a
database object using its persist() method, which applies it in turn
to all its members. These actions will then read, update or insert
database objects, create the schema, or propagate transaction
outcomes.
Note
|
For brevity, our example uses public members. There is nothing that
prevents you to encapsulate your state in private members and provide
accessor methods. You may even define the persistence method in terms of
accessor methods by differentiating between read and write actions.
|
Now that we have a mapping definition for our User class, we can
start a database session, create our schema (if necessary) and add a
user to the database.
Let us walk through the code for doing this.
(tutorial1.C continued)
void run()
{
/*
* Setup a session, would typically be done once at application startup.
*/
dbo::backend::Sqlite3 sqlite3("blog.db");
dbo::Session session;
session.setConnection(sqlite3);
...
The Session object is a long
living object that provides access to our database objects. You will
typically create a Session object for the entire lifetime of an
application session, and one per user. None of the
Wt::Dbo classes are thread safe (except for
the connection pools), and session objects are not shared between
sessions.
The lack of thread-safety is not simply a consequence of laziness on
our part. It coincides with the promises made by transactional
integrity on the database: you will not want to see the changes made
by one session in another session while its transaction has not been
committed (Read-Committed transaction isolation level). It might make
sense however to implement a copy-on-write strategy in the future, to
allow sharing of the bulk of database objects between sessions.
The session is given a connection which it may use to communicate with
the database. A session will use a connection only during a
transaction, and thus does not really need a dedicated
connection. When you are planning for multiple concurrent sessions, it
makes sense to use a connection pool instead, and a session may also
be initialized with a reference to a
connection
pool.
Wt::Dbo uses an abstraction layer for database access, and currently
supports Postgres
and Sqlite3 as
backends.
(tutorial1.C continued)
...
session.mapClass<User>("user");
/*
* Try to create the schema (will fail if already exists).
*/
session.createTables();
...
Next, we use
mapClass()
to register each database class with the session, indicating the
database table onto which the class must be mapped.
Certainly during development, but also for initial deployment, it is
convenient to let Wt::Dbo create or drop the database schema.
This generates the following SQL:
begin transaction
create table "user" (
"id" integer primary key autoincrement,
"version" integer not null,
"name" text not null,
"password" text not null,
"role" integer not null,
"karma" integer not null
)
commit transaction
As you can see, next to the four columns that map to C++ fields, by
default, Wt::Dbo adds another two columns: id and version. The
id is a surrogate primary key, and version is used for version-based
optimistic locking. Since Wt 3.1.4, Wt::Dbo you can suppress the
version field, and provide natural keys of any type instead of the
surrogate primary key, see Customizing the mapping.
Finally, we can add a user to the database. All database operations
happen within a transaction.
(tutorial1.C continued)
...
/*
* A unit of work happens always within a transaction.
*/
dbo::Transaction transaction(session);
User *user = new User();
user->name = "Joe";
user->password = "Secret";
user->role = User::Visitor;
user->karma = 13;
dbo::ptr<User> userPtr = session.add(user);
transaction.commit();
}
A call to
Session::add()
adds an object to the database. This call returns a
ptr<Dbo> to reference a database
object of type Dbo. This is a shared pointer which also keeps
track of the persistence state of the referenced object. Within each
session, a database object will be loaded at most once: the session
keeps track of loaded database objects and returns an existing object
whenever a query to the database requires this. When the last pointer
to a database object goes out of scope, the transient (in-memory)
copy of the database object is also deleted (unless it was modified,
in which case the transient copy will only be be deleted after changes
have been successfully committed to the database).
The session also keeps track of objects that have been modified and
which need to be flushed (using SQL statements) to the
database. Flushing happens automatically when committing the
transaction, or whenever needed to maintain consistency between the
transient objects and the database copy (e.g. before doing a query).
This generates the following SQL:
begin transaction
insert into "user" ("version", "name", "password", "role", "karma") values (?, ?, ?, ?, ?)
commit transaction
All SQL statements are prepared once (per connection) and reused
later, which has the benefit of avoiding SQL injection problems, and
allows potentially better performance.
