As of October 2023, there are around a dozen mature PostgreSQL libraries, all levels of abstractions, from low-level queries to the type level. More than enough for anybody. (Shout out to everyone who says Haskell has no libraries.)
Which one to use? Let’s see. Foreach library, we’ll talk about features and basics like writing queries, complexity, pitfalls, (everyone’s favorite topic) errors, and so on.
We assume you are familiar with the fundamentals of SQL and PostgreSQL.
Before integrating with the database, let’s discuss the data and the setup.
💡 If you want to follow allow at home, the repository contains all the imports and data types — we omit most of them from the tutorial for simplicity.
Let’s imagine that we’re building a (tiny) warehouse management system:
- Warehouse has multiple products, some quantity of each.
- Product has a label and a description.
- Product can belong to categories (many-to-many).
- Category has a label.
The scripts/create_tables.sql contains all the definitions:
CREATE TABLE product (
id SERIAL PRIMARY KEY,
label TEXT NOT NULL,
description TEXT,
UNIQUE (label)
);
CREATE TABLE category (
id SERIAL PRIMARY KEY,
label TEXT NOT NULL,
UNIQUE (label)
);
CREATE TABLE product_category (
category_id INT NOT NULL,
product_id INT NOT NULL,
PRIMARY KEY (category_id, product_id),
FOREIGN KEY (product_id) REFERENCES product(id),
FOREIGN KEY (category_id) REFERENCES category(id)
);
CREATE TABLE warehouse (
id SERIAL PRIMARY KEY,
product_id INT NOT NULL,
quantity INT NOT NULL,
created TIMESTAMP,
modified TIMESTAMP,
FOREIGN KEY (product_id) REFERENCES product(id)
);
Here are a few things you need to know to play along at home.
🐘 If you don’t care, don’t like docker, or already have an easily-accessible postgres server, feel free to skip this section.
First, install docker.
docker compose up (docker-compose up on older versions) starts PostgreSQL (see docker-compose.yml) and initializes the databases and tables (using the scripts/create_tables.sql). It mounts the data to the postgres-data/ (in case you need to wipe it or something).
Some of the hardcoded things that the code relies on:
environment:
- POSTGRES_DB=warehouse
- POSTGRES_USER=postgres
- POSTGRES_PASSWORD=password
ports:
- 5432:5432
💡 These hardcoded values are also hardcoded in the code in
Hardcoded.hs(for laziness reasons)
You can connect to the container and run arbitrary queries using psql:
docker exec -it elephants-postgres-1 psql -U postgres -d warehouseNote: elephants-postgres-1 is a container name, which might be different for you; check with docker ps to get the correct container id (or name). We also pass a couple of flags: -U postgres for the user name and -d warehouse for the database name.
docker compose down to stop and remove the containers.
If you have stack installed, stack build to build and stack run to run.
💡 We use lts-21.7 (ghc-9.4.5), published on 2023-08-14.
🤷 To build the backend, you might need the
libpqdevelopment libraries installed (e.g.,libpq-devon Debian-based distributions).
Note that most of the libraries rely on using various extensions. Here is a quick overview of the most important ones.
OverloadedStrings
Used to simplify the construction of query values — we can use literal strings, like "SELECT * FROM user", instead of manually constructing the whole type; for example, Query . toByteString . stringUtf8 $ "SELECT * FROM user" (see Query in postgresql-simple).
TemplateHaskell
Template Haskell (TH) is fairly useful for generating boilerplate code. Some libraries provide the TH machinery to derive typeclass instances and/or generate custom type-safe data types at the compile time.
DeriveAnyClass and DeriveGeneric enable alternative ways to get free typeclass instances.
We’ll use DerivingStrategies to make the derivation explicit, for example:
data Category = Category {label :: Text}
deriving (Show, Generic) -- Derive Generic instance,
deriving anyclass (ToRow, FromRow) -- used for these instancesQuasiQuotes
Some libraries (e.g., postgresql-simple) provide quasi quoters for less tedious sql construction in Haskell code. For example:
[sql| select label from product |]DuplicateRecordFields
Only one postgres library requires it. But it’s a stylistic choice — we use this extension because we have multiple data types with a label field.
OverloadedRecordDot
To access record fields as well as specific columns of the tables (for example, product.label)
Note that the project uses GHC2021, which includes DeriveGeneric, TypeApplications, and many other extensions required for some libraries to work. Also, for the tutorial's sake, we’ll use the included ScopedTypeVariables to demonstrate some intermediate types.
Let’s start simple. postgresql-simple describes itself as “Mid-Level PostgreSQL client library“.
In other words, we’ll write raw sql queries, and the library will deal with security and stuff.
To get started, we add postgresql-simple to dependencies. We’re using v0.7.0.0 published in 2023.
We use connect to acquire a connection. It accepts ConnectInfo, which we can get by using defaultConnectInfo and overriding some defaults.
getConnection :: IO Connection
getConnection =
connect $
defaultConnectInfo
{ connectHost = Hardcoded.host
, connectDatabase = Hardcoded.database
, connectUser = Hardcoded.user
, connectPassword = Hardcoded.password
}Eventually, we have to close connection. But you will probably not need to do it manually because you can use withConnect, bracket, or (better) a connection pool.
The library doesn’t support pools, but you can use the resource-pool package (or something similar).
We use execute and execute_ to insert, update, and delete data.
The version with the _ suffix is simpler — it doesn’t perform any query substitutions. We can use it with hardcoded values or with straightforward queries such as truncating the tables:
cleanUp :: Connection -> IO ()
cleanUp connection =
void $ execute_ connection "truncate warehouse, product_category, product, category"Both execute functions return the number of affected rows, which isn’t relevant in case of truncate (it’s 0) and safe to ignore (void ignores the result of evaluation).
We can use execute to make a proper insert and pass some values for substitutions. The simplest way is to pass a tuple:
insert1 <-
execute
connection
"insert into product (label, description) values (?, ?)"
("Wood Screw Kit 1" :: Text, "245-pieces" :: Text)Sometimes, we must be explicit about types; for example, in cases like this, when we use string literals with OverloadedStrings or numeric literals (like 245).
