Chapter 6: A camada de abstração de banco de dados - TODO

The database abstraction layer



web2py comes with a Database Abstraction Layer (DAL), an API that maps Python objects into database objects such as queries, tables, and records. The DAL dynamically generates the SQL in real time using the specified dialect for the database back end, so that you do not have to write SQL code or learn different SQL dialects (the term SQL is used generically), and the application will be portable among different types of databases. At the time of this writing, the supported databases are SQLite (which comes with Python and thus web2py), PostgreSQL, MySQL, Oracle, MSSQL, FireBird, DB2, Informix, Ingres, MongoDB, and the Google App Engine (SQL and NoSQL). Experimentally we support more databases. Please check on the web2py web site and mailing list for more recent adapters. Google NoSQL is treated as a particular case in Chapter 13.

The Windows binary distribution works out of the box with SQLite and MySQL. The Mac binary distribution works out of the box with SQLite. To use any other database back-end, run from the source distribution and install the appropriate driver for the required back end.

database drivers

Once the proper driver is installed, start web2py from source, and it will find the driver. Here is a list of drivers:


databasedrivers (source)
SQLitesqlite3 or pysqlite2 or zxJDBC [zxjdbc] (on Jython)
PostgreSQLpsycopg2 [psycopg2] or pg8000 [pg8000] or zxJDBC [zxjdbc] (on Jython)
MySQLpymysql [pymysql] or MySQLdb [mysqldb]
Oraclecx_Oracle [cxoracle]
MSSQLpyodbc [pyodbc]
FireBirdkinterbasdb [kinterbasdb] or fdb or pyodbc
DB2pyodbc [pyodbc]
Informixinformixdb [informixdb]
Ingresingresdbi [ingresdbi]
Cubridcubriddb [cubridb] [cubridb]
SybaseSybase [Sybase]
Teradatapyodbc [Teradata]
MongoDBpymongo [pymongo]
IMAPimaplib [IMAP]

sqlite3, pymysql, pg8000, and imaplib ship with web2py. Support of MongoDB is experimental. The IMAP option allows to use DAL to access IMAP.

web2py defines the following classes that make up the DAL:

DAL represents a database connection. For example:

db = DAL('sqlite://storage.db')

Table represents a database table. You do not directly instantiate Table; instead, DAL.define_table instantiates it.

db.define_table('mytable', Field('myfield'))

The most important methods of a Table are:


.insert, .truncate, .drop, and .import_from_csv_file.


Field represents a database field. It can be instantiated and passed as an argument to DAL.define_table.


DAL Rows

is the object returned by a database select. It can be thought of as a list of Row rows:

rows = db(db.mytable.myfield!=None).select()

Row contains field values.

for row in rows:
    print row.myfield

Query is an object that represents a SQL "where" clause:

myquery = (db.mytable.myfield != None) | (db.mytable.myfield > 'A')

Set is an object that represents a set of records. Its most important methods are count, select, update, and delete. For example:

myset = db(myquery)
rows =

Expression is something like an orderby or groupby expression. The Field class is derived from the Expression. Here is an example.

myorder = db.mytable.myfield.upper() |
db().select(db.table.ALL, orderby=myorder)

Connection strings

connection strings

A connection with the database is established by creating an instance of the DAL object:

>>> db = DAL('sqlite://storage.db', pool_size=0)

db is not a keyword; it is a local variable that stores the connection object DAL. You are free to give it a different name. The constructor of DAL requires a single argument, the connection string. The connection string is the only web2py code that depends on a specific back-end database. Here are examples of connection strings for specific types of supported back-end databases (in all cases, we assume the database is running from localhost on its default port and is named "test"):


Notice that in SQLite the database consists of a single file. If it does not exist, it is created. This file is locked every time it is accessed. In the case of MySQL, PostgreSQL, MSSQL, FireBird, Oracle, DB2, Ingres and Informix the database "test" must be created outside web2py. Once the connection is established, web2py will create, alter, and drop tables appropriately.

It is also possible to set the connection string to None. In this case DAL will not connect to any back-end database, but the API can still be accessed for testing. Examples of this will be discussed in Chapter 7.

Some times you may need to generate SQL as if you had a connection but without actually connecting to the database. This can be done with

db = DAL('...', do_connect=False)

In this case you will be able to call _select, _insert, _update, and _delete to generate SQL but not call select, insert, update, and delete. In most of the cases you can use do_connect=False even without having the required database drivers.

Notice that by default web2py uses utf8 character encoding for databases. If you work with existing databases that behave differently, you have to change it with the optional parameter db_codec like

db = DAL('...', db_codec='latin1')

otherwise you'll get UnicodeDecodeErrors tickets.

Connection pooling

connection pooling

The second argument of the DAL constructor is the pool_size; it defaults to zero.

As it is rather slow to establish a new database connection for each request, web2py implements a mechanism for connection pooling. Once a connection is established and the page has been served and the transaction completed, the connection is not closed but goes into a pool. When the next http request arrives, web2py tries to obtain a connection from the pool and use that for the new transaction. If there are no available connections in the pool, a new connection is established.

The pool_size parameter is ignored by SQLite and Google App Engine.

Connections in the pools are shared sequentially among threads, in the sense that they may be used by two different but not simultaneous threads. There is only one pool for each web2py process.

When web2py starts, the pool is always empty. The pool grows up to the minimum between the value of pool_size and the max number of concurrent requests. This means that if pool_size=10 but our server never receives more than 5 concurrent requests, then the actual pool size will only grow to 5. If pool_size=0 then connection pooling is not used.

Connection pooling is ignored for SQLite, since it would not yield any benefit.

Connection failures

If web2py fails to connect to the database it waits 1 seconds and tries again up to 5 times before declaring a failure. In case of connection pooling it is possible that a pooled connection that stays open but unused for some time is closed by the database end. Thanks to the retry feature web2py tries to re-establish these dropped connections.

When using connection pooling a connection is used, put back in the pool and then recycled. It is possible that while the connection is idle in pool the connection is closed by the database server. This can be because of a malfunction or a timeout. When this happens web2py detects it and re-establish the connection.

Replicated databases

The first argument of DAL(...) can be a list of URIs. In this case web2py tries to connect to each of them. The main purpose for this is to deal with multiple database servers and distribute the workload among them). Here is a typical use case:

db = DAL(['mysql://...1','mysql://...2','mysql://...3'])

In this case the DAL tries to connect to the first and, on failure, it will try the second and the third. This can also be used to distribute load in a database master-slave configuration. We will talk more about this in Chapter 13 in the context of scalability.

Reserved keywords

reserved Keywords

There is also another argument that can be passed to the DAL constructor to check table names and column names against reserved SQL keywords in target back-end databases.

This argument is check_reserved and it defaults to None.

This is a list of strings that contain the database back-end adapter names.

The adapter name is the same as used in the DAL connection string. So if you want to check against PostgreSQL and MSSQL then your connection string would look as follows:

db = DAL('sqlite://storage.db',
         check_reserved=['postgres', 'mssql'])

The DAL will scan the keywords in the same order as of the list.

There are two extra options "all" and "common". If you specify all, it will check against all known SQL keywords. If you specify common, it will only check against common SQL keywords such as SELECT, INSERT, UPDATE, etc.

For supported back-ends you may also specify if you would like to check against the non-reserved SQL keywords as well. In this case you would append _nonreserved to the name. For example:

check_reserved=['postgres', 'postgres_nonreserved']

The following database backends support reserved words checking.


DAL, Table, Field

The best way to understand the DAL API is to try each function yourself. This can be done interactively via the web2py shell, although ultimately, DAL code goes in the models and controllers.

Start by creating a connection. For the sake of example, you can use SQLite. Nothing in this discussion changes when you change the back-end engine.

>>> db = DAL('sqlite://storage.db')

The database is now connected and the connection is stored in the global variable db.

At any time you can retrieve the connection string.

>>> print db._uri

and the database name

>>> print db._dbname

The connection string is called a _uri because it is an instance of a Uniform Resource Identifier.

The DAL allows multiple connections with the same database or with different databases, even databases of different types. For now, we will assume the presence of a single database since this is the most common situation.


The most important method of a DAL is define_table:

>>> db.define_table('person', Field('name'))

It defines, stores and returns a Table object called "person" containing a field (column) "name". This object can also be accessed via db.person, so you do not need to catch the return value.

Do not declare a field called "id", because one is created by web2py anyway. Every table has a field called "id" by default. It is an auto-increment integer field (starting at 1) used for cross-reference and for making every record unique, so "id" is a primary key. (Note: the id's starting at 1 is back-end specific. For example, this does not apply to the Google App Engine NoSQL.)

named id field

Optionally you can define a field of type='id' and web2py will use this field as auto-increment id field. This is not recommended except when accessing legacy database tables. With some limitation, you can also use different primary keys and this is discussed in the section on "Legacy databases and keyed tables".

Tables can be defined only once but you can force web2py to redefine an existing table:

db.define_table('person', Field('name'))
db.define_table('person', Field('name'), redefine=True)

The redefinition may trigger a migration if field content is different.

Because usually in web2py models are executed before controllers, it is possible that some table are defined even if not needed. It is therefore necessary to speed up the code by making table definitions lazy. This is done by setting the DAL(...,lazy_tables=True) attributes. Tables will be actually created only when accessed.

Record representation

It is optional but recommended to specify a format representation for records:

>>> db.define_table('person', Field('name'), format='%(name)s')


>>> db.define_table('person', Field('name'), format='%(name)s %(id)s')

or even more complex ones using a function:

>>> db.define_table('person', Field('name'),
       format=lambda r: or 'anonymous')

The format attribute will be used for two purposes:

  • To represent referenced records in select/option drop-downs.
  • To set the db.othertable.person.represent attribute for all fields referencing this table. This means that SQLTABLE will not show references by id but will use the format preferred representation instead.
Field constructor

These are the default values of a Field constructor:

Field(name, 'string', length=None, default=None,
      required=False, requires='<default>',
      ondelete='CASCADE', notnull=False, unique=False,
      uploadfield=True, widget=None, label=None, comment=None,
      writable=True, readable=True, update=None, authorize=None,
      autodelete=False, represent=None, compute=None,

Not all of them are relevant for every field. "length" is relevant only for fields of type "string". "uploadfield" and "authorize" are relevant only for fields of type "upload". "ondelete" is relevant only for fields of type "reference" and "upload".

  • length sets the maximum length of a "string", "password" or "upload" field. If length is not specified a default value is used but the default value is not guaranteed to be backward compatible. To avoid unwanted migrations on upgrades, we recommend that you always specify the length for string, password and upload fields.
  • default sets the default value for the field. The default value is used when performing an insert if a value is not explicitly specified. It is also used to pre-populate forms built from the table using SQLFORM. Note, rather than being a fixed value, the default can instead be a function (including a lambda function) that returns a value of the appropriate type for the field. In that case, the function is called once for each record inserted, even when multiple records are inserted in a single transaction.
  • required tells the DAL that no insert should be allowed on this table if a value for this field is not explicitly specified.
  • requires is a validator or a list of validators. This is not used by the DAL, but it is used by SQLFORM. The default validators for the given types are shown in the following table:
field typedefault field validators
stringIS_LENGTH(length) default length is 512
integerIS_INT_IN_RANGE(-1e100, 1e100)
doubleIS_FLOAT_IN_RANGE(-1e100, 1e100)
decimal(n,m)IS_DECIMAL_IN_RANGE(-1e100, 1e100)
reference <table>IS_IN_DB(db,table.field,format)
list:reference <table>IS_IN_DB(db,table.field,format,multiple=True)

Decimal requires and returns values as Decimal objects, as defined in the Python decimal module. SQLite does not handle the decimal type so internally we treat it as a double. The (n,m) are the number of digits in total and the number of digits after the decimal point respectively.