(tutorial1.C continued)
dbo::ptr<User> joe = session.find<User>().where("name = ?").bind("Joe");
std::cerr << "Joe has karma: " << joe->karma << std::endl;
All queries use prepared statements with positional argument
binding. The Session::find<T>() method returns a
Query< ptr<T> > object. The Query object
can be used to refine the search by defining Sql where, order by
and group by definitions, and allows binding of parameters using
Query::bind(). In
this case the query should expect a single result and is casted
directly to a database object pointer.
Note
|
Since Wt 3.1.3, the Query class has a second parameter
BindStrategy which has two possible values, corresponding two
different query implementations.
The default strategy is DynamicBinding and allows the query to be a
long lived object associated with the session which may be run
multiple times. It also allows you to modify the query by changing
only the order order the limit/offsets.
An alternative strategy is DirectBinding which passes bound
parameters directly on to an underlying prepared statement. This
corresponds to the old behavior of a Query object. Such a query can be
run only once, but has the benefit of having less (C++) overhead
because the parameters values are directly passed on to the backend
instead of stored within the query object.
|
The query formulated to the database is:
select id, version, "name", "password", "role", "karma"
from "user"
where (name = ?)
The more general way for querying uses
Session::query<Result>(sql),
which supports not only database objects as results. The query of
above is equivalent to:
(tutorial1.C continued)
dbo::ptr<User> joe2 = session.query< dbo::ptr<User> >("select u from user u").where("name = ?").bind("Joe");
And this generates similar SQL:
select u.id, u.version, u."name", u."password", u."role", u."karma"
from user u
where (name = ?)
The sql statement passed to the method may be arbitrary sql which
returns results that are compatible with the Result type. The
select part of the SQL query may be rewritten (as in the example
above) to return the individual fields of a queried database object.
To illustrate that Session::query<Result>() may be used to return
other types, consider the query below where an int result is
returned.
(tutorial1.C continued)
int count = session.query<int>("select count(1) from user").where("name = ?").bind("Joe");
The queries above were expecting unique results, but queries can also
have multiple results. A Session::query<Result>() may therefore in
general return a dbo::collection< Result > (for multiple results)
and in the examples above they were coerced to a single unique
Result for convenience. Similarly, Session::find<Dbo>() may
return a collection< ptr<Dbo> > or a unique ptr<_Dbo>. If a
unique result is asked, but the query found multiple results, a
NoUniqueResultException
will be thrown.
collection<T> is an
STL-compatible collection which has iterators that implement the
InputIterator requirements. Thus, you can only iterate through the
results of a collection once. After the results have been iterated the
collection can no longer be used (but the Query object can be
reused unless a DirectBinding bind strategy was used).
The following code shows how you may multiple results of a query may
be iterated:
(tutorial1.C continued)
typedef dbo::collection< dbo::ptr<User> > Users;
Users users = session.find<User>();
std::cerr << "We have " << users.size() << " users:" << std::endl;
for (Users::const_iterator i = users.begin(); i != users.end(); ++i)
std::cerr << " user " << (*i)->name
<< " with karma of " << (*i)->karma << std::endl;
This code will perform two database queries: one for the call to
collection::size() and one for iterating the results:
select count(1) from "user"
select id, version, "name", "password", "role", "karma" from "user"
Warning
|
A query uses a prepared statement to execute, and prepares a new
statement if no statement was yet prepared for that query. Because a
prepared statement is usually not reentrant and at the same time a
query will use an existing statement if one exists, you need to be
careful to not have two collections with the same statement busy at
the same time. Thus while iterating the results of a query you cannot
use that same query again. Therefore it may be necessary to copy the
results into a standard container (such as std::vector) before
iterating them. Since Wt 3.1.3, concurrent use will be detected and
an exception will be thrown saying:
A collection for '...' is already in use. Reentrant statement use is not yet implemented.
We plan to remove this restriction in later versions, cloning
prepared statements as necessary.
|
Unlike most other smart pointers, ptr<Dbo> is read-only by
default: it returns a const Dbo*. To modify a database object, you
need to call the ptr::modify() method, which returns a non-const
object. This mark the object as dirty and the modifications will later
be synchronized to the database.
(tutorial1.C continued)
dbo::ptr<User> joe = session.find<User>().where("name = ?").bind("Joe");
joe.modify()->karma++;
joe.modify()->password = "public";
Database synchronization does not happen instantaneously, instead,
they are delayed until explicitly asked, using
ptr<Dbo>::flush()
or
Session::flush(),
until a query is executed whose results may be affected by the changes
made, or until the transaction is committed.