Because there is no tuple of 1, the library provides a custom type Only:
insert2 <-
execute
connection
"insert into product (label) values (?)"
(Only "Wood Screw Kit 2" :: Only Text)Alternatively, we can use lists for any number of values:
insert3 <-
execute
connection
"insert into product (label) values (?)"
["Wood Screw Kit 3" :: Text]But preferable, we use dedicated types:
execute
connection
"insert into product (label, description) values (?, ?)"
(BasicProduct "Wood Screw Kit 4" (Just "245-pieces"))A record can be turned into a list of substitutions via the ToRow typeclass, which is derivable using GHC generics:
data BasicProduct = BasicProduct {label :: Text, description :: Maybe Text}
deriving (Show, Generic)
deriving anyclass (ToRow, FromRow)If we want to modify multiple rows, we can use executeMany:
insert5 <-
executeMany
connection
insert into category (label) values (?)"
[Category "Screws", Category "Wood Screws", Category "Concrete Screws"]The execute functions can’t return any results (other than the number of affected rows), so we have to use the query functions.
Similar to execute_, query_ takes a query with no substitutes:
query1 :: [(Int64, Text, Maybe Text)] <-
query_ connection "select id, label, description from product"Note that we must be explicit about return types — the library can’t guess what we want. In this case, we expect an id of type Int64 that corresponds to Serial, required Text label, and optional description.
We can specify a record return type if we derive FromRow (recall ToRow from the previous section). For example, let’s get a BasicProduct list by label using query:
query2 :: [BasicProduct] <-
query
connection
"select label, description from product where label = ? "
(Only "Wood Screw Kit 2" :: Only Text)If we want to use the in-clause, the library provides a dedicated wrapper:
query3 :: [BasicProduct] <-
query connection "select label, description from product where label in ?" $
Only (In ["Wood Screw Kit 2" :: Text, "Wood Screw Kit 3"])Imagine we want to atomically insert a new listing: product, category, and quantity. This touches multiple tables and requires a transaction. Additionally, because we have a many-to-many relationship, we must first insert the product and category and then use their new ids to create a mapping.
We can use returning to get ids of created rows:
productIds :: [Only Int64] <-
query
connection
"insert into product (label, description) values (?, ?) returning id"
(BasicProduct "Drywall Screws Set" (Just "8000pcs"))
categoryIds :: [Only Int64] <-
query
connection
"insert into category (label) values (?) returning id"
(Category "Drywall Screws")Note that we must use query and not execute because these queries return results.
We can use withTransaction to wrap multiple queries in a transaction:
withTransaction connection $ do
productIds :: [Only Int64] <- query ...
categoryIds :: [Only Int64] <- query ...
void $ case (productIds, categoryIds) of
([Only productId], [Only categoryId]) -> do
_ <-
execute
connection
"insert into warehouse (product_id, quantity, created, modified) values (?, ?, now(), now())"
(productId, 10 :: Int)
execute
connection
"insert into product_category (category_id, product_id) values (?, ?)"
(categoryId, productId)
_ ->
throwIO $ userError "Failed to insert product/category"Any error will rollback the transaction (and the exception will be rethrown). In this example, we throw an explicit error if we don’t get the expected ids for some reason.
Note that in case of a sql error, the exception will not only rollback the transaction but, if uncaught, will propagate further (killing everything on its way and potentially crashing the whole app). So, we should (at least) wrap transactions in the exception handler(s); we’ll see how to do this later.
When you need to, you can also use granular transaction functions: begin, commit, and rollback.
To read all these tables at once, we need to query using a few joins. The library provides a quasi-quoter that makes writing big queries easier — we can format the query and not worry about whitespaces:
result :: [Listing] <-
query
connection
[sql|
select
w.quantity,
p.label,
p.description,
c.label
from warehouse as w
inner join product as p on w.product_id = p.id
left outer join product_category as pc on p.id = pc.product_id
left outer join category as c on c.id = pc.category_id
where w.quantity > (?)|]
[3 :: Int]In postgresql-simple, all the programmer errors (in sql or library usage) are (runtime) exceptions.
If the query string is not formatted correctly, we get FormatError. For instance, if we have a mismatching number of substitutions (? and actual values):
execute
connection
"INSERT INTO category (label) VALUES (?)"
("One" :: Text, "Two" :: Text)
FormatError {fmtMessage = "1 single '?' characters, but 2 parameters", fmtQuery = "INSERT INTO category (label) VALUES (?)", fmtParams = ["One","Two"]}
Similarly, on the return side, if the number of columns doesn’t match the number of elements in the result type (in a list, a tuple, or a record), we get ResultError. The most likely variants are Incompatible and UnexpectedNull.
If we forget to wrap a nullable type on the Haskell side, we get UnexpectedNull. For instance, if we try to get description (which is nullable) as Text and not Maybe Text:
let result :: IO [(Text, Text)] = query_ connection "select label, description from product"
UnexpectedNull {errSQLType = "text", errSQLTableOid = Just (Oid 16386), errSQLField = "description", errHaskellType = "Text", errMessage = ""}
If we mistype the types, we get Incompatible. For instance, if we try to parse just id into BasicProduct:
let result :: IO [BasicProduct] = query_ connection "select id from product"
Incompatible {errSQLType = "int4", errSQLTableOid = Just (Oid 16386), errSQLField = "id", errHaskellType = "Text", errMessage = "types incompatible"}
On top of that, if we misuse the library — by mistaking query for execute or vice verse — we get QueryError. For example, if we use execute with insert query that has returning:
execute_
connection
"INSERT INTO category (label) VALUES (Screws) returning id"
QueryError {qeMessage = "execute resulted in Col 1-column result", qeQuery = "INSERT INTO category (label) VALUES ('Deck Screws') returning id"}
And last but not least, any sql errors from postgres, will come back as SqlError:
let result :: IO [BasicProduct] = query_ connection "select I have no idea what I'm doing"
Wrong sql: SqlError {sqlState = "42601", sqlExecStatus = FatalError, sqlErrorMsg = "syntax error at or near \"no\"", sqlErrorDetail = "", sqlErrorHint = ""}
The errors are pretty good but still not the most descriptive — if you try to write big queries, you have to concentrate on projecting the error information to the query.
