Skip to content


A beam model is any single-constructor Haskell record type parameterized by a type of kind * -> *. The model must have an instance of Generic, Beamable, and Table. Generic can be derived using the DeriveGeneric extension of GHC. Beamable must be given an empty instance declaration (instance Beamable Tbl for a table of type Tbl). Table is discussed next.

Each field in the record type must either be a sub-table (another parameterized type with a Beamable instance) or an explicit column. A column is specified using the Columnar type family applied to the type's parameter and the underlying Haskell type of the field.

The Table type class

Table is a type class that must be instantiated for all types that you would like to use as a table. It has one associated data instance and one function.

You must create a type to represent the primary key of the table. The primary key of a table Tbl is the associated data type PrimaryKey Tbl. Like Tbl, it takes one type parameter of kind * -> *. It must have only one constructor which can hold all fields in the primary key. The constructor need not be a record constructor (although it can be).

You must also write a function primaryKey that takes an instance of Tbl (parameterized over any functor f) and returns the associated PrimaryKey type. It is sometimes easiest to use the Applicative instance for r -> to write this function. For example, if tblField1 and tblField2 are part of the primary key, you can write

instance Table Tbl where
  data PrimaryKey Tbl f = TblKey (Columnar f ..) (Columnar f ..)
  primaryKey t = TblKey (tblField1 t) (tblField2 t)

more simply as

instance Table Tbl where
  data PrimaryKey Tbl f = TblKey (Columnar f ..) (Columnar f ..)
  primaryKey = TblKey <$> tblField1 <*> tblField2

The Identity trick

Beam table types are commonly suffixed by a T to indicate the name of the generic table type. Usually, a type synonym named by leaving out the T is defined by applying the table to Identity. Recall each field in the table is either another table or an application of Columnar to the type parameter. When the type is parameterized by Identity, every column is also parameterized by Identity.

Columnar is a type family defined such that Columnar Identity x ~ x. Thus, when parameterized over Identity, every field in the table type takes on the underlying Haskell type.

Suppose you have a table type ModelT and a type synonym type Model = ModelT Identity. Notice that deriving Show, Eq, and other standard Haskell type classes won't generally work for ModelT. However, you can use the standalone deriving mechanism to derive these instances for Model.

data ModelT f = Model { .. } deriving (Generic, Beamable)

-- deriving instance Show (ModelT f) -- Won't work because GHC won't get the constraints right

type Model = ModelT Identity
deriving instance Show Model
deriving instance Eq Model
deriving instance Ord Model

Allowed data types

Any data type can be used within a Columnar. Beam does no checking that a field can be used against a particular database when the data type is defined. Instead, type errors will occur when the table is being used as a query. For example, the following is allowed, even though many backends will not work with array data types.

import qualified Data.Vector as V

data ArrayTable f
    = ArrayTable
    { arrayTablePoints :: Columnar f (V.Vector Int32)
    } deriving Generic

You can construct values of type ArrayTable Identity and even write queries over it (relying on type inference to get the constraints right). However, if you attempt to solve the constraints over a database that doesn't support columns of type V.Vector Int32, GHC will throw an error. Thus, it's important to understand the limits of your backend when deciding which types to use. In general, numeric, floating-point, and text types are well supported.

Maybe types

Optional fields (those that allow a SQL NULL) can usually be given a Maybe type. However, you cannot use Maybe around an embedded table (you will be unable to instantiate Beamable).

Beam offers a way around this. Instead of embedding the table applied to the type parameter f, apply it to Nullable f. Columnar (Nullable f) a ~ Maybe (Columnar f a) for all a. Thus, this will make every column in the embedded table take on the corresponding Maybe type.


Nullable will nest Maybes. That is Columnar (Nullable f) (Maybe a) ~ Maybe (Maybe a). This is bad from a SQL perspective, since SQL has no concept of a nested optional type. Beam treats a Nothing at any 'layer' of the Maybe stack as a corresponding SQL NULL. When marshalling data back, a SQL NULL is read in as a top-level Nothing.

The reasons for this misfeature is basically code simplicity. Fixing this is a top priority of future versions of beam.

Column tags

Above, we saw that applying Identity to a table type results in a type whose columns are the underlying Haskell type. Beam uses other column tags for querying and describing databases. Below is a table of common column tags and their meaning.