The big-id and, big-reference are only supported by some of the database engines and are experimental. They are not normally used as field types unless for legacy tables, however, the DAL constructor has a bigint_id argument that when set to True makes the id fields and reference fields big-id and big-reference respectively.

The list: fields are special because they are designed to take advantage of certain denormalization features on NoSQL (in the case of Google App Engine NoSQL, the field types ListProperty and StringListProperty) and back-port them all the other supported relational databases. On relational databases lists are stored as a text field. The items are separated by a | and each | in string item is escaped as a ||. They are discussed in their own section.

The json field type is pretty much explanatory. It can store any json serializable object. It is designed to work specifically for MongoDB and backported to the other database adapters for portability.

Notice that requires=... is enforced at the level of forms, required=True is enforced at the level of the DAL (insert), while notnull, unique and ondelete are enforced at the level of the database. While they sometimes may seem redundant, it is important to maintain the distinction when programming with the DAL.

  • ondelete translates into the "ON DELETE" SQL statement. By default it is set to "CASCADE". This tells the database that when it deletes a record, it should also delete all records that refer to it. To disable this feature, set ondelete to "NO ACTION" or "SET NULL".
  • notnull=True translates into the "NOT NULL" SQL statement. It prevents the database from inserting null values for the field.
  • unique=True translates into the "UNIQUE" SQL statement and it makes sure that values of this field are unique within the table. It is enforced at the database level.
  • uploadfield applies only to fields of type "upload". A field of type "upload" stores the name of a file saved somewhere else, by default on the filesystem under the application "uploads/" folder. If uploadfield is set, then the file is stored in a blob field within the same table and the value of uploadfield is the name of the blob field. This will be discussed in more detail later in the context of SQLFORM.
  • uploadfolder defaults to the application's "uploads/" folder. If set to a different path, files will uploaded to a different folder. For example, uploadfolder=os.path.join(request.folder,'static/temp') will upload files to the web2py/applications/myapp/static/temp folder.
  • uploadseparate if set to True will upload files under different subfolders of the uploadfolder folder. This is optimized to avoid too many files under the same folder/subfolder. ATTENTION: You cannot change the value of uploadseparate from True to False without breaking the system. web2py either uses the separate subfolders or it does not. Changing the behavior after files have been uploaded will prevent web2py from being able to retrieve those files. If this happens it is possible to move files and fix the problem but this is not described here.
  • uploadfs allows you specify a different file system where to upload files, including an Amazon S3 storage or a remote FTP storage. This option requires PyFileSystem installed. uploadfs must point to PyFileSystem.
  • widget must be one of the available widget objects, including custom widgets, for example: SQLFORM.widgets.string.widget. A list of available widgets will be discussed later. Each field type has a default widget.
  • label is a string (or something that can be serialized to a string) that contains the label to be used for this field in auto-generated forms.
  • comment is a string (or something that can be serialized to a string) that contains a comment associated with this field, and will be displayed to the right of the input field in the autogenerated forms.
  • writable if a field is writable, it can be edited in autogenerated create and update forms.
  • readable if a field is readable, it will be visible in read-only forms. If a field is neither readable nor writable, it will not be displayed in create and update forms.
  • update contains the default value for this field when the record is updated.
  • compute is an optional function. If a record is inserted or updated, the compute function will be executed and the field will be populated with the function result. The record is passed to the compute function as a dict, and the dict will not include the current value of that, or any other compute field.
  • authorize can be used to require access control on the corresponding field, for "upload" fields only. It will be discussed more in detail in the context of Authentication and Authorization.
  • autodelete determines if the corresponding uploaded file should be deleted when the record referencing the file is deleted. For "upload" fields only.
  • represent can be None or can point to a function that takes a field value and returns an alternate representation for the field value. Examples: = lambda name,row: name.capitalize()
db.mytable.other_id.represent = lambda id,row: row.myfield
db.mytable.some_uploadfield.represent = lambda value,row:     A('get it', _href=URL('download', args=value))

"blob" fields are also special. By default, binary data is encoded in base64 before being stored into the actual database field, and it is decoded when extracted. This has the negative effect of using 25% more storage space than necessary in blob fields, but has two advantages. On average it reduces the amount of data communicated between web2py and the database server, and it makes the communication independent of back-end-specific escaping conventions.

Most attributes of fields and tables can be modified after they are defined:

db.person._format = '%(name)s/%(id)s' = 'anonymous'

(notice that attributes of tables are usually prefixed by an underscore to avoid conflict with possible field names).

You can list the tables that have been defined for a given database connection:

>>> print db.tables

You can also list the fields that have been defined for a given table:

>>> print db.person.fields
['id', 'name']

You can query for the type of a table:

>>> print type(db.person)
<class 'pydal.objects.Table'>

and you can access a table from the DAL connection using:

>>> print type(db['person'])
<class 'pydal.objects.Table'>

Similarly you can access fields from their name in multiple equivalent ways:

>>> print type(
<class 'pydal.objects.Field'>
>>> print type(db.person['name'])
<class 'pydal.objects.Field'>
>>> print type(db['person']['name'])
<class 'pydal.objects.Field'>

Given a field, you can access the attributes set in its definition:

>>> print
>>> print
>>> print
>>> print

including its parent table, tablename, and parent connection:

>>> == db.person
>>> == 'person'
>>> == db

A field also has methods. Some of them are used to build queries and we will see them later. A special method of the field object is validate and it calls the validators for the field.


which returns a tuple (value, error). error is None if the input passes validation.



define_table checks whether or not the corresponding table exists. If it does not, it generates the SQL to create it and executes the SQL. If the table does exist but differs from the one being defined, it generates the SQL to alter the table and executes it. If a field has changed type but not name, it will try to convert the data (If you do not want this, you need to redefine the table twice, the first time, letting web2py drop the field by removing it, and the second time adding the newly defined field so that web2py can create it.). If the table exists and matches the current definition, it will leave it alone. In all cases it will create the db.person object that represents the table.

We refer to this behavior as a "migration". web2py logs all migrations and migration attempts in the file "databases/sql.log".

The first argument of define_table is always the table name. The other unnamed arguments are the fields (Field). The function also takes an optional last argument called "migrate" which must be referred to explicitly by name as in:

>>> db.define_table('person', Field('name'), migrate='person.table')

The value of migrate is the filename (in the "databases" folder for the application) where web2py stores internal migration information for this table. These files are very important and should never be removed while the corresponding tables exist. In cases where a table has been dropped and the corresponding file still exist, it can be removed manually. By default, migrate is set to True. This causes web2py to generate the filename from a hash of the connection string. If migrate is set to False, the migration is not performed, and web2py assumes that the table exists in the datastore and it contains (at least) the fields listed in define_table. The best practice is to give an explicit name to the migrate table.

There may not be two tables in the same application with the same migrate filename.

The DAL class also takes a "migrate" argument, which determines the default value of migrate for calls to define_table. For example,

>>> db = DAL('sqlite://storage.db', migrate=False)

will set the default value of migrate to False whenever db.define_table is called without a migrate argument.

Notice that web2py only migrates new columns, removed columns, and changes in column type (not in sqlite). web2py does not migrate changes in attributes such as changes in the values of default, unique, notnull, and ondelete.

Migrations can be disabled for all tables at the moment of connection:

db = DAL(...,migrate_enabled=False)

This is the recommended behavior when two apps share the same database. Only one of the two apps should perform migrations, the other should disabled them.

Fixing broken migrations


There are two common problems with migrations and there are ways to recover from them.

One problem is specific with SQLite. SQLite does not enforce column types and cannot drop columns. This means that if you have a column of type string and you remove it, it is not really removed. If you add the column again with a different type (for example datetime) you end up with a datetime column that contains strings (junk for practical purposes). web2py does not complain about this because it does not know what is in the database, until it tries to retrieve records and fails.

If web2py returns an error in the gluon.sql.parse function when selecting records, this is the problem: corrupted data in a column because of the above issue.

The solution consists in updating all records of the table and updating the values in the column in question with None.

The other problem is more generic but typical with MySQL. MySQL does not allow more than one ALTER TABLE in a transaction. This means that web2py must break complex transactions into smaller ones (one ALTER TABLE at the time) and commit one piece at the time. It is therefore possible that part of a complex transaction gets committed and one part fails, leaving web2py in a corrupted state. Why would part of a transaction fail? Because, for example, it involves altering a table and converting a string column into a datetime column, web2py tries to convert the data, but the data cannot be converted. What happens to web2py? It gets confused about what exactly is the table structure actually stored in the database.

The solution consists of disabling migrations for all tables and enabling fake migrations:


This will rebuild web2py metadata about the table according to the table definition. Try multiple table definitions to see which one works (the one before the failed migration and the one after the failed migration). Once successful remove the fake_migrate=True attribute.

Before attempting to fix migration problems it is prudent to make a copy of "applications/yourapp/databases/*.table" files.

Migration problems can also be fixed for all tables at once:

db = DAL(...,fake_migrate_all=True)

Although if this fails, it will not help in narrowing down the problem.


Given a table, you can insert records

>>> db.person.insert(name="Alex")
>>> db.person.insert(name="Bob")

Insert returns the unique "id" value of each record inserted.

You can truncate the table, i.e., delete all records and reset the counter of the id.

>>> db.person.truncate()

Now, if you insert a record again, the counter starts again at 1 (this is back-end specific and does not apply to Google NoSQL):

>>> db.person.insert(name="Alex")

Notice you can pass parameters to truncate, for example you can tell SQLITE to restart the id counter.

db.person.truncate('RESTART IDENTITY CASCADE')

The argument is in raw SQL and therefore engine specific.


web2py also provides a bulk_insert method

>>> db.person.bulk_insert([{'name':'Alex'}, {'name':'John'}, {'name':'Tim'}])

It takes a list of dictionaries of fields to be inserted and performs multiple inserts at once. It returns the IDs of the inserted records. On the supported relational databases there is no advantage in using this function as opposed to looping and performing individual inserts but on Google App Engine NoSQL, there is a major speed advantage.

commit and rollback

No create, drop, insert, truncate, delete, or update operation is actually committed until you issue the commit command

>>> db.commit()

To check it let's insert a new record:

>>> db.person.insert(name="Bob")

and roll back, i.e., ignore all operations since the last commit:

>>> db.rollback()

If you now insert again, the counter will again be set to 2, since the previous insert was rolled back.

>>> db.person.insert(name="Bob")

Code in models, views and controllers is enclosed in web2py code that looks like this:

     execute models, controller function and view
     rollback all connections
     log the traceback
     send a ticket to the visitor
     commit all connections
     save cookies, sessions and return the page

There is no need to ever call commit or rollback explicitly in web2py unless one needs more granular control.