The docs are also simple; the library covers all the primary blocks, describes the functions, and provides some examples. Outside, a few blog posts cover similar things, mainly targeting beginners.
And you don’t need more than that — if you know how to write one simple query, you know how to write them all.
The library has a companion package, postgresql-migration.
🗂️ This is a fork of the archived
postgresql-simple-migration.
postgresql-simple is a library for all levels and a great option if you love writing sql by hand and don’t need reusability.
It doesn’t parse or validate the queries, so we must pay attention to what we write: sql queries, haskell types (type-safety is our responsibility), the order of parameters, and so on.
The next “obvious” step is to add more type-safety.
According to the readme, Hasql “is a highly efficient PostgreSQL driver for Haskell with a typesafe yet flexible mapping API; it is production-ready, actively maintained, and the API is pretty stable. It's used by many companies and most notably by the Postgrest project.“
Hasql is an ecosystem of libraries. To keep it simple, let’s limit ourselves to core hasql, hasql-transaction, and hasql-th. We’re using hasql 1.6.3.2 published in 2023.
We’ll also use contravariant-extras, vector, profunctors, and tuple packages to make a few things tidier (this isn’t required; it’s all copy-paste anyway).
💡 (It’s not very important, but) We assume you’ve seen the part on
postgresql-simple, which covers the same topics but at a slower pace.
First, we get a connection:
Right connection <- getConnectiongetConnection :: IO (Either ConnectionError Connection)
getConnection =
acquire $ settings Hardcoded.host Hardcoded.portNumber Hardcoded.user Hardcoded.password Hardcoded.databaseNote the Either. But for now, let’s just pattern-match and not worry about possible errors…
In reality/production, we should probably use hasql-pool to work with a pool of connections.
Let’s see the leading players through the clean-up query:
cleanUp :: Connection -> IO (Either QueryError ())
cleanUp connection = run cleanUpSession connection
where
cleanUpSession :: Session ()
cleanUpSession = statement () cleanUpStatement
cleanUpStatement :: Statement () ()
cleanUpStatement = Statement rawSql E.noParams D.noResult True
rawSql = "truncate warehouse, product_category, product, category"Sessionis a batch of actions to be executed in the context of a connection (a query).Statementis a specification of a strictly single-statement query, which can be parameterized and prepared (how to make a query).Statementconsists of SQL template, params encoder, result decoder, and a flag that determines whether it’s prepared.statementcreates aSessionfrom aStatementand input parameters.runexecutes a bunch of commands (statements) on the provided connection.
💡 Remember that you can see the complete code in the repo.
We have a simple query with no parameters and no result — we don’t need to encode or decode anything. That’s what E.noParams D.noResult for. If we want to pass parameters, we need to supply a decoder.
The first option, is tuples of primitive types and manually written decoders:
insertProductSql = "insert into product (label, description) values ($1, $2)"insertProduct1 :: Statement (Text, Maybe Text) Int64
insertProduct1 = Statement insertProductSql rawParams D.rowsAffected True
rawParams =
(fst >$< E.param (E.nonNullable E.text))
<> (snd >$< E.param (E.nullable E.text))statement ("Wood Screw Kit 1", Just "245-pieces") insertProduct1rawParams is the encoder for our parameters. We use contramap operator (>$<) and append (<>) to compose multiple parameters. D.rowsAffected is the decoder for the result when we want to return the number of affected rows.
💡 Instead of
fstandsnd, you can use thecontrazipfamily of functions from thecontravariant-extraspackage to reduce boilerplate.
Another option, is using records:
insertProduct2 :: Statement BasicProduct Int64
insertProduct2 = Statement insertProductSql basicProductParams D.rowsAffected True
basicProductParams :: E.Params BasicProduct
basicProductParams =
((.label) >$< E.param (E.nonNullable E.text))
<> ((.description) >$< E.param (E.nullable E.text))statement (BasicProduct "Wood Screw Kit 2" Nothing) insertProduct2If we want to modify multiple rows, we have to use the postgres unnest function:
insertManyCategories :: Statement (Vector Category) Int64
insertManyCategories = Statement insertManyCategoriesSql categoryParams D.rowsAffected True
insertManyCategoriesSql = "insert into category (label) select * from unnest ($1)"
categoryParams :: E.Params (Vector Category)
categoryParams =
E.param
$ E.nonNullable
$ E.array
$ E.dimension List.foldl'
$ categoryArray
categoryArray :: E.Array Category
categoryArray = (.label) >$< (E.element $ E.nonNullable E.text)categoryParams is an encoder that allows us to pass a vector of categories to insert.
let categories = [Category "Screws", Category "Wood Screws", Category "Concrete Screws"]
statement (fromList categories) insertManyCategoriesNote that
unnestis more efficient than executing a single-row insert statement multiple times.
Querying data is similar:
session1 :: Session [(Int64, Text, Maybe Text)]
session1 =
statement ()
$ Statement
"select id, label, description from product"
E.noParams
decoder1
True
decoder1 =
D.rowList
$ (,,)
<$> D.column (D.nonNullable D.int8)
<*> D.column (D.nonNullable D.text)
<*> D.column (D.nullable D.text)We need to provide a decoder for the result (to specify how each row results maps into the expected type). If this sounds tedious, we can ask Template Haskell to do the work for us:
In this case, we use singletonStatement that expects one result. There are other variants that we’ll see later.
session2 :: Session (Text, Maybe Text)
session2 = statement () statement2
statement2 :: Statement () (Text, Maybe Text)
statement2 =
[singletonStatement|
select label :: text, description :: text? from product limit 1
|]We write the query and specify the types, hasql-th handles the codecs for us.
