3. Connecting to a Database
Alright, let’s get going.
In this chapter we start from the beginning. First we write a program that connects to a database and returns a value, and then we run that program in the REPL. We also touch on composing small programs to construct larger ones.
Our First Program
Before we can use doobie we need to import some symbols. We will use the doobie.imports
module here as a convenience; it exposes the most commonly-used symbols when working with the high-level API. We will also import the
scalaz core, as well as Task
from scalaz-concurrent.
import doobie.imports._
import scalaz._, Scalaz._
In the doobie high level API the most common types we will deal with have the form ConnectionIO[A]
, specifying computations that take place in a context where a java.sql.Connection
is available, ultimately producing a value of type A
.
So let’s start with a ConnectionIO
program that simply returns a constant.
scala> val program1: ConnectionIO[Int] = 42.pure[ConnectionIO]
program1: doobie.imports.ConnectionIO[Int] = Return(42)
This is a perfectly respectable doobie program, but we can’t run it as-is; we need a Connection
first. There are several ways to do this, but here let’s use a Transactor
.
val xa = DriverManagerTransactor[IOLite](
"org.postgresql.Driver", "jdbc:postgresql:world", "postgres", ""
)
A Transactor
is simply a structure that knows how to connect to a database, hand out connections, and clean them up; and with this knowledge it can transform ConnectionIO ~> IOLite
, which gives us something we can run. Specifically it gives us an IOLite
that, when run, will connect to the database and run our program in a single transaction.
Scala does not have a standard IO, so the examples in this book use the simple IOLite
data type provided by doobie. This type is not very feature-rich but is safe and performant and fine to use. Similar monadic types like scalaz.effect.IO
, scalaz.concurrent.Task
, fs2.Task
, and monix.Task
will also work fine.
In fact, you can use any Monad M[_]
as long as there is a Catchable[M]
and Capture[M]
available. See Using Your Own Target Monad at the end of this capter for more details.
The DriverManagerTransactor
simply delegates to the java.sql.DriverManager
to allocate connections, which is fine for development but inefficient for production use. In a later chapter we discuss other approaches for connection management.
Right, so let’s do this.
scala> val task: IOLite[Int] = program1.transact(xa)
task: doobie.imports.IOLite[Int] = doobie.util.iolite$IOLite$$anon$4@531423d
scala> task.unsafePerformIO
res0: Int = 42
Hooray! We have computed a constant. It’s not very interesting because we never ask the database to perform any work, but it’s a first step.
Keep in mind that all the code in this book is pure except the calls to
IOLite.unsafePerformIO
, which is the “end of the world” operation that typically appears only at your application’s entry points. In the REPL we use it to force a computation to “happen”.
Right. Now let’s try something more interesting.
Our Second Program
Let’s use the sql
string interpolator to construct a query that asks the database to compute a constant. We will cover this construction in great detail later on, but the meaning of program2
is “run the query, interpret the resultset as a stream of Int
values, and yield its one and only element.”
scala> val program2: ConnectionIO[Int] = sql"select 42".query[Int].unique
program2: doobie.imports.ConnectionIO[Int] = Gosub(Gosub(Suspend(PrepareStatement(select 42)),<function1>),<function1>)
scala> val task2: IOLite[Int] = program2.transact(xa)
task2: doobie.imports.IOLite[Int] = doobie.util.iolite$IOLite$$anon$4@17eefbf0
scala> task2.unsafePerformIO
res1: Int = 42
Ok! We have now connected to a database to compute a constant. Considerably more impressive.
Our Third Program
What if we want to do more than one thing in a transaction? Easy! ConnectionIO
is a monad, so we can use a for
comprehension to compose two smaller programs into one larger program.
val program3: ConnectionIO[(Int, Double)] =
for {
a <- sql"select 42".query[Int].unique
b <- sql"select random()".query[Double].unique
} yield (a, b)
And behold!
scala> program3.transact(xa).unsafePerformIO
res2: (Int, Double) = (42,0.9716776320710778)
The astute among you will note that we don’t actually need a monad to do this; an applicative functor is all we need here. So we could also write program3
as:
val program3a = {
val a: ConnectionIO[Int] = sql"select 42".query[Int].unique
val b: ConnectionIO[Double] = sql"select random()".query[Double].unique
(a |@| b).tupled
}
And lo, it was good:
scala> program3a.transact(xa).unsafePerformIO
res3: (Int, Double) = (42,0.9884093310683966)
And of course this composition can continue indefinitely.
scala> program3a.replicateM(5).transact(xa).unsafePerformIO.foreach(println)
(42,0.2739106295630336)
(42,0.5807683779858053)
(42,0.254662926774472)
(42,0.386529924813658)
(42,0.4502082602120936)
Diving Deeper
You do not need to know this, but if you’re a scalaz user you might find it helpful.