Timing queries

All queries are automatically timed by web2py. The variable db._timings is a list of tuples. Each tuple contains the raw SQL query as passed to the database driver and the time it took to execute in seconds. This variable can be displayed in views using the toolbar:



The DAL allows you to explicitly issue SQL statements.

>>> print db.executesql('SELECT * FROM person;')
[(1, u'Massimo'), (2, u'Massimo')]

In this case, the return values are not parsed or transformed by the DAL, and the format depends on the specific database driver. This usage with selects is normally not needed, but it is more common with indexes. executesql takes four optional arguments: placeholders, as_dict, fields and colnames. placeholders is an optional sequence of values to be substituted in or, if supported by the DB driver, a dictionary with keys matching named placeholders in your SQL.

If as_dict is set to True, and the results cursor returned by the DB driver will be converted to a sequence of dictionaries keyed with the db field names. Results returned with as_dict = True are the same as those returned when applying .as_list() to a normal select.

[{field1: value1, field2: value2}, {field1: value1b, field2: value2b}]

The fields argument is a list of DAL Field objects that match the fields returned from the DB. The Field objects should be part of one or more Table objects defined on the DAL object. The fields list can include one or more DAL Table objects in addition to or instead of including Field objects, or it can be just a single table (not in a list). In that case, the Field objects will be extracted from the table(s).

Instead of specifying the fields argument, the colnames argument can be specified as a list of field names in tablename.fieldname format. Again, these should represent tables and fields defined on the DAL object.

It is also possible to specify both fields and the associated colnames. In that case, fields can also include DAL Expression objects in addition to Field objects. For Field objects in "fields", the associated colnames must still be in tablename.fieldname format. For Expression objects in fields, the associated colnames can be any arbitrary labels.

Notice, the DAL Table objects referred to by fields or colnames can be dummy tables and do not have to represent any real tables in the database. Also, note that the fields and colnames must be in the same order as the fields in the results cursor returned from the DB.


Whether SQL was executed manually using executesql or was SQL generated by the DAL, you can always find the SQL code in db._lastsql. This is useful for debugging purposes:

>>> rows = db().select(db.person.ALL)
>>> print db._lastsql
SELECT, FROM person;

web2py never generates queries using the "*" operator. web2py is always explicit when selecting fields.


Finally, you can drop tables and all data will be lost:

>>> db.person.drop()


Currently the DAL API does not provide a command to create indexes on tables, but this can be done using the executesql command. This is because the existence of indexes can make migrations complex, and it is better to deal with them explicitly. Indexes may be needed for those fields that are used in recurrent queries.

Here is an example of how to create an index using SQL in SQLite:

>>> db = DAL('sqlite://storage.db')
>>> db.define_table('person', Field('name'))
>>> db.executesql('CREATE INDEX IF NOT EXISTS myidx ON person (name);')

Other database dialects have very similar syntaxes but may not support the optional "IF NOT EXISTS" directive.

Legacy databases and keyed tables

web2py can connect to legacy databases under some conditions.

The easiest way is when these conditions are met:

  • Each table must have a unique auto-increment integer field called "id"
  • Records must be referenced exclusively using the "id" field.

When accessing an existing table, i.e., a table not created by web2py in the current application, always set migrate=False.

If the legacy table has an auto-increment integer field but it is not called "id", web2py can still access it but the table definition must contain explicitly as Field('....','id') where ... is the name of the auto-increment integer field.

keyed table

Finally if the legacy table uses a primary key that is not an auto-increment id field it is possible to use a "keyed table", for example:

  • primarykey is a list of the field names that make up the primary key.
  • All primarykey fields have a NOT NULL set even if not specified.
  • Keyed tables can only reference other keyed tables.
  • Referencing fields must use the reference tablename.fieldname format.
  • The update_record function is not available for Rows of keyed tables.

Note that currently this is only available for DB2, MS-SQL, Ingres and Informix, but others can be easily added.

At the time of writing, we cannot guarantee that the primarykey attribute works with every existing legacy table and every supported database backend. For simplicity, we recommend, if possible, creating a database view that has an auto-increment id field.

Distributed transaction

distributed transactions

At the time of writing this feature is only supported by PostgreSQL, MySQL and Firebird, since they expose API for two-phase commits.

Assuming you have two (or more) connections to distinct PostgreSQL databases, for example:

db_a = DAL('postgres://...')
db_b = DAL('postgres://...')

In your models or controllers, you can commit them concurrently with:

DAL.distributed_transaction_commit(db_a, db_b)

On failure, this function rolls back and raises an Exception.

In controllers, when one action returns, if you have two distinct connections and you do not call the above function, web2py commits them separately. This means there is a possibility that one of the commits succeeds and one fails. The distributed transaction prevents this from happening.

More on uploads

Consider the following model:

>>> db.define_table('myfile',
    Field('image', 'upload', default='path/'))

In the case of an 'upload' field, the default value can optionally be set to a path (an absolute path or a path relative to the current app folder) and the default image will be set to a copy of the file at the path. A new copy is made for each new record that does not specify an image.

Normally an insert is handled automatically via a SQLFORM or a crud form (which is a SQLFORM) but occasionally you already have the file on the filesystem and want to upload it programmatically. This can be done in this way:

>>> stream = open(filename, 'rb')
>>> db.myfile.insert(, filename))

It is also possible to insert a file in a simpler way and have the insert method call store automatically:

>>> stream = open(filename, 'rb')
>>> db.myfile.insert(image=stream)

In this case the filename is obtained from the stream object if available.

The store method of the upload field object takes a file stream and a filename. It uses the filename to determine the extension (type) of the file, creates a new temp name for the file (according to web2py upload mechanism) and loads the file content in this new temp file (under the uploads folder unless specified otherwise). It returns the new temp name, which is then stored in the image field of the db.myfile table.

Note, if the file is to be stored in an associated blob field rather than the file system, the store() method will not insert the file in the blob field (because store() is called before the insert), so the file must be explicitly inserted into the blob field:

>>> db.define_table('myfile',
        Field('image', 'upload', uploadfield='image_file'),
        Field('image_file', 'blob'))
>>> stream = open(filename, 'rb')
>>> db.myfile.insert(, filename),

The opposite of .store is .retrieve:

>>> row = db(db.myfile).select().first()
>>> (filename, stream) = db.myfile.image.retrieve(row.image)
>>> import shutil
>>> shutil.copyfileobj(stream,open(filename,'wb'))

Query, Set, Rows

Let's consider again the table defined (and dropped) previously and insert three records:

>>> db.define_table('person', Field('name'))
>>> db.person.insert(name="Alex")
>>> db.person.insert(name="Bob")
>>> db.person.insert(name="Carl")

You can store the table in a variable. For example, with variable person, you could do:

>>> person = db.person

You can also store a field in a variable such as name. For example, you could also do:

>>> name =

You can even build a query (using operators like ==, !=, <, >, <=, >=, like, belongs) and store the query in a variable q such as in:

>>> q = name=='Alex'

When you call db with a query, you define a set of records. You can store it in a variable s and write:

>>> s = db(q)

Notice that no database query has been performed so far. DAL + Query simply define a set of records in this db that match the query. web2py determines from the query which table (or tables) are involved and, in fact, there is no need to specify that.


Given a Set, s, you can fetch the records with the command select:


>>> rows =

It returns an iterable object of class pydal.objects.Rows whose elements are Row objects. pydal.objects.Row objects act like dictionaries, but their elements can also be accessed as attributes, like former differ from the latter because its values are read-only.

The Rows object allows looping over the result of the select and printing the selected field values for each row:

>>> for row in rows:
1 Alex

You can do all the steps in one statement:

>>> for row in db('Alex').select():

The select command can take arguments. All unnamed arguments are interpreted as the names of the fields that you want to fetch. For example, you can be explicit on fetching field "id" and field "name":

>>> for row in db().select(,

The table attribute ALL allows you to specify all fields:

>>> for row in db().select(db.person.ALL):

Notice that there is no query string passed to db. web2py understands that if you want all fields of the table person without additional information then you want all records of the table person.

An equivalent alternative syntax is the following:

>>> for row in db( > 0).select():

and web2py understands that if you ask for all records of the table person (id > 0) without additional information, then you want all the fields of table person.

Given one row

row = rows[0]

you can extract its values using multiple equivalent expressions:

>>> row['name']
>>> row('')

The latter syntax is particularly handy when selecting en expression instead of a column. We will show this later.

You can also do

rows.compact = False

to disable the notation


and enable, instead, the less compact notation:


Yes this is unusual and rarely needed.


DAL shortcuts

The DAL supports various code-simplifying shortcuts. In particular:

myrecord = db.mytable[id]

returns the record with the given id if it exists. If the id does not exist, it returns None. The above statement is equivalent to

myrecord = db(

You can delete records by id:

del db.mytable[id]

and this is equivalent to


and deletes the record with the given id, if it exists.

You can insert records:

db.mytable[0] = dict(myfield='somevalue')

It is equivalent to


and it creates a new record with field values specified by the dictionary on the right hand side.

You can update records:

db.mytable[id] = dict(myfield='somevalue')

which is equivalent to


and it updates an existing record with field values specified by the dictionary on the right hand side.

Fetching a Row

Yet another convenient syntax is the following:

record = db.mytable(id)
record = db.mytable(
record = db.mytable(id,myfield='somevalue')

Apparently similar to db.mytable[id] the above syntax is more flexible and safer. First of all it checks whether id is an int (or str(id) is an int) and returns None if not (it never raises an exception). It also allows to specify multiple conditions that the record must meet. If they are not met, it also returns None.

Recursive selects

recursive selects

Consider the previous table person and a new table "thing" referencing a "person":

>>> db.define_table('thing',
        Field('owner','reference person'))

and a simple select from this table:

>>> things = db(db.thing).select()

which is equivalent to

>>> things = db(db.thing._id>0).select()

where ._id is a reference to the primary key of the table. Normally db.thing._id is the same as and we will assume that in most of this book.


For each Row of things it is possible to fetch not just fields from the selected table (thing) but also from linked tables (recursively):

>>> for thing in things: print,

Here requires one database select for each thing in things and it is therefore inefficient. We suggest using joins whenever possible instead of recursive selects, nevertheless this is convenient and practical when accessing individual records.

You can also do it backwards, by selecting the things referenced by a person:

person =  db.person(id)
for thing in
    print, 'owns',

In this last expressions person.thing is a shortcut for


i.e. the Set of things referenced by the current person. This syntax breaks down if the referencing table has multiple references to the referenced table. In this case one needs to be more explicit and use a full Query.

Serializing Rows in views

Given the following action containing a query

def index()
    return dict(rows = db(query).select())

The result of a select can be displayed in a view with the following syntax:

{{extend 'layout.html'}}

Which is equivalent to:

{{extend 'layout.html'}}

SQLTABLE converts the rows into an HTML table with a header containing the column names and one row per record. The rows are marked as alternating class "even" and class "odd". Under the hood, Rows is first converted into a SQLTABLE object (not to be confused with Table) and then serialized. The values extracted from the database are also formatted by the validators associated to the field and then escaped.

Yet it is possible and sometimes convenient to call SQLTABLE explicitly.