But we still need to handle the conversions if we use custom types instead of tuples. The result of the statement has a Profunctor instance, which allows us to modify (input) parameters and (output) results. In other words, we use lmap to map parameters, rmap — result, and dimap — both. For example, let’s return BasicProduct:
session3 :: Session (Maybe BasicProduct)
session3 = statement "Wood Screw Kit 2" statement3
statement3 :: Statement Text (Maybe BasicProduct)
statement3 =
rmap
(fmap (uncurryN BasicProduct))
[maybeStatement|
select label :: text, description :: text?
from product
where label = $1 :: text
|]💡 (fmap (uncurryN BasicProduct)) is a concise way to write the following (using tuples package):
(\result -> fmap (\(a, b) -> (BasicProduct a b)) result)hasql doesn’t have "special support" for an array as a parameter for the IN operator, we should use Any:
session4 :: Session (Vector BasicProduct)
session4 = statement (fromList ["Wood Screw Kit 1", "Wood Screw Kit 2"]) statement4
statement4 :: Statement (Vector Text) (Vector BasicProduct)
statement4 =
rmap
(fmap (uncurryN BasicProduct))
[vectorStatement|
select label :: text, description :: text?
from product
where label = ANY($1 :: text[])
|]We can use returning to get ids of created rows:
insertProduct :: Statement (Text, Maybe Text) Int64
insertProduct =
[singletonStatement|
insert into product (label, description) values ($1 :: text, $2 :: text?) returning id :: int8
|]To wrap multiple queries in a transaction, we can use hasql-transaction. First, we compose the statements:
insertAll :: FullProduct -> Transaction Int64
insertAll listing = do
productId <- Transaction.statement (listing.label, listing.description) insertProduct
categoryId <- Transaction.statement listing.category insertCategory
_ <- Transaction.statement (productId) insertListing
ids <- Transaction.statement (productId, categoryId) insertMapping
pure ids
insertProduct :: Statement (Text, Maybe Text) Int64
insertCategory :: Statement Text Int64
insertListing :: Statement Int64 ()
insertMapping :: Statement (Int64, Int64) Int64Then we run the transaction using the relevant isolation level and mode:
insertWithTransaction :: Connection -> IO ()
insertWithTransaction connection = do
let listing = FullProduct "Drywall Screws Set" (Just "8000pcs") "Drywall Screws"
mapping <- run (transaction Serializable Write $ insertAll listing) connection
putStrLn $ "Insert with transaction: " <> show mappingWe can query these tables using a few joins. There should be nothing unexpected here:
listings :: Statement Int32 (Vector Listing)
listings =
rmap
(fmap (uncurryN Listing))
[vectorStatement|
select
w.quantity :: int,
p.label :: text,
p.description :: text?,
c.label :: text?
from warehouse as w
inner join product as p on w.product_id = p.id
left outer join product_category as pc on p.id = pc.product_id
left outer join category as c on c.id = pc.category_id
where w.quantity > $1 :: int4
|]We’ve been neglecting this until now, but all error reporting is explicit and is presented using Either.
💡 Just a reminder, don’t ignore errors. And don’t pattern match only on Right, or you will end up with this:
user error (Pattern match failure in 'do' block at ...)The other good thing is that the hasql-th parser is pretty good at error reporting and catching typos at compile time (and most of the time, it’s more explicit than postgres’ syntax error at or near). This won’t compile:
[singletonStatement|
select I have no idea what I'm doing
|]The library doesn’t accept (doesn’t compile) if you forget to specify one of the types. For instance, if we omit type of label, we get a somewhat generic error:
[singletonStatement|
select label, description :: text? from product
|]
Result expression is missing a typecast
This ensures that most input and result type (including nullability) mismatches are caught in the compile time. For example, if we forget an input type and return a wrong result type:
statement :: Statement () (Text, Int32)
statement =
[singletonStatement|
select label :: text, description :: text?
-- ^^^^^
-- Couldn't match type ‘Int32’ with ‘Maybe Text’
from product where label = $1 :: text
-- ^^^^
-- Couldn't match type ‘Text’ with ‘()’
|]However, we’re not safe from programming errors. We should use correct statement functions not to get a runtime error. For example, if we use singletonStatement for statements that might not return a result (instead of maybeStatement):
do
failure <- run (statement () failsBecauseNoResults) connection
putStrLn $ "Wrong statement function: " <> show failure
where
failsBecauseNoResults :: Statement () (Text)
failsBecauseNoResults =
[singletonStatement|
select label :: text from product where 1 = 0
|]
Wrong statement function: Left (QueryError "SELECT label :: text FROM product WHERE 1 = 0" [] (ResultError (UnexpectedAmountOfRows 0)))
Or if we use singletonStatement with () result (instead of resultlessStatement):
do
failure <- run (statement () failsBecauseResultless) connection
putStrLn $ "Wrong statement function: " <> show failure
where
failsBecauseResultless :: Statement () ()
failsBecauseResultless =
[singletonStatement|
insert into product (label) values ('this insert fails')
|]
Wrong statement function: Left (QueryError "INSERT INTO product (label) VALUES ('this insert fails')" [] (ResultError (UnexpectedResult "Unexpected result status: CommandOk")))
In case of runtime sql error, for instance, if we violate a constraint, we get a similar error:
inserProduct :: Statement Text ()
inserProduct =
[singletonStatement|
insert into product (label) values ($1 :: text)
|]run (statement "Duplicate screw" inserProduct) connection
>> run (statement "Duplicate screw" inserProduct) connection
Wrong statement function (Left): QueryError "INSERT INTO product (label) VALUES ('Duplicate')" [] (ResultError (ServerError "23505" "duplicate key value violates unique constraint \"product_label_key\"" (Just "Key (label)=(Duplicate screw) already exists.") Nothing Nothing))
Core readme has a good overview and example. The library has simple docs, a couple of tutorials, and talks from the author.
hasql-migrations tool is a port of postgresql-simple-migration for use with hasql.
Overall, hasql is a great choice for writing raw sql queries with more type safety and compile-time syntax checks. The ecosystem comes with other whistles like connection pools and transactions.
The TemplateHaskell module and compile-time checks are optional — if you want, you can deal with the encoders and decoders yourself.
The library requires basic/intermediate knowledge of Haskell and ecosystems. To be comfortable and productive, you must be familiar with vectors, contravariant functors, etc. Other than that, the library is relatively straightforward.
If that was not enough, it’s time to move to the type level.