The previous code will generate the following SQL:
select id, version, "name", "password", "role", "karma"
from "user"
where (name = ?)
update "user" set "version" = ?, "name" = ?, "password" = ?, "role" = ?, "karma" = ?
where "id" = ? and "version" = ?
We already saw how using Session::add(ptr<Dbo>), we added a new
object to the database. The opposite operation is
ptr<Dbo>::remove(): it deletes the object in the database.
(tutorial1.C continued)
dbo::ptr<User> joe = session.find<User>().where("name = ?").bind("Joe");
joe.remove();
After removing an object, the transient object can still be used, and
can even be re-added to the database.
Note
|
Like modify(), also the add() and remove() operations defer
synchronization with the database, and therefore the following code
does not actually have any effect on the database:
(tutorial1.C continued)
dbo::ptr<User> silly = session.add(new User());
silly.modify()->name = "Silly";
silly.remove();
|
6.1. Many-to-One relations
Let’s add posts to our blogging example, and define a Man-to-One
relation between posts and users. In the code below, we limit
ourselves to the statements important for defining the relationship.
Many-to-One relation (tutorial2.C)
#include <Wt/Dbo/Dbo>
#include <string>
namespace dbo = Wt::Dbo;
class User;
class Post {
public:
...
dbo::ptr<User> user;
template<class Action>
void persist(Action& a)
{
...
dbo::belongsTo(a, user, "user");
}
};
class User {
public:
...
dbo::collection< dbo::ptr<Post> > posts;
template<class Action>
void persist(Action& a)
{
...
dbo::hasMany(a, posts, dbo::ManyToOne, "user");
}
};
At the Many-side, we add a reference to a user, and in the
persist() method we call
belongsTo(). This
allows us to reference the user to which this post belongs. The last
argument will correspond to the name of the database column which
defines the relationship.
At the One-side, we add a collection of posts, and in the
persist() method we call
hasMany(). The
join field must be the same name as in reciproce belongsTo() method
call.
If we add the Post class too to our session using
Session::mapClass(), and create the schema, the following SQL is
generated:
create table "user" (
...
-- table user is unaffected by the relationship
);
create table "post" (
...
"user_id" bigint,
constraint "fk_post_user" foreign key ("user_id") references "user" ("id")
)
Note the user_id field which corresponds to the join name “user”.
At the Many-side, you may read or write the ptr to set a user to
which this post belongs.
The collection at the One-side allows us to retrieve all associated
elements, and also insert() and remove() elements, which has the same
effect as setting the ptr on the Many-side.
(tutorial2.C continued)
dbo::ptr<Post> post = session.add(new Post());
post.modify()->user = joe; // or joe.modify()->posts.insert(post);
// will print 'Joe has 1 post(s).'
std::cerr << "Joe has " << joe->posts.size() << " post(s)." << std::endl;
As you can see, as soon as joe is set as user for the new post, the
post is reflected in the posts collection of joe, and vice-versa.
Warning
|
The collection uses a prepared statement to execute. Collections will
try to share a single prepared statement, but prepared statements are
usually not reentrant. As a result, you need to be careful to not have
two collections with the same statement busy at the same time. Thus
while iterating a collection, you need to be sure you will not
reentrantly iterate the same collection (of the same or another
object). Therefore it may be necessary to copy the results into a
standard container (such as std::vector) before iterating them.
We plan to remove this restriction in later versions, cloning
prepared statements as necessary.
|
6.2. Many-to-Many relations
To illustrate Many-to-Many relations, we will add tags to our
blogging example, and define an Many-to-Many relation between posts
and tags. In the code below, we again limit ourselves to the
statements important for defining the relationship.
Many-to-Many relation (tutorial2.C)
#include <Wt/Dbo/Dbo>
#include <string>
namespace dbo = Wt::Dbo;
class Tag;
class Post {
public:
...
dbo::collection< dbo::ptr<Tag> > tags;
template<class Action>
void persist(Action& a)
{
...
dbo::hasMany(a, tags, dbo::ManyToMany, "post_tags");
}
};
class Tag {
public:
...
dbo::collection< dbo::ptr<Post> > posts;
template<class Action>
void persist(Action& a)
{
...
dbo::hasMany(a, posts, dbo::ManyToMany, "post_tags");
}
};
As expected, the relationship is reflected in almost the same way in
both classes: they both have a collection of database objects of the
related class, and in the persist() method we call hasMany(). The
join field in this case will correspond to the name of a join-table
used to persist the relation.