The SQLTABLE constructor takes the following optional arguments:

  • linkto the URL or an action to be used to link reference fields (default to None)
  • upload the URL or the download action to allow downloading of uploaded files (default to None)
  • headers a dictionary mapping field names to their labels to be used as headers (default to {}). It can also be an instruction. Currently we support headers='fieldname:capitalize'.
  • truncate the number of characters for truncating long values in the table (default is 16)
  • columns the list of fieldnames to be shown as columns (in tablename.fieldname format). Those not listed are not displayed (defaults to all).
  • **attributes generic helper attributes to be passed to the most external TABLE object.

Here is an example:

{{extend 'layout.html'}}


SQLTABLE is useful but there are times when one needs more. SQLFORM.grid is an extension of SQLTABLE that creates a table with search features and pagination, as well as ability to open detailed records, create, edit and delete records. SQLFORM.smartgrid is a further generalization that allows all of the above but also creates buttons to access referencing records.

Here is an example of usage of SQLFORM.grid:

def index():
    return dict(grid=SQLFORM.grid(query))

and the corresponding view:

{{extend 'layout.html'}}

SQLFORM.grid and SQLFORM.smartgrid should be preferred to SQLTABLE because they are more powerful although higher level and therefore more constraining. They will be explained in more detail in chapter 8.

orderby, groupby, limitby, distinct, having

The select command takes five optional arguments: orderby, groupby, limitby, left and cache. Here we discuss the first three.

You can fetch the records sorted by name:


>>> for row in db().select(

You can fetch the records sorted by name in reverse order (notice the tilde):

>>> for row in db().select(

You can have the fetched records appear in random order:

>>> for row in db().select(
        db.person.ALL, orderby='<random>'):

The use of orderby='<random>' is not supported on Google NoSQL. However, in this situation and likewise in many others where built-ins are insufficient, imports can be used:

import random
rows=db(...).select().sort(lambda row: random.random())

And you can sort the records according to multiple fields by concatenating them with a "|":

>>> for row in db().select(

Using groupby together with orderby, you can group records with the same value for the specified field (this is back-end specific, and is not on the Google NoSQL):

>>> for row in db().select(

You can use having in conjunction with groupby to group conditionally (only those having the condition are grouped.

>>> print db(query1).select(db.person.ALL,, having=query2)

Notice that query1 filters records to be displayed, query2 filters records to be grouped.


With the argument distinct=True, you can specify that you only want to select distinct records. This has the same effect as grouping using all specified fields except that it does not require sorting. When using distinct it is important not to select ALL fields, and in particular not to select the "id" field, else all records will always be distinct.

Here is an example:

>>> for row in db().select(, distinct=True):

Notice that distinct can also be an expression for example:

>>> for row in db().select(,

With limitby=(min, max), you can select a subset of the records from offset=min to but not including offset=max (in this case, the first two starting at zero):

>>> for row in db().select(db.person.ALL, limitby=(0, 2)):

Logical operators

Queries can be combined using the binary AND operator "&":


>>> rows = db(('Alex') & (>3)).select()
>>> for row in rows: print,
4 Alex

and the binary OR operator "|":

>>> rows = db(('Alex') | (>3)).select()
>>> for row in rows: print,
1 Alex

You can negate a query (or sub-query) with the "!=" binary operator:

>>> rows = db((!='Alex') | (>3)).select()
>>> for row in rows: print,
2 Bob
3 Carl

or by explicit negation with the "~" unary operator:

>>> rows = db(~('Alex') | (>3)).select()
>>> for row in rows: print,
2 Bob
3 Carl

Due to Python restrictions in overloading "and" and "or" operators, these cannot be used in forming queries. The binary operators "&" and "|" must be used instead. Note that these operators (unlike "and" and "or") have higher precedence than comparison operators, so the "extra" parentheses in the above examples are mandatory. Similarly, the unary operator "~" has higher precedence than comparison operators, so ~-negated comparisons must also be parenthesized.

It is also possible to build queries using in-place logical operators:

>>> query =!='Alex'
>>> query &=>3
>>> query |='John'

count, isempty, delete, update

You can count records in a set:


>>> print db( > 0).count()

Notice that count takes an optional distinct argument which defaults to False, and it works very much like the same argument for select. count has also a cache argument that works very much like the equivalent argument of the select method.

Sometimes you may need to check if a table is empty. A more efficient way than counting is using the isempty method:

>>> print db( > 0).isempty()

or equivalently:

>>> print db(db.person).isempty()

You can delete records in a set:

>>> db( > 3).delete()

And you can update all records in a set by passing named arguments corresponding to the fields that need to be updated:

>>> db( > 3).update(name='Ken')


The value assigned an update statement can be an expression. For example consider this model

>>> db.define_table('person',
        Field('visits', 'integer', default=0))
>>> db( == 'Massimo').update(
        visits = db.person.visits + 1)

The values used in queries can also be expressions

>>> db.define_table('person',
        Field('visits', 'integer', default=0),
        Field('clicks', 'integer', default=0))
>>> db(db.person.visits == db.person.clicks + 1).delete()


An expression can contain a case clause for example:

>>> db.define_table('person',Field('name'))
>>> condition ='M')
>>> yes_or_no ='Yes','No')
>>> for row in db().select(, yes_or_no):
...     print,  row(yes_or_no)
Max Yes
John No



web2py also allows updating a single record that is already in memory using update_record

>>> row = db(
>>> row.update_record(name='Curt')

update_record should not be confused with

>>> row.update(name='Curt')

because for a single row, the method update updates the row object but not the database record, as in the case of update_record.

It is also possible to change the attributes of a row (one at a time) and then call update_record() without arguments to save the changes:

>>> row = db( > 2).select().first()
>>> = 'Curt'
>>> row.update_record() # saves above change

The update_record method is available only if the table's id field is included in the select, and cacheable is not set to True.

Inserting and updating from a dictionary

A common issue consists of needing to insert or update records in a table where the name of the table, the field to be updated, and the value for the field are all stored in variables. For example: tablename, fieldname, and value.

The insert can be done using the following syntax:


The update of record with given id can be done with:



Notice we used table._id instead of In this way the query works even for tables with a field of type "id" which has a name other than "id".

first and last


Given a Rows object containing records:

>>> rows = db(query).select()
>>> first_row = rows.first()
>>> last_row = rows.last()

are equivalent to

>>> first_row = rows[0] if len(rows)>0 else None
>>> last_row = rows[-1] if len(rows)>0 else None

as_dict and as_list


A Row object can be serialized into a regular dictionary using the as_dict() method and a Rows object can be serialized into a list of dictionaries using the as_list() method. Here are some examples:

>>> rows = db(query).select()
>>> rows_list = rows.as_list()
>>> first_row_dict = rows.first().as_dict()

These methods are convenient for passing Rows to generic views and or to store Rows in sessions (since Rows objects themselves cannot be serialized since contain a reference to an open DB connection):

>>> rows = db(query).select()
>>> session.rows = rows # not allowed!
>>> session.rows = rows.as_list() # allowed!

Combining rows

Row objects can be combined at the Python level. Here we assume:

>>> print rows1
>>> print rows2

You can do a union of the records in two set of rows:

>>> rows3 = rows1 & rows2
>>> print rows3

You can do a union of the records removing duplicates:

>>> rows3 = rows1 | rows2
>>> print rows3

find, exclude, sort


There are times when one needs to perform two selects and one contains a subset of a previous select. In this case it is pointless to access the database again. The find, exclude and sort objects allow you to manipulate a Rows objects and generate another one without accessing the database. More specifically:

  • find returns a new set of Rows filtered by a condition and leaves the original unchanged.
  • exclude returns a new set of Rows filtered by a condition and removes them from the original Rows.
  • sort returns a new set of Rows sorted by a condition and leaves the original unchanged.

All these methods take a single argument, a function that acts on each individual row.

Here is an example of usage:

>>> db.define_table('person',Field('name'))
>>> db.person.insert(name='John')
>>> db.person.insert(name='Max')
>>> db.person.insert(name='Alex')
>>> rows = db(db.person).select()
>>> for row in rows.find(lambda row:[0]=='M'):
>>> print len(rows)
>>> for row in rows.exclude(lambda row:[0]=='M'):
>>> print len(rows)
>>> for row in rows.sort(lambda row:

They can be combined:

>>> rows = db(db.person).select()
>>> rows = rows.find(
        lambda row: 'x' in
            lambda row:
>>> for row in rows:

Sort takes an optional argument reverse=True with the obvious meaning.

The find method as an optional limitby argument with the same syntax and functionality as the Set select method.

Other methods



Some times you need to perform an insert only if there is no record with the same values as those being inserted. This can be done with


The record will be inserted only of there is no other user called John born in Chicago.

You can specify which values to use as a key to determine if the record exists. For example:


and if there is John his birthplace will be updated else a new record will be created.

validate_and_insert, validate_and_update


The function

ret = db.mytable.validate_and_insert(field='value')

works very much like

id = db.mytable.insert(field='value')

except that it calls the validators for the fields before performing the insert and bails out if the validation does not pass. If validation does not pass the errors can be found in ret.error. If it passes, the id of the new record is in Mind that normally validation is done by the form processing logic so this function is rarely needed.


ret = db(query).validate_and_update(field='value')

works very much the same as

num = db(query).update(field='value')

except that it calls the validators for the fields before performing the update. Notice that it only works if query involves a single table. The number of updated records can be found in res.updated and errors will be ret.errors.

smart_query (experimental)

There are times when you need to parse a query using natural language such as

name contain m and age greater than 18

The DAL provides a method to parse this type of queries:

search = 'name contain m and age greater than 18'
rows = db.smart_query([db.person],search).select()

The first argument must be a list of tables or fields that should be allowed in the search. It raises a RuntimeError if the search string is invalid. This functionality can be used to build RESTful interfaces (see chapter 10) and it is used internally by the SQLFORM.grid and SQLFORM.smartgrid.

In the smartquery search string, a field can be identified by fieldname only and or by tablename.fieldname. Strings may be delimited by double quotes if they contain spaces.

Computed fields


DAL fields may have a compute attribute. This must be a function (or lambda) that takes a Row object and returns a value for the field. When a new record is modified, including both insertions and updates, if a value for the field is not provided, web2py tries to compute from the other field values using the compute function. Here is an example:

>>> db.define_table('item',
            compute=lambda r: r['unit_price']*r['quantity']))
>>> r = db.item.insert(unit_price=1.99, quantity=5)
>>> print r.total_price

Notice that the computed value is stored in the db and it is not computed on retrieval, as in the case of virtual fields, described later. Two typical applications of computed fields are:

  • in wiki applications, to store the processed input wiki text as HTML, to avoid re-processing on every request
  • for searching, to compute normalized values for a field, to be used for searching.

Virtual fields

virtual fields

Virtual fields are also computed fields (as in the previous subsection) but they differ from those because they are virtual in the sense that they are not stored in the db and they are computed each time records are extracted from the database. They can be used to simplify the user's code without using additional storage but they cannot be used for searching.

New style virtual fields

web2py provides a new and easier way to define virtual fields and lazy virtual fields. This section is marked experimental because they APIs may still change a little from what is described here.