According to the readme, ”Persistent's goal is to catch every possible error at compile-time, and it comes close to that. It is also designed to be adaptable to any datastore”. As a result, ”a major limitation for SQL databases is that the persistent library does not directly provide joins”.
However, we can use Esqueleto (”a bare bones, type-safe EDSL for SQL queries”) with Persistent's serialization to write type-safe SQL queries. It’s unlikely that you want to use Persistent by itself with SQL, so let’s use and review them together.
We’re using persistent (2.14.5.1), persistent-postgresql (2.13.5.2), and esqueleto (3.5.10.1), all published in 2023. Additionally, we’ll use the experimental style, which will become the new "default" in esqueleto-4.0.0.0.
We’ll also use mtl, monad-logger, unliftio-core, time, and exceptions.
The libraries require additional extensions: DataKinds, GADTs, TypeFamilies, and UndecidableInstances.
💡 (It’s not very important, but) We assume you’ve seen the part on
postgresql-simple, which covers the same topics but at a slower pace.
Database.Persist.Postgresql provides various ways to connect to postgres with and without a connection pool.
First, we need a libpq connection string, which looks like this "host=localhost port=5432 user=postgres dbname=warehouse password=password".
We create the pool and run actions on it using withPostgresqlPool and passing the connection string, number of connections, and action(s) to be executed. We use liftSqlPersistMPool to run an action/transaction on a pool. And “finally”, use runNoLoggingT (runStdoutLoggingT, or alternative) to run with appropriate logging.
runNoLoggingT $ P.withPostgresqlPool Hardcoded.connectionString 3 $ \pool -> do
le runWithPool = flip liftSqlPersistMPool pool
runWithPool transaction1
runWithPool transaction2
...💡 We can use
runStdoutLoggingTto see what sql queries get executed.
Persistent takes care of creating and matching Haskell datatypes and PersistEntity instances; we need to declare the entities by passing them to mkPersist:
mkPersist
sqlSettings
[persistLowerCase|
Product
label Text
description Text Maybe
UniqueLabel label
deriving Eq Show
Category
label Text
UniqueCategory label
deriving Eq Show
ProductCategory
productId ProductId
categoryId CategoryId
Primary productId categoryId
deriving Eq Show
Warehouse
productId ProductId
quantity Int
created UTCTime default=CURRENT_TIME
modified UTCTime default=CURRENT_TIME
deriving Eq Show
|]persistLowerCase states that SomeTable corresponds to the SQL table some_table.
Even though it’s not encouraged, we can always execute raw sql; for example, we can truncate tables with rawExecute:
cleanUp :: (MonadIO m) => SqlPersistT m ()
cleanUp = rawExecute "truncate warehouse, product_category, product, category" []What’s SqlPersistT m ()? Let’s say it’s something that can be executed with runWithPool and returns ().
💡 Note that we can also use deleteWhere to delete all the records from a table:
deleteWhere ([] :: [Filter Product]))Because we’ve done all the groundwork, we use records right away (no tuples):
insertStuff :: (MonadIO m) => SqlPersistT m ()
insertStuff = do
newId <- insert $ Product "Wood Screw Kit 1" (Just "245-pieces")
liftIO $ putStrLn $ "Insert 1: " <> show newId
newIds <- insertMany [Category "Screws", Category "Wood Screws", Category "Concrete Screws"]
liftIO $ putStrLn $ "Insert 2: " <> show newIdsThat’s it! Persistent is concise when it comes to inserts. Note that insert returns the id, and insertMany returns multiple ids.
We can use liftIO to execute IO operations like printing “inside” SqlPersistT.
This is the part where esqueleto comes in.
The first query takes a label and returns a list of product entities:
query1 :: Text -> SqlPersistT m [Entity Product]
query1 label = select $ do
aProduct <- from $ table @Product
where_ (aProduct.label ==. val label)
pure aProductIt returns an Entity instead of a value — an Entity combines a database id and a value.
This is an experimental syntax that mimics sql. We use the TypeApplications extensions to make the table explicit, OverloadedRecordDot to select the field/column value, the ==. operator to check for equality, and val to “lift” haskell value into “sql query land”.
💡 Note that there are other alternatives for field projections (instead of
OverloadedRecordDot), such as the(^.)operator andOverloadedLabels.
We can select multiple labels using in_:
query2 :: [Text] -> SqlPersistT m [Entity Product]
query2 lables = select $ do
aProduct <- from $ table @Product
where_ $ aProduct.label `in_` valList lables
pure aProductWe’ve been kind-of using transactions all this time. Everything inside a single call to liftSqlPersistMPool (and other versions, with and without pooling) runs in a single transaction.
insertWithTransaction :: (MonadIO m, MonadCatch m) => SqlPersistT m ()
insertWithTransaction = handle (\(SomeException _) -> pure ()) $ do
productId <- insert $ Product "Drywall Screws Set" (Just "8000pcs")
categoryId <- insert $ Category "Drywall Screws"
time <- liftIO getCurrentTime
_ <- insert_ $ Warehouse productId 10 time time
_ <- insert_ $ ProductCategory productId categoryId
liftIO $ putStrLn $ "Insert with transaction"This time, we handle exceptions (any SomeException).
💡 We generally want to split the queries into transactions and catch exceptions on each transaction. We dive deeper into error handling in the errors section.
And this is the part where experimental syntax comes in handy:
query quantity = select $ do
(warehouse :& aProduct :& _ :& category) <-
from
$ table @Warehouse
`innerJoin` table @Product
`on` do \(w :& p) -> w.productId ==. p.id
`LeftOuterJoin` table @ProductCategory
`on` do \(_ :& p :& pc) -> just p.id ==. pc.productId
`LeftOuterJoin` table @Category
`on` do \(_ :& _ :& pc :& c) -> pc.categoryId ==. c.id
where_ (warehouse.quantity >. val quantity)
pure $ (warehouse.quantity, aProduct.label, aProduct.description, category.label)The on clauses are attached directly to the relevant join. The ON clause lambda has all the available tables — only the tables we have already joined into are in scope.
We use the :& operator to pattern match against the joined tables. We use _ placeholder to ignore the previous references to the table.
