Ribbit
Ribbit is yet another type safe relational database
library for Haskell, heavily inspired by the amazing
Servant library. The goal
is to create a type-level language for defining table schemas "as a type",
queries that operate on those schemas, and, tangentially, "backends" that
can do something useful with those types like talk to an actual database.
Using Ribbit, you might expect to see something like this:
data PeopleTable
instance Table PeopleTable where
type Name PeopleTable = "people"
type DBSchema PeopleTable =
Field "id" Int
:> Field "name" Text
:> Field "age" Int
type MyQuery = Select '["id", "name"] `From` PeopleTable `Where` "age" `Equals` (?)
matchingPeople <-
query
dbConn
(Proxy :: Proxy MyQuery)
(Only 21) -- argument that fills in the (?) placeholder
:: IO [Only Int :> Only Text]
Status
The status of Ribbit "Very Incomplete". My goal is to take a "depth
first" approach, where every feature added is production ready
before moving on to the next feature. Featured back-ends include
postgresql-simple
at this time.
Current Features
These are the features that are currently implemented.
Basic @Select .. From ..
We support queries of the form:
type MyQuery = Select '["field1", "field2"] `From` MyTable
Cross product
We support queries of the form:
type MyQuery = Select '["t1.field1", "t2.field2"] `From` '[MyTable1 `As` "t1", MyTable2 `As` "t2"]
Conditionals
We support queries of the form:
type MyQuery = Select '["field1", "field2"] `From` MyTable `Where` <condition>
Where <condition>
can include:
a `And` b
: Basic intersection.
a `Or` b
: Basic union.
"field" `Equals` (?)
: Test for equality. This introduces a query parameter that must be supplied at runtime.
"field1" `Equals` "field2"
: Test the equality of two fields (that both must exist in the schema)
a `Lt` b
: Less than operator.
a `Lte` b
: Less than or equal to operator.
a `Gt` b
: Greater than operator.
a `Gte` b
: Greater than or equal to operator.
Not a
: Not operator.
"field" `NotEquals` (?)
: Test for inequality against a query parameter.
"field1" `NotEquals` "field2"
: Test for inequality between two fields.
Limited CREATE TABLE
support.
The postgresql-simple backend supports creating tables in the
database. This support is "limited" because it misses the following
features:
-
PostgreSQL column types are know for only a small number of Haskell types.
This is extensible by the user (by implementing a type class), but it would
be nice to have a more comprehensive set of mappings out of the box.
-
Foreign key constraints are not yet supported.
-
Arbitrary non-primary-key indexes are not yet supported.
What IS supported already is:
-
Determining whether a field is nullable, based on whether the corresponding
Haskell type is wrapped in a Maybe
.
-
Compound primary keys. I.e. primary keys consisting of more than one
component.
INSERT INTO
support.
Basic inserts are supported:
type MyInsert = InsertInto PeopleTable '["id", "name", "age"]
execute
conn
(Proxy :: Proxy MyInsert)
(Only 1 :> Only "Bob Marley" :> Only 36)
DELETE FROM
support.
Basic deletes are supported:
type MyDeleteEveryone = DeleteFrom PeopleTable
type MyDeleteById = DeleteFrom PeopleTable `Where` id `Equals` (?)
execute
conn
(Proxy :: Proxy MyDeleteEveryone)
()
execute
conn
(Proxy :: Proxy MyDeleteById)
(Only 1)
UPDATE TABLE
support.
Basic updates are supported:
execute
conn
(Proxy :: Proxy (
Update PeopleTable
'[
"name", -- each field adds a query parameter so you can
"age" -- provide the new value at run time.
]
`Where` (
"id" `Equals` (?)
)
))
(Only newName :> Only newAge :> Only targetEmployeeId)
Roadmap
This is what I plan to work on next:
Evaluate the usefulness of what I've got so far.
I'm am half pleased with the way the query language turned out, because
someone with understanding of SQL can at least read and modify queries,
even if they would find it hard to construct them from scratch without
examples, but I still feel they could be more ergonomic in at least two ways:
- Better error messages for compiler-guided development.
- Fewer backticks.
I have tried playing around with defining the query language in terms
of value level constructs, which I think can help improve both of these
issues, but I haven't yet gotten that to work without requiring so much
type annotation that it defeats the purpose of ergonomics. I have some
vague idea about how I might solve that with generic instances, but I
am reluctant to do anything that smacks of "magic" without exhausting
all other possibilities first.
Flesh out Haskell to PostgreSQL type mapping.
As mentioned above, only a small number of Haskell types are mapped
to PostgreSQL types out of the box. We would like to make this a more
comprehensive mapping out of the box.
How it compares with other libraries.
The short answer is there are a lot of other libraries and I'm not sure.
Persistent and esquelleto are ones I've used, but if you search "relational" or
"sql" in Hackage there seems to be a lot of other options. Part of the goals
for this library are to flesh out this approach myself, so I can have a better
context for understanding everything else available. In other words, it is part
research project. With that in mind, there are at least a couple of specific
goals I have in mind:
-
Avoid template Haskell. Persistent is amazing, but the use of Template
Haskell makes certain things difficult, like documenting (or for large
projects even understanding) everything that is produced by the Template
Haskell.
-
Make the language easy to understand. If you have some basic SQL knowledge,
it should be immediately obvious what is going on even if you are a beginner
Haskeller.
-
Try to make as much stuff happen at the type level as possible. The ability
to write your own type classes or type families over Servant API types is, I feel,
part of what makes Servant so amazing. I want to replicate that success here.
So, for instance, if someone somewhere defines a schema type that looks like
this:
type MySchema =
Field "id" Int
:> Field "name" Text
:> Field "address" (Maybe Text)
Then you would be free to deconstruct this type (using type families),
transform it into another schema, generate customized CREATE TABLE
statements if the ones provided aren't good enough for your back-end or
use case... that sort of thing. As a somewhat contrived example: maybe,
for who knows what reason, you never want to allow null values in your
database. You can write a type family that can inspect every field in
an arbitrary schema, replacing all the Maybe a
with just a
, like:
-- With -XPolyKinds
type family NoNulls schema where
NoNulls (Field name (Maybe typ)) = Field name typ
NoNulls (a :> b) = NoNulls a :> NoNulls b
NoNulls a = a
NoNulls MySchema
-- Same as:
-- Field "id" Int
-- :> Field "name" Text
-- :> Field "address" Text <--- note the lack of Maybe
The name: Ribbit
The name means nothing except I kindof like the sound of it. There are so many
"sql", "relational", "query", etc. package names already that I didn't want to:
- get lost in the mix.
- step on anyone's toes by choosing too similar a name.
- create confusion by seeming to be associated with some other package with
which I am not.