Here we will consider the same example as in the previous subsection. In particular we consider the following model:

>>> db.define_table('item',

One can define a total_price virtual field as

>>> db.item.total_price = Field.Virtual(
    lambda row: row.item.unit_price*row.item.quantity)

i.e. by simply defining a new field total_price to be a Field.Virtual. The only argument of the constructor is a function that takes a row and returns the computed values.

A virtual field defined as the one above is automatically computed for all records when the records are selected:

>>> for row in db(db.item).select(): print row.total_price

It is also possible to define method fields which are calculated on-demand, when called. For example:

>>> db.item.discounted_total = Field.Method(lambda row, discount=0.0:        row.item.unit_price*row.item.quantity*(1.0-discount/100))

In this case row.discounted_total is not a value but a function. The function takes the same arguments as the function passed to the Method constructor except for row which is implicit (think of it as self for rows objects).

The lazy field in the example above allows one to compute the total price for each item:

>>> for row in db(db.item).select(): print row.discounted_total()

And it also allows to pass an optional discount percentage (15%):

>>> for row in db(db.item).select(): print row.discounted_total(15)

Virtual and Method fields can also be defined in place when a table is defined:

>>> db.define_table('item',
        Field.Virtual('total_price', lambda row: ...),
        Field.Method('discounted_total', lambda row, discount=0.0: ...))

Mind that virtual fields do not have the same attributes as the other fields (default, readable, requires, etc) and they do not appear in the list of db.table.fields and are not visualized by default in tables (TABLE) and grids (SQLFORM.grid, SQLFORM.smartgrid).

Old style virtual fields

In order to define one or more virtual fields, you can also define a container class, instantiate it and link it to a table or to a select. For example, consider the following table:

>>> db.define_table('item',

One can define a total_price virtual field as

>>> class MyVirtualFields(object):
        def total_price(self):
            return self.item.unit_price*self.item.quantity
>>> db.item.virtualfields.append(MyVirtualFields())

Notice that each method of the class that takes a single argument (self) is a new virtual field. self refers to each one row of the select. Field values are referred by full path as in self.item.unit_price. The table is linked to the virtual fields by appending an instance of the class to the table's virtualfields attribute.

Virtual fields can also access recursive fields as in

>>> db.define_table('item',
>>> db.define_table('order_item',
        Field('item','reference item'),
>>> class MyVirtualFields(object):
        def total_price(self):
            return self.order_item.item.unit_price                 * self.order_item.quantity
>>> db.order_item.virtualfields.append(MyVirtualFields())

Notice the recursive field access self.order_item.item.unit_price where self is the looping record.

They can also act on the result of a JOIN

>>> db.define_table('item',
>>> db.define_table('order_item',
        Field('item','reference item'),
>>> rows = db(
>>> class MyVirtualFields(object):
        def total_price(self):
            return self.item.unit_price                 * self.order_item.quantity
>>> rows.setvirtualfields(order_item=MyVirtualFields())
>>> for row in rows: print row.order_item.total_price

Notice how in this case the syntax is different. The virtual field accesses both self.item.unit_price and self.order_item.quantity which belong to the join select. The virtual field is attached to the rows of the table using the setvirtualfields method of the rows object. This method takes an arbitrary number of named arguments and can be used to set multiple virtual fields, defined in multiple classes, and attach them to multiple tables:

>>> class MyVirtualFields1(object):
        def discounted_unit_price(self):
            return self.item.unit_price*0.90
>>> class MyVirtualFields2(object):
        def total_price(self):
            return self.item.unit_price                 * self.order_item.quantity
        def discounted_total_price(self):
            return self.item.discounted_unit_price                 * self.order_item.quantity
>>> rows.setvirtualfields(
>>> for row in rows:
        print row.order_item.discounted_total_price

Virtual fields can be lazy; all they need to do is return a function and access it by calling the function:

>>> db.define_table('item',
>>> class MyVirtualFields(object):
        def lazy_total_price(self):
            def lazy(self=self):
                return self.item.unit_price                     * self.item.quantity
            return lazy
>>> db.item.virtualfields.append(MyVirtualFields())
>>> for item in db(db.item).select():
        print item.lazy_total_price()

or shorter using a lambda function:

>>> class MyVirtualFields(object):
        def lazy_total_price(self):
            return lambda self=self: self.item.unit_price                 * self.item.quantity

One to many relation

one to many

To illustrate how to implement one to many relations with the web2py DAL, define another table "thing" that refers to the table "person" which we redefine here:

>>> db.define_table('person',
>>> db.define_table('thing',
                    Field('owner', 'reference person'),

Table "thing" has two fields, the name of the thing and the owner of the thing. The "owner" field id a reference field. A reference type can be specified in two equivalent ways:

Field('owner', 'reference person')
Field('owner', db.person)

The latter is always converted to the former. They are equivalent except in the case of lazy tables, self references or other types of cyclic references where the former notation is the only allowed notation.

When a field type is another table, it is intended that the field reference the other table by its id. In fact, you can print the actual type value and get:

>>> print db.thing.owner.type
reference person

Now, insert three things, two owned by Alex and one by Bob:

>>> db.thing.insert(name='Boat', owner=1)
>>> db.thing.insert(name='Chair', owner=1)
>>> db.thing.insert(name='Shoes', owner=2)

You can select as you did for any other table:

>>> for row in db(db.thing.owner==1).select():

Because a thing has a reference to a person, a person can have many things, so a record of table person now acquires a new attribute thing, which is a Set, that defines the things of that person. This allows looping over all persons and fetching their things easily:

>>> for person in db().select(db.person.ALL):
        for thing in
            print '    ',

Inner joins

Another way to achieve a similar result is by using a join, specifically an INNER JOIN. web2py performs joins automatically and transparently when the query links two or more tables as in the following example:

inner join

>>> rows = db(
>>> for row in rows:
        print, 'has',
Alex has Boat
Alex has Chair
Bob has Shoes

Observe that web2py did a join, so the rows now contain two records, one from each table, linked together. Because the two records may have fields with conflicting names, you need to specify the table when extracting a field value from a row. This means that while before you could do:

and it was obvious whether this was the name of a person or a thing, in the result of a join you have to be more explicit and say:


There is an alternative syntax for INNER JOINS:

>>> rows = db(db.person).select(join=db.thing.on(
>>> for row in rows:
    print, 'has',
Alex has Boat
Alex has Chair
Bob has Shoes

While the output is the same, the generated SQL in the two cases can be different. The latter syntax removes possible ambiguities when the same table is joined twice and aliased:

>>> db.define_table('thing',
        Field('owner1','reference person'),
        Field('owner2','reference person'))
>>> rows = db(db.person).select(

The value of join can be list of db.table.on(...) to join.

Left outer join

Notice that Carl did not appear in the list above because he has no things. If you intend to select on persons (whether they have things or not) and their things (if they have any), then you need to perform a LEFT OUTER JOIN. This is done using the argument "left" of the select command. Here is an example:

left outer join
outer join

>>> rows=db().select(
        db.person.ALL, db.thing.ALL,
>>> for row in rows:
        print, 'has',
Alex has Boat
Alex has Chair
Bob has Shoes
Carl has None


left = db.thing.on(...)

does the left join query. Here the argument of db.thing.on is the condition required for the join (the same used above for the inner join). In the case of a left join, it is necessary to be explicit about which fields to select.

Multiple left joins can be combined by passing a list or tuple of db.mytable.on(...) to the left attribute.

Grouping and counting

When doing joins, sometimes you want to group rows according to certain criteria and count them. For example, count the number of things owned by every person. web2py allows this as well. First, you need a count operator. Second, you want to join the person table with the thing table by owner. Third, you want to select all rows (person + thing), group them by person, and count them while grouping:

>>> count =
>>> for row in db(, count,
        print, row[count]
Alex 2
Bob 1

Notice the count operator (which is built-in) is used as a field. The only issue here is in how to retrieve the information. Each row clearly contains a person and the count, but the count is not a field of a person nor is it a table. So where does it go? It goes into the storage object representing the record with a key equal to the query expression itself. The count method of the Field object has an optional distinct argument. When set to True it specifies that only distinct values of the field in question are to be counted.

Many to many


In the previous examples, we allowed a thing to have one owner but one person could have many things. What if Boat was owned by Alex and Curt? This requires a many-to-many relation, and it is realized via an intermediate table that links a person to a thing via an ownership relation.

Here is how to do it:

>>> db.define_table('person',
>>> db.define_table('thing',
>>> db.define_table('ownership',
                    Field('person', 'reference person'),
                    Field('thing', 'reference thing'))

the existing ownership relationship can now be rewritten as:

>>> db.ownership.insert(person=1, thing=1) # Alex owns Boat
>>> db.ownership.insert(person=1, thing=2) # Alex owns Chair
>>> db.ownership.insert(person=2, thing=3) # Bob owns Shoes

Now you can add the new relation that Curt co-owns Boat:

>>> db.ownership.insert(person=3, thing=1) # Curt owns Boat too

Because you now have a three-way relation between tables, it may be convenient to define a new set on which to perform operations:

>>> persons_and_things = db(
        (         & (

Now it is easy to select all persons and their things from the new Set:

>>> for row in
Alex Boat
Alex Chair
Bob Shoes
Curt Boat

Similarly, you can search for all things owned by Alex:

>>> for row in persons_and_things('Alex').select():

and all owners of Boat:

>>> for row in persons_and_things('Boat').select():

A lighter alternative to Many 2 Many relations is tagging. Tagging is discussed in the context of the IS_IN_DB validator. Tagging works even on database backends that do not support JOINs like the Google App Engine NoSQL.

list:<type>, and contains


web2py provides the following special field types:

list:reference <table>

They can contain lists of strings, of integers and of references respectively.

On Google App Engine NoSQL list:string is mapped into StringListProperty, the other two are mapped into ListProperty(int). On relational databases they all mapped into text fields which contain the list of items separated by |. For example [1,2,3] is mapped into |1|2|3|.

For lists of string the items are escaped so that any | in the item is replaced by a ||. Anyway this is an internal representation and it is transparent to the user.

You can use list:string, for example, in the following way:

>>> db.define_table('product',
>>> db.product.colors.requires=IS_IN_SET(('red','blue','green'))
>>> db.product.insert(name='Toy Car',colors=['red','green'])
>>> products = db(db.product.colors.contains('red')).select()
>>> for item in products:
        print, item.colors
Toy Car ['red', 'green']

list:integer works in the same way but the items must be integers.

As usual the requirements are enforced at the level of forms, not at the level of insert.

For list:<type> fields the contains(value) operator maps into a non trivial query that checks for lists containing the value. The contains operator also works for regular string and text fields and it maps into a LIKE '%value%'.

The list:reference and the contains(value) operator are particularly useful to de-normalize many-to-many relations. Here is an example:

>>> db.define_table('tag',Field('name'),format='%(name)s')
>>> db.define_table('product',
        Field('tags','list:reference tag'))
>>> a = db.tag.insert(name='red')
>>> b = db.tag.insert(name='green')
>>> c = db.tag.insert(name='blue')
>>> db.product.insert(name='Toy Car',tags=[a, b, c])
>>> products = db(db.product.tags.contains(b)).select()
>>> for item in products:
        print, item.tags
Toy Car [1, 2, 3]
>>> for item in products:
        print, db.product.tags.represent(item.tags)
Toy Car red, green, blue

Notice that a list:reference tag field get a default constraint

requires = IS_IN_DB(db,'',db.tag._format,multiple=True)

that produces a SELECT/OPTION multiple drop-box in forms.