This generates this query:
SELECT
"warehouse"."quantity",
"product"."label",
"product"."description",
"category"."label"
FROM
"warehouse"
INNER JOIN "product" ON "warehouse"."product_id" = "product"."id"
LEFT OUTER JOIN "product_category" ON "product"."id" = "product_category"."product_id"
LEFT OUTER JOIN "category" ON "product_category"."category_id" = "category"."id"
WHERE
"warehouse"."quantity" > ?It’s possible to write type-checked queries that fail at runtime, but most typical sql errors are caught as compile-time errors.
Sometimes, mistakes in queries will result in error messages that refer to library internals (for example, you might see PersistUniqueRead backend0, Database.Esqueleto.Internal.Internal.SqlExpr, PersistRecordBackend backend val, ‘BaseBackend backend0’, ‘SqlBackend’). This takes some time to get used to. Help the type inference, and it will help you.
Nobody is safe from runtime sql errors. For example, if we violate the uniqueness constraint, we get an exception that we need to deal with:
errors :: (MonadIO m, MonadCatch m) => SqlPersistT m ()
errors = do
let duplicateScrew = Product "Duplicate screw" Nothing
void $ insert duplicateScrew
(void $ insert duplicateScrew)
`catch` (\(SomeException err) -> liftIO $ putStrLn $ "Caught SQL Error: " <> displayException err)
Caught SQL Error: SqlError {sqlState = "23505", sqlExecStatus = FatalError, sqlErrorMsg = "duplicate key value violates unique constraint \"product_label_key\"", sqlErrorDetail = "Key (label)=(Duplicate screw) already exists.", sqlErrorHint = ""}
Note that we use the exceptions package to handle exceptions. (We don’t use the exceptions from Control.Exception as we did in postgresql-simple because we don’t want to be limited to IO).
persistent is well documented as part of the yesod book, and esqueleto has good readme and docs. The catch is that you have to keep an eye on multiple packages simultaneously.
On top of that, (currently) esqueleto supports legacy and experimental syntax, and you have to be aware that some tutorials and examples use less safe legacy syntax (or a mix of both) — the good news is that the compiler will warn you if you’re on the wrong path.
persistent can automatically create tables and do migrations. However, the book discourages that:
“Using automated database migrations is only recommended in development environments. Allowing your application to modify your database schema in a production environment is very strongly discouraged.”
You should consider persistent with esqueleto if you mainly have a lot of simple queries, are tired of writing raw sql, but want moderately more type-safe and composable sql.
The persistent is a (very) generalized library, meaning you should be comfortable using abstractions. And you should be familiar with mtl, monad-logger, lifting/unlifting IO, and so on.
“Most kinds of errors committed when writing SQL are caught as compile-time errors — although it is possible to write type-checked
esqueletoqueries that fail at runtime”
If you look around, some tutorials and comparisons might say that esqueleto joins might lead to to runtime errors. Don’t worry — this refers to legacy syntax — use new/experimental syntax.
Tired of sql and Template Haskell?
Beam “is a highly-general library for accessing any kind of database with Haskell”. Beam makes extensive use of GHC's Generics mechanism — no Template Haskell.
First, install beam-core (0.10.1.0 released in 2023) and beam-postgres(0.5.3.1).
A few additional extensions: GADTs and TypeFamilies.
beam-postgres is built on top of postgresql-simple, which is used for connection management, transaction support, serialization, and deserialization.
💡 We assume that you’ve seen the part on
postgresql-simple.
We use postgresql-simple straight away. Reminder:
connectionInfo :: ConnectInfo
connectionInfo =
defaultConnectInfo
{ connectHost = Hardcoded.host
, connectDatabase = Hardcoded.database
, connectUser = Hardcoded.user
, connectPassword = Hardcoded.password
}Simple.withConnect connectionInfo $ \connection -> do
doFoo connection
doBar connectionLet’s look at the definition of the product table:
data ProductT f = Product
{ id :: Columnar f Int64
, label :: Columnar f Text
, description :: Columnar f (Maybe Text)
}
deriving (Generic)
deriving anyclass (Beamable)
type Product = ProductT Identity
deriving instance Show Product
instance Table ProductT where
data PrimaryKey ProductT f = ProductId (Columnar f Int64)
deriving (Generic)
deriving anyclass (Beamable)
primaryKey = ProductId . (.id)ProductT is a beam table. All beam tables must implement the Beamable typeclass (derived via generics) and the Table typeclass. The Table instance declares the type of primary keys for the table and a function that extracts them. We can use Product to construct values of type Product.
💡 For details, see beam tutorial.
All the other tables look quite similar; see the repo for the rest of the boilerplate. One interesting bit is foreign keys / referencing other primary keys; for example, product_id and category_id in the mapping table look like are defined as PrimaryKey ProductT f (not Columnar f Int64):
data ProductCategoryT f = ProductCategory
{ product_id :: PrimaryKey ProductT f
, category_id :: PrimaryKey CategoryT f
}
deriving (Generic)
deriving anyclass (Beamable)After declaring all the tables, we describe our database:
data WarehouseDb f = WarehouseDb
{ product :: f (TableEntity ProductT)
, category :: f (TableEntity CategoryT)
, product_category :: f (TableEntity ProductCategoryT)
, warehouse :: f (TableEntity WarehouseT)
}
deriving (Generic)
deriving anyclass (Database Postgres)
warehouseDb :: DatabaseSettings Postgres WarehouseDb
warehouseDb =
defaultDbSettings
`withDbModification` dbModification
{ product_category =
modifyTableFields
tableModification
{ category_id = CategoryId (fieldNamed "category_id")
, product_id = ProductId (fieldNamed "product_id")
}
, warehouse =
modifyTableFields @WarehouseT
tableModification
{ product_id = ProductId (fieldNamed "product_id")
}
}WarehouseDb needs to define all the tables and an instance of Database _.
💡 Note that you don’t need to hardcode
Postgresand can keep the database more generic.
If you don’t have an existing database, you might get away with only defaultDbSettings as DatabaseSettings. Beam can guess a lot about the tables if we follow their conventions. But we need to override a few generated table fields in our case.