Converting between tags

Suppose you have a Beamable type parameterized over a tag f and needed one parameterized over a tag g. Given a function conv :: forall a. Columnar f a -> Columnar g a, you can use changeBeamRep to convert between the tables.

There is one caveat however -- since Columnar is a type family, the type of conv is actually ambiguous. We need a way to carry the type of f, g, and a into the code. For this reason, conv must actually be written over the Columnar'(notice the tick) newtype. Columnar' is a newtype defined as such

newtype Columnar' f a = Columnar' (Columnar f a)

Notice that, unlike Columnar (a non-injective type family), Columnar' is a full type. The type of conv' :: forall a. Columnar' f a -> Columnar' g a is now unambiguous. You can easily use conv to implement conv':

conv' (Columnar' a) = Columnar' (conv a)

The Beamable type class

All beam tables, primary keys, and shared data fields must be instances of the Beamable class. You cannot override the methods of Beamable. Rather, they are derived using GHC's generics mechanism. Once you've declared your data type, you can simply write instance Beamable <your-type-name> to instantiate the correct Beamable instance for your type.

The Table type class

All Beamable data types that you want to include as a TableEntity in your database must be members of the Table type class. The Table type class defines one associated type family PrimaryKey and a function primaryKey that takes a table over an arbitrary column tag and produces that table's PrimaryKey. For example, if you have a model

data PersonT f
    = Person
    { personEmail     :: Columnar f Text
    , personFirstName :: Columnar f Text
    , personLastName  :: Columnar f Text
    , personAge       :: Columnar f Int32
    } deriving (Generic, Beamable)

and you want the personEmail field to form the primary key, you would define a Table instance as such

instance Table PersonT where
  data PrimaryKey PersonT f
      = PersonKey (Columnar f Text) deriving (Generic, Beamable)
  primaryKey person = PersonKey <$> personEmail


Many people find it useful to use the Applicative instance for (->) a to write primaryKey. For example, we could have written the above primaryKey person = PersonKey (personFirstName person) (personLastName person) as primaryKey = PersonKey <$> personFirstName <*> personLastName.


Typing Columnar may become tiresome. Database.Beam also exports C as a type alias for Columnar, which may make writing models easier. Since C may cause name clashes, all examples are given using Columnar.

Many also like defining type synonyms for their table and primary key types. For example, for the table PersonT above, a programmer may define.

type Person = PersonT Identity
type PersonKey = PrimaryKey PersonT Identity
deriving instance Show Person; deriving instance Eq Person
deriving instance Show PersonKey; deriving instance Eq PersonKey

By convention, beam table types are suffixed with T to distinguish their type names from the same type parameterized over Identity (the 'regular' Haskell data type).

What about tables without primary keys?

Tables without primary keys are considered bad style. However, sometimes you need to use beam with a schema that you have no control over. To declare a table without a primary key, simply instantiate the Table class and set PrimaryKey tbl to a type with no fields. Then just produce this type in primaryKey.

For example

data BadT f
  = BadT
  { badFirstName :: C f Text
  , badLastName  :: C f Text
  } deriving (Generic, Beamable)
instance Beamable BadT
instance Table BadT where
  data PrimaryKey BadT f = BadNoId
    deriving (Generic, Beamable)
  primaryKey _ = BadNoId

Foreign references

Foreign references are also easily supported in models by simply embedding the PrimaryKey of the referred to table directly in the parent. For example, suppose we want to create a new model representing a post by a user.

data PostT f
    = Post
    { postId       :: Columnar f (SqlSerial Int32)
    , postPostedAt :: Columnar f LocalTime
    , postContent  :: Columnar f Text
    , postPoster   :: PrimaryKey PersonT f
    } deriving (Generic, Beamable)

instance Table PostT where
  data PrimaryKey PostT f
      = PostId (Columnar f (SqlSerial Int32)) deriving (Generic, Beamable)
  primaryKey = PostId . postId

type Post = PostT Identity
type PostId = PrimaryKey PostT Identity
deriving instance Show Post; deriving instance Eq Post
deriving instance Show PostId; deriving instance Eq PostId

Nullable foreign references

Above, any non-bottom value of type PostT Identity must carry a concrete value of PrimaryKey PersonT Identity. Sometimes, you may want to optionally include a foreign key. You can make a foreign key nullable by embedding the primary key and adding the Nullable column tag modifier.