All of the doobie monads are implemented via Free
and have no operational semantics; we can only “run” a doobie program by transforming FooIO
(for some carrier type java.sql.Foo
) to a monad that actually has some meaning.
Out of the box all of the doobie provides an interpreter from its free monads to Kleisli[M, Foo, ?]
given Monad[M]
, Catchable[M]
, and Capture[M]
(we will discuss Capture
shortly, standby).
scala> import doobie.free.connection.ConnectionOp
import doobie.free.connection.ConnectionOp
scala> import java.sql.Connection
import java.sql.Connection
scala> import scalaz.~>
import scalaz.$tilde$greater
scala> val interpreter: ConnectionOp ~> Kleisli[IOLite, Connection, ?] = KleisliInterpreter[IOLite].ConnectionInterpreter
interpreter: scalaz.~>[doobie.free.connection.ConnectionOp,[γ$0$]scalaz.Kleisli[[A]doobie.util.iolite.IOLite[A],java.sql.Connection,γ$0$]] = doobie.free.KleisliInterpreter$$anon$10@63dbddf5
scala> val kleisli: Kleisli[IOLite, Connection, Int] = program1.foldMap(interpreter)
kleisli: scalaz.Kleisli[doobie.imports.IOLite,java.sql.Connection,Int] = Kleisli(<function1>)
scala> // >>= is simply flatMap and kleisli.run is (Connection) => IOLite[Int]
| val task: IOLite[Int] = IOLite.primitive(null: java.sql.Connection) >>= kleisli.run
task: doobie.imports.IOLite[Int] = doobie.util.iolite$IOLite$$anon$4@7eba6a5a
scala> task.unsafePerformIO // sneaky; program1 never looks at the connection
res6: Int = 42
So the interpreter above is used to transform a ConnectionIO[A]
program into a Kleisli[IOLite, Connection, A]
. Then we construct an IOLite[Connection]
(returning null
) and bind it through the Kleisli
, yielding our IOLite[Int]
. This of course only works because program1
is a pure value that does not look at the connection.
The Transactor
that we defined at the beginning of this chapter is basically a utility that allows us to do the same as above using program1.transact(xa)
.
There is a bit more going on when calling transact
(we add commit/rollback handling and ensure that the connection is closed in all cases) but fundamentally it’s just a natural transformation and a bind.
The Capture Typeclass
Currently scalaz has no typeclass for monads with effect-capturing unit, so that’s all Capture
does; it’s simply (=> A) => M[A]
that is referentially transparent for all expressions, even those with side-effects. This allows us to sequence the same effect multiple times in the same program. This is exactly the behavior you expect from IO
for example.
doobie provides Capture
instances for Task
and IO
, and the implementations are simply delay
and apply
, respectively.
Note that
scala.concurrent.Future
does not have an effect-capturing constructor and thus cannot be used as a target type for doobie programs. AlthoughFuture
is very commonly used for side-effecting operations, doing so is not referentially transparent.Future
has nothing at all to say about side-effects. It is well-behaved in a functional sense only for pure computations.
Using Your Own Target Monad
As mentioned earlier, you can use any monad M[_]
when using a Transactor
as long as there is a scalaz.Catchable[M]
and doobie.util.capture.Capture[M]
available. For example, for monix.eval.Task
, you could have the following:
import doobie.util.capture.Capture
import monix.eval.Task
import scala.util.{Failure, Success}
import scalaz.{Catchable, \/, \/-, -\/}
implicit object monixTaskCatchable extends Catchable[Task] with Capture[Task] {
def apply[A](a: => A): Task[A] =
Task.delay(a)
def pure[A](a: A): Task[A] =
Task.pure(a)
def flatMap[A, B](a: Task[A])(f: A => Task[B]): Task[B] =
a.flatMap(f)
def fail[A](err: Throwable): Task[A] =
Task.raiseError(err)
def attempt[A](fa: Task[A]): Task[Throwable \/ A] =
fa.materialize.map({
case Success(v) => \/-(v)
case Failure(err) => -\/(err)
})
def suspend[A](fa: => Task[A]): Task[A] =
Task.suspend(fa)
}