Remember that we have a couple of foreign keys? Beam adds a suffix __id to these, meaning if we have a record field product_id, generated queries will try to use the column product_id__id. So, we must override these in the product_category* mapping and warehouse tables.
💡 See beam defaults for more information.
For raw queries, we can use postgresql-simple:
cleanUp :: Connection -> IO ()
cleanUp connection =
void $ Simple.execute_ connection "truncate warehouse, product_category, product, category"Let’s insert some products:
insert1 :: Connection -> IO ()
insert1 connection =
runBeamPostgres connection
$ runInsert
$ insert (warehouseDb.product)
$ insertValues
[ Product 1 "Wood Screw Kit 1" (Just "245-pieces")
, Product 2 "Wood Screw Kit 2" Nothing
]We construct the statement using insert, which accepts a table and values. We use insertValues to supply concrete values (including ids). runInsert runs the statement (in MonadBeam), which runBeamPostgres executes using the given connection.
💡 Note that we can use
runBeamPostgresDebug putStrLninstead ofrunBeamPostgresto see what sql queries get executed.
runInsert doesn’t return anything (no affected rows, no ids, nothing). When we want some confirmation back, we can use runInsertReturningList:
insert2 :: Connection -> IO ()
insert2 connection = do
result :: [Category] <-
runBeamPostgres connection
$ runInsertReturningList
$ insert (warehouseDb.category)
$ insertExpressions
[ Category default_ "Screws"
, Category default_ "Wood Screws"
, Category default_ "Concrete Screws"
]
putStrLn $ "Inserted categories: " <> show resultWe can use insertExpressions function to insert arbitrary sql expressions. In this case, we pass default_ to ask the database to give us default ids.
Instead of talking about Q monads and MonadBeam, let’s look at the examples. First, query all the products:
query1 :: (MonadBeam Postgres m) => m [Product]
query1 = do
let allProducts = all_ (warehouseDb.product)
runSelectReturningList $ select allProductsrunBeamPostgres connection queryImportant bits:
- build a query;
- pass it into
select; - run it in
MonadBeam(usingrunSelectReturningList,runSelectReturningOne, etc); - execute using
runBeamPostgres connection.
For example, to build a query, we can use all_ to introduce all entries of a table together with guard_ to filter the results:
query2 label = runSelectReturningList $ select $ do
aProduct <- all_ warehouseDb.product
guard_ (aProduct.label ==. val_ label)
pure (aProduct.label, aProduct.description)filter_ is built on top of guard_ and allows us to use the in_ clause:
query3 labels =
runSelectReturningList
$ select
$ filter_ (\p -> p.label `in_` predicate)
$ all_ warehouseDb.product
where
predicate = val_ <$> labelsNote that we use val_ to “lift” haskell values into “sql query land”.
We use postgresql-simple for transactions:
insertWithTransaction :: Connection -> IO ()
insertWithTransaction connection = Simple.withTransaction connection $ do
[newProduct] :: [Product] <-
runBeamPostgres connection
$ runInsertReturningList
$ insert (warehouseDb.product)
$ insertExpressions [Product default_ "Drywall Screws Set" (just_ "8000pcs")]
[newCategory] <-
runBeamPostgres connection
$ runInsertReturningList
$ insert (warehouseDb.category)
$ insertExpressions [Category default_ "Drywall Screws"]
runBeamPostgresDebug putStrLn connection
$ runInsert
$ insert (warehouseDb.product_category)
$ insertValues [ProductCategory (pk newProduct) (pk newCategory)]
runBeamPostgres connection
$ runInsert
$ insert (warehouseDb.warehouse)
$ insertExpressions [Warehouse default_ (val_ (pk newProduct)) 10 currentTimestamp_ currentTimestamp_]
putStrLn $ "Insert with transaction"We use currentTimestamp_ to ask the database for the current time and pk to get the entity's primary key. For example, we pass pk newProduct into the ProductCategory mapping.
There are various ways to get data from multiple tables using Beam.
For example, we can use related_ to get all entries of the given table referenced by the given primary key and leftJoin_ to introduce a table using a left join:
query1 quantity = runBeamPostgres connection
$ runSelectReturningList
$ select
$ do
warehouse <- all_ warehouseDb.warehouse
aProduct <- related_ warehouseDb.product warehouse.product_id
mapping <-
leftJoin_
(all_ warehouseDb.product_category)
(\pc -> pc.product_id ==. primaryKey aProduct)
category <-
leftJoin_
(all_ warehouseDb.category)
(\c -> just_ (primaryKey c) ==. mapping.category_id)
guard_ (warehouse.quantity >. quantity)
pure (warehouse.quantity, aProduct.label, aProduct.description, category.label)Which generates the following query:
SELECT
"t0"."quantity" AS "res0",
"t1"."label" AS "res1",
"t1"."description" AS "res2",
"t3"."label" AS "res3"
FROM
"warehouse" AS "t0"
INNER JOIN "product" AS "t1" ON ("t0"."product_id") = ("t1"."id")
LEFT JOIN "product_category" AS "t2" ON ("t2"."product_id") = ("t1"."id")
LEFT JOIN "category" AS "t3" ON ("t3"."id") IS NOT DISTINCT
FROM
("t2"."category_id")
WHERE
("t0"."quantity") > (3)We can also use the manyToMany_ construct to fetch sides of a many-to-many relationship.
productCategoryRelationship :: ManyToMany Postgres WarehouseDb ProductT CategoryT
productCategoryRelationship =
manyToMany_ (warehouseDb.product_category) (.product_id) (.category_id)query2 quantity = runBeamPostgres connection
$ runSelectReturningList
$ select
$ do
warehouse <- all_ warehouseDb.warehouse
products <- related_ warehouseDb.product warehouse.product_id
categories <- all_ warehouseDb.category
(aProduct, category) <- productCategoryRelationship (pure products) (pure categories)
guard_ (warehouse.quantity >. quantity)
pure (warehouse.quantity, aProduct.label, aProduct.description, category.label)Which generates the following query:
SELECT
"t0"."quantity" AS "res0",
"t1"."label" AS "res1",
"t1"."description" AS "res2",
"t2"."label" AS "res3"
FROM
"warehouse" AS "t0"
INNER JOIN "product" AS "t1" ON ("t0"."product_id") = ("t1"."id")
CROSS JOIN "category" AS "t2"
INNER JOIN "product_category" AS "t3" ON (
("t3"."product_id") = ("t1"."id")
)
AND (
("t3"."category_id") = ("t2"."id")
)
WHERE
("t0"."quantity") > (3)It’s not possible to write an invalid sql query, but this comes at a cost — compile-time errors.