Adding the Post class to our session using Session::mapClass(), we
now get the following SQL for creating the schema:
create table "post" (
...
-- table post is unaffected by the relationship
)
create table "tag" (
...
-- table tag is unaffected by the relationship
)
create table "post_tags" (
"post_id" bigint not null,
"tag_id" bigint not null,
primary key ("post_id", "tag_id"),
constraint "fk_post_tags_key1" foreign key ("post_id") references "post" ("id"),
constraint "fk_post_tags_key2" foreign key ("tag_id") references "tag" ("id")
)
create index "post_tags_post" on "post_tags" ("post_id")
create index "post_tags_tag" on "post_tags" ("tag_id")
The collection at either side of the Many-to-Many relation allows us
to retrieve all associated elements. Unlike a collection in a
Many-to-One relation however, we may now also
insert()
and
erase()
items from the collection. To define a relation between a post and a
tag, you need to add the post to the tag’s posts collection, or the
tag to the post’s tags collection. You may not do both! The change
will automatically be reflected in the reciproce collection. Likewise,
to undo the relation between a post and a tag, you should remove the
tag from the post’s tags collection, or the post from the tag’s
posts collection, but not both.
(tutorial2.C continued)
dbo::ptr<Post> post = ...
dbo::ptr<Tag> cooking = session.add(new Tag());
cooking.modify()->name = "Cooking";
post.modify()->tags.insert(cooking);
// will print '1 post(s) tagged with Cooking.'
std::cerr << cooking->posts.size() << " post(s) tagged with Cooking." << std::endl;
Warning
|
The same warning as above applies here as well.
|
6.3. One-to-One relations
One-to-One relations are currently not supported, but can be
simulated using Many-to-One relations as they have the same database
schema structure.
By default, Wt::Dbo will add an auto-incrementing surrogate primary
(id) key and a version field (version) to each mapped table.
While these defaults make sense for a new project, you can tailor the
mapping so that you can map to virtually any existing database schema.
7.1. Changing or disabling the surrogate primary key "id" field
To change the field name used for the surrogate primary key for a
mapped class, or to disable the surrogate primary key for a class and use a nautral key instead, you need to specialize
Wt::Dbo::dbo_traits<C>.
For example, the code below changes the primary key field for class
Post from id to post_id:
Changing the "id" field name (tutorial3.C)
#include <Wt/Dbo/Dbo>
namespace dbo = Wt::Dbo;
class Post {
public:
...
};
namespace Wt {
namespace Dbo {
template<>
struct dbo_traits<Post> : public dbo_default_traits {
static const char *surrogateIdField() {
return "post_id";
}
};
}
}
7.2. Changing or disabling the "version" field
To change the field name used for the optimistic concurrency control
version field (version), or to disable optimistic concurrency
control for a class alltoghether, you need to specialize
Wt::Dbo::dbo_traits<C>.
For example, the code below disables optimistic concurrency control
for class Post:
Disabling the "version" field name (tutorial4.C)
#include <Wt/Dbo/Dbo>
namespace dbo = Wt::Dbo;
class Post {
public:
...
};
namespace Wt {
namespace Dbo {
template<>
struct dbo_traits<Post> : public dbo_default_traits {
static const char *versionField() {
return 0;
}
};
}
}
7.3. Specifying a natural primary key
Instead of using a auto-incrementing surrogate primary key, you may
want to use a different primary key.
For example, the code below changes the primary key for the User table
to a string (his username) which maps onto a varchar (20) field
user_name:
Using a natural key (tutorial5.C)
#include <Wt/Dbo/Dbo>
namespace dbo = Wt::Dbo;
class User {
public:
std::string userId;
template<class Action>
void persist(Action& a)
{
dbo::id(a, userId, "user_id", 20);
}
};
namespace Wt {
namespace Dbo {
template<>
struct dbo_traits<User> : public dbo_default_traits {
typedef std::string IdType;
static IdType invalidId() {
return std::string();
}
static const char *surrogateIdField() { return 0; }
};
}
}
The id() function has the same syntax as the field() function.