Also notice that this field gets a default represent attribute which represents the list of references as a comma-separated list of formatted references. This is used in read forms and SQLTABLEs.

While list:reference has a default validator and a default representation, list:integer and list:string do not. So these two need an IS_IN_SET or an IS_IN_DB validator if you want to use them in forms.

Other operators

web2py has other operators that provide an API to access equivalent SQL operators. Let's define another table "log" to store security events, their event_time and severity, where the severity is an integer number.


>>> db.define_table('log', Field('event'),
                           Field('event_time', 'datetime'),
                           Field('severity', 'integer'))

As before, insert a few events, a "port scan", an "xss injection" and an "unauthorized login". For the sake of the example, you can log events with the same event_time but with different severities (1, 2, and 3 respectively).

>>> import datetime
>>> now =
>>> print db.log.insert(
        event='port scan', event_time=now, severity=1)
>>> print db.log.insert(
        event='xss injection', event_time=now, severity=2)
>>> print db.log.insert(
        event='unauthorized login', event_time=now, severity=3)

like, regexp, startswith, contains, upper, lower


Fields have a like operator that you can use to match strings:

>>> for row in db('port%')).select():
        print row.event
port scan

Here "port%" indicates a string starting with "port". The percent sign character, "%", is a wild-card character that means "any sequence of characters".

The like operator is case-insensitive but it can be made case-sensitive with'value',case_sensitive=True)

web2py also provides some shortcuts:


which are equivalent respectively to'value%')'%value%')

Notice that contains has a special meaning for list:<type> fields and it was discussed in a previous section.

The contains method can also be passed a list of values and an optional boolean argument all to search for records that contain all values:

db.mytable.myfield.contains(['value1','value2'], all=True)

or any value from the list

db.mytable.myfield.contains(['value1','value2'], all=false)

There is a also a regexp method that works like the like method but allows regular expression syntax for the look-up expression. It is only supported by PostgreSQL and SQLite.

The upper and lower methods allow you to convert the value of the field to upper or lower case, and you can also combine them with the like operator:


>>> for row in db(db.log.event.upper().like('PORT%')).select():
        print row.event
port scan

year, month, day, hour, minutes, seconds


The date and datetime fields have day, month and year methods. The datetime and time fields have hour, minutes and seconds methods. Here is an example:

>>> for row in db(db.log.event_time.year()==2009).select():
        print row.event
port scan
xss injection
unauthorized login


The SQL IN operator is realized via the belongs method which returns true when the field value belongs to the specified set (list of tuples):

>>> for row in db(db.log.severity.belongs((1, 2))).select():
        print row.event
port scan
xss injection

The DAL also allows a nested select as the argument of the belongs operator. The only caveat is that the nested select has to be a _select, not a select, and only one field has to be selected explicitly, the one that defines the set.

nested select
>>> bad_days = db(db.log.severity==3)._select(db.log.event_time)
>>> for row in db(db.log.event_time.belongs(bad_days)).select():
        print row.event
port scan
xss injection
unauthorized login

In those cases where a nested select is required and the look-up field is a reference we can also use a query as argument. For example:

db.define_table('thing',Field('owner'),Field('owner','reference thing'))

In this case it is obvious that the next select only needs the field referenced by the db.thing.owner field so we do not need the more verbose _select notation.


A nested select can also be used as insert/update value but in this case the syntax is different:

lazy = db('Jonathan').nested_select(
db( = lazy)

In this case lazy is a nested expression that computes the id of person "Jonathan". The two lines result in one single SQL query.

sum, avg, min, max and len

Previously, you have used the count operator to count records. Similarly, you can use the sum operator to add (sum) the values of a specific field from a group of records. As in the case of count, the result of a sum is retrieved via the store object:

>>> sum = db.log.severity.sum()
>>> print db().select(sum).first()[sum]

You can also use avg, min, and max to the average, minimum, and maximum value respectively for the selected records. For example:

>>> max = db.log.severity.max()
>>> print db().select(max).first()[max]

.len() computes the length of a string, text or boolean fields.

Expressions can be combined to form more complex expressions. For example here we are computing the sum of the length of all the severity strings in the logs, increased of one:

>>> sum = (db.log.severity.len()+1).sum()
>>> print db().select(sum).first()[sum]


One can build an expression to refer to a substring. For example, we can group things whose name starts with the same three characters and select only one from each group:

db(db.thing).select(distinct =[:3])

Default values with coalesce and coalesce_zero

There are times when you need to pull a value from database but also need a default values if the value for a record is set to NULL. In SQL there is a keyword, COALESCE, for this. web2py has an equivalent coalesce method:

>>> db.define_table('sysuser',Field('username'),Field('fullname'))
>>> db.sysuser.insert(username='max',fullname='Max Power')
>>> db.sysuser.insert(username='tim',fullname=None)
print db(db.sysuser).select(db.sysuser.fullname.coalesce(db.sysuser.username))
Max Power

Other times you need to compute a mathematical expression but some fields have a value set to None while it should be zero. coalesce_zero comes to the rescue by defaulting None to zero in the query:

>>> db.define_table('sysuser',Field('username'),Field('points'))
>>> db.sysuser.insert(username='max',points=10)
>>> db.sysuser.insert(username='tim',points=None)
>>> print db(db.sysuser).select(db.sysuser.points.coalesce_zero().sum())

Generating raw sql

raw SQL

Sometimes you need to generate the SQL but not execute it. This is easy to do with web2py since every command that performs database IO has an equivalent command that does not, and simply returns the SQL that would have been executed. These commands have the same names and syntax as the functional ones, but they start with an underscore:

Here is _insert


>>> print db.person._insert(name='Alex')
INSERT INTO person(name) VALUES ('Alex');

Here is _count


>>> print db('Alex')._count()
SELECT count(*) FROM person WHERE'Alex';

Here is _select


>>> print db('Alex')._select()
SELECT, FROM person WHERE'Alex';

Here is _delete


>>> print db('Alex')._delete()

And finally, here is _update


>>> print db('Alex')._update()
UPDATE person SET  WHERE'Alex';

Moreover you can always use db._lastsql to return the most recent SQL code, whether it was executed manually using executesql or was SQL generated by the DAL.

Exporting and importing data


CSV (one Table at a time)

When a DALRows object is converted to a string it is automatically serialized in CSV:

>>> rows = db(
>>> print rows,,,,thing.owner

You can serialize a single table in CSV and store it in a file "test.csv":

>>> open('test.csv', 'w').write(str(db(

This is equivalent to

>>> rows = db(
>>> rows.export_to_csv_file(open('test.csv', 'w'))

You can read the CSV file back with:

>>> db.person.import_from_csv_file(open('test.csv', 'r'))

When importing, web2py looks for the field names in the CSV header. In this example, it finds two columns: "" and "". It ignores the "person." prefix, and it ignores the "id" fields. Then all records are appended and assigned new ids. Both of these operations can be performed via the appadmin web interface.

CSV (all tables at once)

In web2py, you can backup/restore an entire database with two commands:

To export:

>>> db.export_to_csv_file(open('somefile.csv', 'wb'))

To import:

>>> db.import_from_csv_file(open('somefile.csv', 'rb'))

This mechanism can be used even if the importing database is of a different type than the exporting database. The data is stored in "somefile.csv" as a CSV file where each table starts with one line that indicates the tablename, and another line with the fieldnames:

TABLE tablename
field1, field2, field3, ...

Two tables are separated \r\n\r\n. The file ends with the line


The file does not include uploaded files if these are not stored in the database. In any case it is easy enough to zip the "uploads" folder separately.

When importing, the new records will be appended to the database if it is not empty. In general the new imported records will not have the same record id as the original (saved) records but web2py will restore references so they are not broken, even if the id values may change.

If a table contains a field called "uuid", this field will be used to identify duplicates. Also, if an imported record has the same "uuid" as an existing record, the previous record will be updated.

CSV and remote database synchronization

Consider the following model:

db = DAL('sqlite:memory:')
    Field('owner', 'reference person'),

if not db(db.person).count():
    id = db.person.insert(name="Massimo")
    db.thing.insert(owner=id, name="Chair")

Each record is identified by an ID and referenced by that ID. If you have two copies of the database used by distinct web2py installations, the ID is unique only within each database and not across the databases. This is a problem when merging records from different databases.

In order to make a record uniquely identifiable across databases, they must:

  • have a unique id (UUID),
  • have an event_time (to figure out which one is more recent if multiple copies),
  • reference the UUID instead of the id.

This can be achieved without modifying web2py. Here is what to do:

1. Change the above model into:

    Field('uuid', length=64, default=lambda:str(uuid.uuid4())),
    Field('modified_on', 'datetime', default=now),

    Field('uuid', length=64, default=lambda:str(uuid.uuid4())),
    Field('modified_on', 'datetime', default=now),
    Field('owner', length=64),

db.thing.owner.requires = IS_IN_DB(db,'person.uuid','%(name)s')

if not db(
    id = uuid.uuid4()
    db.person.insert(name="Massimo", uuid=id)
    db.thing.insert(owner=id, name="Chair")

Note, in the above table definitions, the default value for the two 'uuid' fields is set to a lambda function, which returns a UUID (converted to a string). The lambda function is called once for each record inserted, ensuring that each record gets a unique UUID, even if multiple records are inserted in a single transaction.

2. Create a controller action to export the database:

def export():
    s = StringIO.StringIO()
    response.headers['Content-Type'] = 'text/csv'
    return s.getvalue()

3. Create a controller action to import a saved copy of the other database and sync records:

def import_and_sync():
    form = FORM(INPUT(_type='file', _name='data'), INPUT(_type='submit'))
    if form.process(session=None).accepted:
        # for every table
        for table in db.tables:
            # for every uuid, delete all but the latest
            items = db(db[table]).select(db[table].id,
            for item in items:
                db((db[table].uuid==item.uuid)&                   (db[table].id!
    return dict(form=form)

Notice that session=None disables the CSRF protection since this URL is intended to be accessed from outside.

4. Create an index manually to make the search by uuid faster.

Notice that steps 2 and 3 work for every database model; they are not specific for this example.


Alternatively, you can use XML-RPC to export/import the file.

If the records reference uploaded files, you also need to export/import the content of the uploads folder. Notice that files therein are already labeled by UUIDs so you do not need to worry about naming conflicts and references.

HTML and XML (one Table at a time)

DALRows objects

DALRows objects also have an xml method (like helpers) that serializes it to XML/HTML:

>>> rows = db( > 0).select()
>>> print rows.xml()
    <tr class="even">
DALRows custom tags

If you need to serialize the DALRows in any other XML format with custom tags, you can easily do that using the universal TAG helper and the * notation:

>>> rows = db( > 0).select()
>>> print TAG.result(*[TAG.row(*[TAG.field(r[f], _name=f)           for f in db.person.fields]) for r in rows])
    <field name="id">1</field>
    <field name="name">Alex</field>

Data representation


The export_to_csv_file function accepts a keyword argument named represent. When True it will use the columns represent function while exporting the data instead of the raw data.