For example, to make the poster optional above.

data PostT f
    = Post
    { postId       :: Columnar f (SqlSerial Int32)
    , postPostedAt :: Columnar f LocalTime
    , postContent  :: Columnar f Text
    , postPoster   :: PrimaryKey PersonT (Nullable f)
    } deriving (Generic, Beamable)

More complicated relationships

This is the extent of beam's support for defining models. Although similar packages in other languages provide support for declaring one-to-many, many-to-one, and many-to-many relationships, beam's focused is providing a direct mapping of relational database concepts to Haskell, not on abstracting away the complexities of database querying. Thus, beam does not use 'lazy-loading' or other tricks that obfuscate performance. Because of this, the bulk of the functionality dealing with different types of relations is found in the querying support, rather than in the model declarations.

Also, notice that beam does not allow you to specify any kind of reference constraints between tables in your data types. This is because references are a property of the database, not a particular table schema. Such relationships can be defined using the beam-migrate package.


Sometimes, we want to declare multiple models with fields in common. Beam allows you to simple embed such fields in common types and embed those directly into models. For example, in the Chinook example schema, we define the following structure for addresses.

data AddressMixin f
  = Address
  { address           :: Columnar f (Maybe Text)
  , addressCity       :: Columnar f (Maybe Text)
  , addressState      :: Columnar f (Maybe Text)
  , addressCountry    :: Columnar f (Maybe Text)
  , addressPostalCode :: Columnar f (Maybe Text)
  } deriving (Generic, Beamable)
type Address = AddressMixin Identity
deriving instance Show (AddressMixin Identity)

We can then use AddressMixin in our models.

data EmployeeT f
  = Employee
  { employeeId        :: Columnar f Int32
  , employeeLastName  :: Columnar f Text
  , employeeFirstName :: Columnar f Text
  , employeeTitle     :: Columnar f (Maybe Text)
  , employeeReportsTo :: PrimaryKey EmployeeT (Nullable f)
  , employeeBirthDate :: Columnar f (Maybe LocalTime)
  , employeeHireDate  :: Columnar f (Maybe LocalTime)
  , employeeAddress   :: AddressMixin f
  , employeePhone     :: Columnar f (Maybe Text)
  , employeeFax       :: Columnar f (Maybe Text)
  , employeeEmail     :: Columnar f (Maybe Text)
  } deriving (Generic, Beamable)
-- ...
data CustomerT f
  = Customer
  { customerId        :: Columnar f Int32
  , customerFirstName :: Columnar f Text
  , customerLastName  :: Columnar f Text
  , customerCompany   :: Columnar f (Maybe Text)
  , customerAddress   :: AddressMixin f
  , customerPhone     :: Columnar f (Maybe Text)
  , customerFax       :: Columnar f (Maybe Text)
  , customerEmail     :: Columnar f Text
  , customerSupportRep :: PrimaryKey EmployeeT (Nullable f)
  } deriving (Generic, Beamable)


Based on your data type declarations, beam can already guess a lot about your tables. For example, it already assumes that the personFirstName field is accessible in SQL as first_name. This defaulting behavior makes it very easy to interact with typical databases.

For the easiest user experience, it's best to follow beam's conventions for declaring models. In particular, the defaulting mechanisms rely on each table type declaring only one constructor which has fields named in the camelCase style.

When defaulting the name of a table field or column, beam un-camelCases the field name (after dropping leading underscores) and drops the first word. The remaining words are joined with underscores. If there is only one component, it is not dropped. Trailing and internal underscores are preserved in the name and if the name consists solely of underscores, beam makes no changes. A summary of these rules is given in the table below.

Haskell field name Beam defaulted column name
personFirstName first_name
_personLastName last_name
name name
first_name first_name
_first_name first_name
___ (three underscores) ___ (no changes)

Note that beam only uses lower case in field names. While typically case does not matter for SQL queries, beam always quotes identifiers. Many DBMS's are case-sensitive for quoted identifiers. Thus, queries can sometimes fail if your tables use mixtures of lower- and upper-case to distinguish between fields.

For information on modifying the defaults, see the next section.