A natural primary key may also be a composite key, a foreign key or
a combination.
7.4. Specifying a composite natural primary key
To use a composite type as a natural primary key, i.e. a type which
consists of more than one field, you need to have a corresponding C++
type.
The type has a number of basic requirements, such as default
constructor, comparison operators (== and <), and a streaming operator.
Using a composite natural primary key (tutorial6.C)
struct Coordinate {
int x, y;
Coordinate()
: x(-1), y(-1) { }
Coordinate(int an_x, int an_y)
: x(an_x), y(an_y) { }
bool operator== (const Coordinate& other) const {
return x == other.x && y == other.y;
}
bool operator< (const Coordinate& other) const {
if (x < other.x)
return true;
else if (x == other.x)
return y < other.y;
else
return false;
}
};
std::ostream& operator<< (std::ostream& o, const Coordinate& c)
{
return o << "(" << c.x << ", " << c.y << ")";
}
Next, you must indicate how the type is persisted, by overloading Dbo’s
field() function for it.
(tutorial6.C continued)
namespace Wt {
namespace Dbo {
template <class Action>
void field(Action& action, Coordinate& coordinate, const std::string& name,
int size = -1)
{
field(action, coordinate.x, name + "_x");
field(action, coordinate.y, name + "_y");
}
}
}
With this in place, we can use the Coordinate type as a natural primary
key type:
(tutorial6.C continued)
class GeoTag;
namespace Wt {
namespace Dbo {
template<>
struct dbo_traits<GeoTag> : public dbo_default_traits
{
typedef Coordinate IdType;
static IdType invalidId() { return Coordinate(); }
static const char *surrogateIdField() { return 0; }
};
}
}
class GeoTag {
public:
Coordinate position;
std::string name;
template <class Action>
void persist(Action& a)
{
dbo::id(a, position, "position");
dbo::field(a, name, "name");
}
};
Note that the composite key may also include foreign keys, by storing
ptr<> objects in the composite, which you map using a belongsTo()
declaration. See tutorial8.C for a complete example.
7.5. Specifying foreign key constraints
The belongsTo() function is overloaded so that you can add foreign
key constraints which are enforced by the database, such as:
-
NotNull: cannot be null
-
OnUpdateCascade: cascade an update of the (natural) primary key to the
foreign keys that reference it
-
OnUpdateSetNull: an update of the (natural) primary key sets referencing
foreign keys to null
-
OnDeleteCascade: cascade a delete of the object to also delete objects
that reference it using a foreign key
-
OnDeleteSetNull: when the object is deleted, set the referencing foreign
keys to null.
In the next chapter we will see how you can specify these foreign key
constraints also for foreign keys that double as primary key.
7.6. Specifying a natural primary key that is also a foreign key
Let’s define a class UserInfo which provides additional data for a
User. We will only allow exactly one UserInfo object per User, and
therefore chose as primary key for the UserInfo a reference to the
User.
Using a foreign key as primary key (tutorial7.C)
#include <Wt/Dbo/Dbo>
#include <Wt/Dbo/backend/Sqlite3>
namespace dbo = Wt::Dbo;
class UserInfo;
class User;
namespace Wt {
namespace Dbo {
template<>
struct dbo_traits<UserInfo> : public dbo_default_traits {
typedef ptr<User> IdType;
static IdType invalidId() {
return ptr<User>();
}
static const char *surrogateIdField() { return 0; }
};
}
}
class User {
public:
std::string name;
dbo::collection< dbo::ptr<UserInfo> > infos;
template<class Action>
void persist(Action& a)
{
dbo::field(a, name, "name");
// In fact, this is really constrained to hasOne() ...
dbo::hasMany(a, infos, dbo::ManyToOne, "user");
}
};
class UserInfo {
public:
dbo::ptr<User> user;
std::string info;
template<class Action>
void persist(Action& a)
{
dbo::id(a, user, "user", dbo::OnDeleteCascade);
dbo::field(a, info, "info");
}
};
void run()
{
/*
* Setup a session, would typically be done once at application startup.
*/
dbo::backend::Sqlite3 sqlite3(":memory:");
sqlite3.setProperty("show-queries", "true");
dbo::Session session;
session.setConnection(sqlite3);
session.mapClass<User>("user");
session.mapClass<UserInfo>("user_info");
/*
* Try to create the schema (will fail if already exists).