The function also accepts a keyword argument named colnames that should contain a list of column names one wish to export. It defaults to all columns.

Both export_to_csv_file and import_from_csv_file accept keyword arguments that tell the csv parser the format to save/load the files:

  • delimiter: delimiter to separate values (default ',')
  • quotechar: character to use to quote string values (default to double quotes)
  • quoting: quote system (default csv.QUOTE_MINIMAL)

Here is some example usage:

>>> import csv
>>> rows = db(query).select()
>>> rows.export_to_csv_file(open('/tmp/test.txt', 'w'),

Which would render something similar to

"hello"|35|"this is the text description"|"2009-03-03"

For more information consult the official Python documentation [quoteall]

Caching selects

The select method also takes a cache argument, which defaults to None. For caching purposes, it should be set to a tuple where the first element is the cache model (cache.ram, cache.disk, etc.), and the second element is the expiration time in seconds.

In the following example, you see a controller that caches a select on the previously defined db.log table. The actual select fetches data from the back-end database no more frequently than once every 60 seconds and stores the result in cache.ram. If the next call to this controller occurs in less than 60 seconds since the last database IO, it simply fetches the previous data from cache.ram.

cache select
def cache_db_select():
    logs = db().select(db.log.ALL, cache=(cache.ram, 60))
    return dict(logs=logs)

The select method has an optional cacheable argument, normally set to False. When cacheable=True the resulting Rows is serializable but The Rows lack update_record and delete_record methods.

If you do not need these methods you can speed up selects a lot by setting the cacheable attribute:

rows = db(query).select(cacheable=True)

The results of a select are normally complex, un-pickleable objects; they cannot be stored in a session and cannot be cached in any other way than the one explained here unless the cache attribute is set or cacheable=True.

When the cache argument is set but cacheable=False (default) only the database results are cached, not the actual Rows object. When the cache argument is used in conjunction with cacheable=True the entire Rows object is cached and this results in much faster caching:

rows = db(query).select(cache=(cache.ram,3600),cacheable=True)

Self-Reference and aliases

self reference

It is possible to define tables with fields that refer to themselves, here is an example:

reference table
    Field('father_id', 'reference person'),
    Field('mother_id', 'reference person'))

Notice that the alternative notation of using a table object as field type will fail in this case, because it uses a variable db.person before it is defined:

    Field('father_id', db.person), # wrong!
    Field('mother_id', db.person)) # wrong!

In general db.tablename and "reference tablename" are equivalent field types, but the latter is the only one allowed for self.references.


If the table refers to itself, then it is not possible to perform a JOIN to select a person and its parents without use of the SQL "AS" keyword. This is achieved in web2py using the with_alias. Here is an example:

>>> Father = db.person.with_alias('father')
>>> Mother = db.person.with_alias('mother')
>>> db.person.insert(name='Massimo')
>>> db.person.insert(name='Claudia')
>>> db.person.insert(name='Marco', father_id=1, mother_id=2)
>>> rows = db().select(,,,
>>> for row in rows:
Massimo None None
Claudia None None
Marco Massimo Claudia

Notice that we have chosen to make a distinction between:

  • "father_id": the field name used in the table "person";
  • "father": the alias we want to use for the table referenced by the above field; this is communicated to the database;
  • "Father": the variable used by web2py to refer to that alias.

The difference is subtle, and there is nothing wrong in using the same name for the three of them:

    Field('father', 'reference person'),
    Field('mother', 'reference person'))
>>> father = db.person.with_alias('father')
>>> mother = db.person.with_alias('mother')
>>> db.person.insert(name='Massimo')
>>> db.person.insert(name='Claudia')
>>> db.person.insert(name='Marco', father=1, mother=2)
>>> rows = db().select(,,,
>>> for row in rows:
Massimo None None
Claudia None None
Marco Massimo Claudia

But it is important to have the distinction clear in order to build correct queries.

Advanced features

Table inheritance


It is possible to create a table that contains all the fields from another table. It is sufficient to pass the other table in place of a field to define_table. For example

db.define_table('person', Field('name'))
db.define_table('doctor', db.person, Field('specialization'))
dummy table

It is also possible to define a dummy table that is not stored in a database in order to reuse it in multiple other places. For example:

signature = db.Table(db, 'signature',
    Field('created_on', 'datetime',,
    Field('created_by', db.auth_user, default=auth.user_id),
    Field('updated_on', 'datetime',,
    Field('updated_by', db.auth_user, update=auth.user_id))

db.define_table('payment', Field('amount', 'double'), signature)

This example assumes that standard web2py authentication is enabled.

Notice that if you use Auth web2py already creates one such table for you:

auth = Auth(db)
db.define_table('payment', Field('amount', 'double'), auth.signature)

When using table inheritance, if you want the inheriting table to inherit validators, be sure to define the validators of the parent table before defining the inheriting table.

filter_in and filter_out


It is possible to define a filter for each field to be called before a value is inserted into the database for that field and after a value is retrieved from the database.

Imagine for example that you want to store a serializable Python data structure in a field in the json format. Here is how it could be accomplished:

>>> from simplejson import loads, dumps
>>> db.define_table('anyobj',Field('name'),Field('data','text'))
>>> = lambda obj, dumps=dumps: dumps(obj)
>>> = lambda txt, loads=loads: loads(txt)
>>> myobj = ['hello', 'world', 1, {2: 3}]
>>> id = db.anyobj.insert(name='myobjname', data=myobj)
>>> row = db.anyobj(id)
['hello', 'world', 1, {2: 3}]

Another way to accomplish the same is by using a Field of type SQLCustomType, as discussed later.

before and after callbacks


Web2py provides a mechanism to register callbacks to be called before and/or after insert, update and delete of records.

Each table stores six lists of callbacks:


You can register callback function by appending it the corresponding function to one of those lists. The caveat is that depending on the functionality, the callback has different signature.

This is best explained via some examples.

>>> db.define_table('person',Field('name'))
>>> def pprint(*args): print args
>>> db.person._before_insert.append(lambda f: pprint(f))
>>> db.person._after_insert.append(lambda f,id: pprint(f,id))
>>> db.person._before_update.append(lambda s,f: pprint(s,f))
>>> db.person._after_update.append(lambda s,f: pprint(s,f))
>>> db.person._before_delete.append(lambda s: pprint(s))
>>> db.person._after_delete.append(lambda s: pprint(s))

Here f is a dict of fields passed to insert or update, id is the id of the newly inserted record, s is the Set object used for update or delete.

>>> db.person.insert(name='John')
({'name': 'John'},)
({'name': 'John'}, 1)
>>> db('Tim')
(<Set ( = 1)>, {'name': 'Tim'})
(<Set ( = 1)>, {'name': 'Tim'})
>>> db(
(<Set ( = 1)>,)
(<Set ( = 1)>,)

The return values of these callback should be None or False. If any of the _before_* callback returns a True value it will abort the actual insert/update/delete operation.


Some times a callback may need to perform an update in the same of a different table and one wants to avoid callbacks calling themselves recursively.

For this purpose there the Set objects have an update_naive method that works like update but ignores before and after callbacks.

Record versioning


It is possible to ask web2py to save every copy of a record when the record is modified. There are different ways to do it and it can be done for all tables at once using the syntax:


this requires Auth and it is discussed in the chapter about authentication. It can also be done for each individual table as discussed below.

Consider the following table:


Notice the hidden boolean field called is_active and defaulting to True.

We can tell web2py to create a new table (in the same or a different database) and store all previous versions of each record in the table, when modified.

This is done in the following way:


or in a more verbose syntax:

    archive_db = db,
    archive_name = 'stored_item_archive',
    current_record = 'current_record',
    is_active = 'is_active')

The archive_db=db tells web2py to store the archive table in the same database as the stored_item table. The archive_name sets the name for the archive table. The archive table has the same fields as the original table stored_item except that unique fields are no longer unique (because it needs to store multiple versions) and has an extra field which name is specified by current_record and which is a reference to the current record in the stored_item table.

When records are deleted, they are not really deleted. A deleted record is copied in the stored_item_archive table (like when it is modified) and the is_active field is set to False. By enabling record versioning web2py sets a custom_filter on this table that hides all fields in table stored_item where the is_active field is set to False. The is_active parameter in the _enable_record_versioning method allows to specify the name of the field used by the custom_filter to determine if the field was deleted or not.

custom_filters are ignored by the appadmin interface.

Common fields and multi-tenancy

common fields
multi tenancy

db._common_fields is a list of fields that should belong to all the tables. This list can also contain tables and it is understood as all fields from the table. For example occasionally you find yourself in need to add a signature to all your tables but the `auth tables. In this case, after you db.define_tables() but before defining any other table, insert


One field is special: "request_tenant". This field does not exist but you can create it and add it to any of your tables (or them all):


For every table with a field called db._request_tenant, all records for all queries are always automatically filtered by:

db.table.request_tenant == db.table.request_tenant.default

and for every record insert, this field is set to the default value. In the example above we have chosen

default = request.env.http_host

i.e. we have chose to ask our app to filter all tables in all queries with

db.table.request_tenant == request.env.http_host

This simple trick allow us to turn any application into a multi-tenant application. i.e. even if we run one instance of the app and we use one single database, if the app is accessed under two or more domains (in the example the domain name is retrieved from request.env.http_host) the visitors will see different data depending on the domain. Think of running multiple web stores under different domains with one app and one database.

You can turn off multi tenancy filters using:


rows = db(query, ignore_common_filters=True).select()

Common filters

A common filter is a generalization of the above multi-tenancy idea. It provides an easy way to prevent repeating of the same query. Consider for example the following table:

    Field('post_text', 'text'),
    Field('is_public', 'boolean'),
    common_filter = lambda query: db.blog_post.is_public==True

Any select, delete or update in this table, will include only public blog posts. The attribute can also be changed in controllers:

db.blog_post._common_filter = lambda query: db.blog_post.is_public == True

It serves both as a way to avoid repeating the "db.blog_post.is_public==True" phrase in each blog post search, and also as a security enhancement, that prevents you from forgetting to disallow viewing of none public posts.

In case you actually do want items left out by the common filter (for example, allowing the admin to see none public posts), you can either remove the filter:

db.blog_post._common_filter = None

or ignore it:

db(query, ignore_common_filters=True).select(...)

Custom Field types (experimental)


It is possible to define new/custom field types. For example we consider here the example if a field that contains binary data in compressed form:

from gluon.dal import SQLCustomType
import zlib

compressed = SQLCustomType(
     type ='text',
     encoder =(lambda x: zlib.compress(x or '')),
     decoder = (lambda x: zlib.decompress(x))

db.define_table('example', Field('data',type=compressed))

SQLCustomType is a field type factory. Its type argument must be one of the standard web2py types. It tells web2py how to treat the field values at the web2py level. native is the name of the field as far as the database is concerned. Allowed names depend on the database engine. encoder is an optional transformation function applied when the data is stored and decoder is the optional reversed transformation function.

This feature is marked as experimental. In practice it has been in web2py for a long time and it works but it can make the code not portable, for example when the native type is database specific. It does not work on Google App Engine NoSQL.