*/
session.createTables();
dbo::Transaction transaction(session);
{
User *user = new User();
user->name = "Joe";
dbo::ptr<User> userPtr = session.add(user);
UserInfo *userInfo = new UserInfo();
userInfo->user = userPtr;
userInfo->info = "great guy";
session.add(userInfo);
transaction.commit();
}
{
dbo::Transaction transaction(session);
dbo::ptr<UserInfo> info = session.find<UserInfo>();
std::cerr << info->user->name << " is a " << info->info << std::endl;
transaction.commit();
}
}
int main(int argc, char **argv)
{
run();
}
As you can see, in this example we would really need a One-to-One
relationship, but this currently not yet supported in Dbo and thus we
emulate it using a Many-to-One relationship (which has the same
representation in SQL).
When run, this should output:
begin transaction
create table "user" (
"id" integer primary key autoincrement,
"version" integer not null,
"name" text not null
)
create table "user_info" (
"version" integer not null,
"user_id" bigint,
"info" text not null,
primary key ("user_id"),
constraint "fk_user_info_user" foreign key ("user_id") references "user" ("id") on delete cascade
)
commit transaction
begin transaction
insert into "user" ("version", "name") values (?, ?)
insert into "user_info" ("version", "user_id", "info") values (?, ?, ?)
commit transaction
begin transaction
select version, "user_id", "info" from "user_info"
select "version", "name" from "user" where "id" = ?
Joe is a great guy
commit transaction
Reading data from the database or flushing changes to the database
require an active transaction. A
Transaction is a RIIA
(Resource-Initialization-is-Acquisition) class which at the same time
provides isolation between concurrent sessions and atomicity for
persisting changes to the database.
The library implements optimistic locking, which allows detection
(rather than avoidance) of concurrent modifications. It is a
recommended and widely used strategy for dealing with concurrency
issues in a scalable manner as no write locks are needed on the
database. To detect a concurrent modification, a version field is
added to each table which is incremented on each modification. When
performing a modification (such as updating or removing an object), it
is checked that the version of the record in the database is the same
as the version of the object that was originally read from the
database.
Note
|
Transaction isolation levels The minimum level of isolation which is required for the library’s
optimistic locking strategy is Read Committed: modifications in a
transaction are only visible to other sessions as soon as they are
committed. This is usually the lowest level of isolation supported by
a database. |
The Transaction class is a light-weight proxy that references a
logical transaction: multiple (usually nested) Transaction objects
may be instantiated simultaneously, which each need to be committed
for the logical transaction to be committed. In this way you can
easily protect individual methods which require database access with
such a transaction object, which will automatically participate in a
wider transaction if that is available. A transaction will in fact
defer opening a real transaction in the database until needed, and
thus there is no penalty for instantiating a transaction to make sure
a unit of work is atomic, even if you are not yet sure that there will
be actual work done.
Transactions may fail and dealing with failing transactions is an
integral aspect of their usage. When the library detects a concurrent
modification, a
StaleObjectException
is thrown. Other exceptions may be thrown, including exceptions in the
backend driver when for example the database schema is not compatible
with the mapping. There may also be problems detected by the business
logic which may raise an exception and cause the transaction to be
rolled back. When a transaction is rolled back, the modified database
objects are not successfully synchronized with the database, but may
possibly be synchronized later in a new transaction.
Obviously, many exceptions will be fatal. One notable exception is the
StaleObjectException however. Different strategies are possible to
deal with this exception. Regardless of the approach, you will at
least need to
reread()
the stale database object(s) before being able to commit changes made
in a new transaction.
Wt::Dbo is included in Wt, and can thus be installed as part of this
library for which there may be standard packages availabe for your
operating system.
The library does however in no way depend on Wt, and can also be
built, installed and used separately from it. Starting from a Wt
source package (and on in a UNIX-like environment), you would do the
following to build and install only Wt::Dbo:
Installing Wt::Dbo from source (UNIX-like)
$ cd wt-xxx
$ mkdir build
$ cd build
$ cmake ../ # extra options may be needed for locating boost, postgres, install directory, ...
$ cd src/Wt/Dbo
$ make
$ sudo make install