Using DAL without define tables

The DAL can be used from any Python program simply by doing this:

from gluon import DAL, Field
db = DAL('sqlite://storage.sqlite',folder='path/to/app/databases')

i.e. import the DAL, Field, connect and specify the folder which contains the .table files (the app/databases folder).

To access the data and its attributes we still have to define all the tables we are going to access with db.define_tables(...).

If we just need access to the data but not to the web2py table attributes, we get away without re-defining the tables but simply asking web2py to read the necessary info from the metadata in the .table files:

from gluon import DAL, Field
db = DAL('sqlite://storage.sqlite',folder='path/to/app/databases',

This allows us to access any db.table without need to re-define it.

PostGIS, SpatiaLite, and MS Geo (experimental)

Geo Extensions

The DAL supports geographical APIs using PostGIS (for PostgreSQL), spatialite (for SQLite), and MSSQL and Spatial Extensions. This is a feature that was sponsored by the Sahana project and implemented by Denes Lengyel.

DAL provides geometry and geography fields types and the following functions:


st_asgeojson (PostGIS only)
st_simplify (PostGIS only)

Here are some examples:

from gluon.dal import DAL, Field, geoPoint, geoLine, geoPolygon
db = DAL("mssql://user:pass@host:db")
sp = db.define_table('spatial', Field('loc','geometry()'))

Below we insert a point, a line, and a polygon:


Notice that

rows = db(>0).select()

Always returns the geometry data serialized as text. You can also do the same more explicitly using st_astext():

print db(>0).select(, sp.loc.st_astext()),spatial.loc.STAsText()
1, "POINT (1 2)"
2, "LINESTRING (100 100, 20 180, 180 180)"
3, "POLYGON ((0 0, 150 0, 150 150, 0 150, 0 0))"

You can ask for the native representation by using st_asgeojson() (in PostGIS only):

print db(>0).select(, sp.loc.st_asgeojson().with_alias('loc')),loc
1, [1, 2]
2, [[100, 100], [20 180], [180, 180]]
3, [[[0, 0], [150, 0], [150, 150], [0, 150], [0, 0]]]

(notice an array is a point, an array of arrays is a line, and an array of array of arrays is a polygon).

Here are example of how to use geographical functions:

query = sp.loc.st_intersects(geoLine((20,120),(60,160)))
query = sp.loc.st_overlaps(geoPolygon((1,1),(11,1),(11,11),(11,1),(1,1)))
query = sp.loc.st_contains(geoPoint(1,1))
print db(query).select(,sp.loc),spatial.loc
3,"POLYGON ((0 0, 150 0, 150 150, 0 150, 0 0))"

Computed distances can also be retrieved as floating point numbers:

dist = sp.loc.st_distance(geoPoint(-1,2)).with_alias('dist')
print db(>0).select(, dist), dist
1 2.0
2 140.714249456
3 1.0

Copy data from one db into another

Consider the situation in which you have been using the following database:

db = DAL('sqlite://storage.sqlite')

and you wish to move to another database using a different connection string:

db = DAL('postgres://username:password@localhost/mydb')

Before you switch, you want to move the data and rebuild all the metadata for the new database. We assume the new database to exist but we also assume it is empty.

Web2py provides a script that does this work for you:

cd web2py
python scripts/    -f applications/app/databases    -y 'sqlite://storage.sqlite'    -Y 'postgres://username:password@localhost/mydb'

After running the script you can simply switch the connection string in the model and everything should work out of the box. The new data should be there.

This script provides various command line options that allows you to move data from one application to another, move all tables or only some tables, clear the data in the tables. for more info try:

python scripts/ -h

Note on new DAL and adapters

The source code of the Database Abstraction Layer was completely rewritten in 2010. While it stays backward compatible, the rewrite made it more modular and easier to extend. Here we explain the main logic.

The file "gluon/" defines, among other, the following classes.

BaseAdapter extends ConnectionPool

Their use has been explained in the previous sections, except for BaseAdapter. When the methods of a Table or Set object need to communicate with the database they delegate to methods of the adapter the task to generate the SQL and or the function call.

For example:




which delegates the adapter by returning:


Here db.mytable._listify converts the dict of arguments into a list of (field,value) and calls the insert method of the adapter. db._adapter does more or less the following:

query = db._adapter._insert(db.mytable,list_of_fields)

where the first line builds the query and the second executes it.

BaseAdapter define the interface for all adapters.

"gluon/" at the moment of writing this book, contains the following adapters:

SQLiteAdapter extends BaseAdapter
JDBCSQLiteAdapter extends SQLiteAdapter
MySQLAdapter extends BaseAdapter
PostgreSQLAdapter extends BaseAdapter
JDBCPostgreSQLAdapter extends PostgreSQLAdapter
OracleAdapter extends BaseAdapter
MSSQLAdapter extends BaseAdapter
MSSQL2Adapter extends MSSQLAdapter
FireBirdAdapter extends BaseAdapter
FireBirdEmbeddedAdapter extends FireBirdAdapter
InformixAdapter extends BaseAdapter
DB2Adapter extends BaseAdapter
IngresAdapter extends BaseAdapter
IngresUnicodeAdapter extends IngresAdapter
GoogleSQLAdapter extends MySQLAdapter
NoSQLAdapter extends BaseAdapter
GoogleDatastoreAdapter extends NoSQLAdapter
CubridAdapter extends MySQLAdapter (experimental)
TeradataAdapter extends DB2Adapter (experimental)
SAPDBAdapter extends BaseAdapter (experimental)
CouchDBAdapter extends NoSQLAdapter (experimental)
MongoDBAdapter extends NoSQLAdapter (experimental)

which override the behavior of the BaseAdapter.

Each adapter has more or less this structure:

class MySQLAdapter(BaseAdapter):

    # specify a diver to use
    driver = globals().get('pymysql',None)

    # map web2py types into database types
    types = {
        'boolean': 'CHAR(1)',
        'string': 'VARCHAR(%(length)s)',
        'text': 'LONGTEXT',

    # connect to the database using driver
    def __init__(self,db,uri,pool_size=0,folder=None,db_codec ='UTF-8',
                credential_decoder=lambda x:x, driver_args={},
        # parse uri string and store parameters in driver_args
        # define a connection function
        def connect(driver_args=driver_args):
            return self.driver.connect(**driver_args)
        # place it in the pool
        # set optional parameters (after connection)
        self.execute('SET FOREIGN_KEY_CHECKS=1;')
        self.execute("SET sql_mode='NO_BACKSLASH_ESCAPES';")

   # override BaseAdapter methods as needed
   def lastrowid(self,table):
        self.execute('select last_insert_id();')
        return int(self.cursor.fetchone()[0])

Looking at the various adapters as examples should be easy to write new ones.

When db instance is created:

db = DAL('mysql://...')

the prefix in the uri string defines the adapter. The mapping is defined in the following dictionary also in "gluon/":

    'sqlite': SQLiteAdapter,
    'sqlite:memory': SQLiteAdapter,
    'mysql': MySQLAdapter,
    'postgres': PostgreSQLAdapter,
    'oracle': OracleAdapter,
    'mssql': MSSQLAdapter,
    'mssql2': MSSQL2Adapter,
    'db2': DB2Adapter,
    'teradata': TeradataAdapter,
    'informix': InformixAdapter,
    'firebird': FireBirdAdapter,
    'firebird_embedded': FireBirdAdapter,
    'ingres': IngresAdapter,
    'ingresu': IngresUnicodeAdapter,
    'sapdb': SAPDBAdapter,
    'cubrid': CubridAdapter,
    'jdbc:sqlite': JDBCSQLiteAdapter,
    'jdbc:sqlite:memory': JDBCSQLiteAdapter,
    'jdbc:postgres': JDBCPostgreSQLAdapter,
    'gae': GoogleDatastoreAdapter, # discouraged, for backward compatibility
    'google:datastore': GoogleDatastoreAdapter,
    'google:sql': GoogleSQLAdapter,
    'couchdb': CouchDBAdapter,
    'mongodb': MongoDBAdapter,

the uri string is then parsed in more detail by the adapter itself.

For any adapter you can replace the driver with a different one:

from gluon.dal import MySQLAdapter
MySQLAdapter.driver = mysqldb

and you can specify optional driver arguments and adapter arguments:

db =DAL(..., driver_args={}, adapter_args={})


SQLite does not support dropping and altering columns. That means that web2py migrations will work up to a point. If you delete a field from a table, the column will remain in the database but be invisible to web2py. If you decide to reinstate the column, web2py will try re-create it and fail. In this case you must set fake_migrate=True so that metadata is rebuilt without attempting to add the column again. Also, for the same reason, SQLite is not aware of any change of column type. If you insert a number in a string field, it will be stored as string. If you later change the model and replace the type "string" with type "integer", SQLite will continue to keep the number as a string and this may cause problem when you try to extract the data.

MySQL does not support multiple ALTER TABLE within a single transaction. This means that any migration process is broken into multiple commits. If something happens that causes a failure it is possible to break a migration (the web2py metadata are no longer in sync with the actual table structure in the database). This is unfortunate but it can be prevented (migrate one table at the time) or it can be fixed a posteriori (revert the web2py model to what corresponds to the table structure in database, set fake_migrate=True and after the metadata has been rebuilt, set fake_migrate=False and migrate the table again).

Google SQL has the same problems as MySQL and more. In particular table metadata itself must be stored in the database in a table that is not migrated by web2py. This is because Google App Engine has a read-only file system. Web2py migrations in Google:SQL combined with the MySQL issue described above can result in metadata corruption. Again, this can be prevented (my migrating the table at once and then setting migrate=False so that the metadata table is not accessed any more) or it can fixed a posteriori (my accessing the database using the Google dashboard and deleting any corrupted entry from the table called web2py_filesystem.


MSSQL does not support the SQL OFFSET keyword. Therefore the database cannot do pagination. When doing a limitby=(a,b) web2py will fetch the first b rows and discard the first a. This may result in a considerable overhead when compared with other database engines.

Oracle also does not support pagination. It does not support neither the OFFSET nor the LIMIT keywords. Web2py achieves pagination by translating a db(...).select(limitby=(a,b)) into a complex three-way nested select (as suggested by official Oracle documentation). This works for simple select but may break for complex selects involving aliased fields and or joins.

MSSQL has problems with circular references in tables that have ONDELETE CASCADE. This is an MSSSQL bug and you work around it by setting the ondelete attribute for all reference fields to "NO ACTION". You can also do it once and for all before you define tables:

db = DAL('mssql://....')
for key in ['reference','reference FK']:
        '%(on_delete_action)s','NO ACTION')

MSSQL also has problems with arguments passed to the DISTINCT keyword and therefore while this works,


this does not


Google NoSQL (Datastore) does not allow joins, left joins, aggregates, expression, OR involving more than one table, the ‘like’ operator searches in "text" fields. Transactions are limited and not provided automatically by web2py (you need to use the Google API run_in_transaction which you can look up in the Google App Engine documentation online). Google also limits the number of records you can retrieve in each one query (1000 at the time of writing). On the Google datastore record IDs are integer but they are not sequential. While on SQL the "list:string" type is mapped into a "text" type, on the Google Datastore it is mapped into a ListStringProperty. Similarly "list:integer" and "list:reference" are mapped into "ListProperty". This makes searches for content inside these fields types are more efficient on Google NoSQL than on SQL databases.