For example, once we forgot to pass a parameter, and this resulted in:
Couldn't match expected type: Q Postgres
WarehouseDb
QBaseScope
a0
with actual type: Q Postgres
WarehouseDb
s0
(ProductT (QExpr Postgres s0))
-> Q Postgres WarehouseDb s0 (CategoryT (QExpr Postgres s0))
-> Q Postgres
WarehouseDb
s0
(ProductT (QExpr Postgres s0),
CategoryT (QExpr Postgres s0))Runtime sql errors are still there, re-exported from postgresql-simple. Review the relevant error section if you need a reminder.
errors :: Connection -> IO ()
errors connection = do
insertDuplicateScrew
insertDuplicateScrew
`catch` (\err@SqlError{} -> putStrLn $ "Caught SQL Error: " <> displayException err)
where
insertDuplicateScrew =
runBeamPostgres connection
$ runInsert
$ insert (warehouseDb.product)
$ insertExpressions [Product default_ "Duplicate screw" nothing_]
Caught SQL Error: SqlError {sqlState = "23505", sqlExecStatus = FatalError, sqlErrorMsg = "duplicate key value violates unique constraint \"product_label_key\"", sqlErrorDetail = "Key (label)=(Duplicate screw) already exists.", sqlErrorHint = ""}
Beam has you covered — it comes with an overview, quick-start guide, tutorial, user guide, and hackage docs.
Spoiler alert: Beam is likely the best-documented library reviewed in this tutorial.
The beam-migrate package provides a migrations framework.
“The
beam-migratetool can generate a beam schema from a pre-existing database, manage migrations for several production databases, automatically generate migrations between two schemas, and much more.”
beam states that if the query compiles, it will generate proper code. Beam uses the GHC Haskell type system and nothing else — no Template Haskell. You don’t have to write raw sql or sql like code. After defining some boilerplate, you write and compose queries in a straightforward Haskell style and get valid SQL.
Regarding complexity, let the types do the talking:
manyToMany_
:: ( Database be db, Table joinThrough
, Table left, Table right
, Sql92SelectSanityCheck syntax
, IsSql92SelectSyntax syntax
, SqlEq (QExpr (Sql92SelectExpressionSyntax syntax) s) (PrimaryKey left (QExpr (Sql92SelectExpressionSyntax syntax) s))
, SqlEq (QExpr (Sql92SelectExpressionSyntax syntax) s) (PrimaryKey right (QExpr (Sql92SelectExpressionSyntax syntax) s)) )
=> DatabaseEntity be db (TableEntity joinThrough)
-> (joinThrough (QExpr (Sql92SelectExpressionSyntax syntax) s) -> PrimaryKey left (QExpr (Sql92SelectExpressionSyntax syntax) s))
-> (joinThrough (QExpr (Sql92SelectExpressionSyntax syntax) s) -> PrimaryKey right (QExpr (Sql92SelectExpressionSyntax syntax) s))
-> Q syntax db s (left (QExpr (Sql92SelectExpressionSyntax syntax) s)) -> Q syntax db s (right (QExpr (Sql92SelectExpressionSyntax syntax) s))
-> Q syntax db s (left (QExpr (Sql92SelectExpressionSyntax syntax) s), right (QExpr (Sql92SelectExpressionSyntax syntax) s))Okay, what if we did something quite similar but quite different?
Squeal “is a type-safe embedding of PostgreSQL in Haskell”, which means “that Squeal embeds both SQL terms and SQL types into Haskell at the term and type levels respectively. This leads to a very high level of type-safety”.
Install squeal-postgresql (0.9.1.3 released in 2023) and generics-sop, which the library uses for generic encodings of Haskell tuples and records.
Enable: DataKinds, GADTs, and OverloadedLabels
💡 (It’s not very important, but) We assume you’ve seen the part on postgresql-simple, which covers the same topics but at a slower pace.
We pass libpq connection string (e.g., "host=localhost port=5432 user=postgres dbname=warehouse password=password") to withConnection:
withConnection Hardcoded.connectionString $
doFoo
& pqThen doBar
& pqThen doBazWe can also create a connection pool using createConnectionPool and use the pool with usingConnectionPool.
First, we define table columns and constraints:
type ProductColumns =
'[ "id" ::: 'Def :=> 'NotNull 'PGint4
, "label" ::: 'NoDef :=> 'NotNull 'PGtext
, "description" ::: 'NoDef :=> 'Null 'PGtext
]
type ProductConstraints = '["pk_product" ::: 'PrimaryKey '["id"]]'Def means that DEFAULT is available for inserts and updates, 'NoDef — unavailable. We specify nullability with 'NotNull and 'Null and the primary key with 'PrimaryKey.
We use the ::: type operators to pair a Symbol with schema types, constraints, column types, etc. We use :=> to specify constraints as well as optionality.
All the other tables look pretty similar (with additional 'ForeignKey constraints here and there); see the repo for the rest of the boilerplate.
Then, we define a schema:
type Schema =
'[ "product" ::: 'Table (ProductConstraints :=> ProductColumns)
, "category" ::: 'Table (CategoryConstraints :=> CategoryColumns)
, "product_category" ::: 'Table (ProductCategoryConstraints :=> ProductCategoryColumns)
, "warehouse" ::: 'Table (WarehouseConstraints :=> WarehouseColumns)
]
type DB = Public SchemaWe use generics to convert between Haskell and PostgreSQL values:
import qualified Generics.SOP as SOP
import qualified GHC.Generics as GHCdata BasicProduct = BasicProduct {label :: Text, description :: Maybe Text}
deriving stock (Show, GHC.Generic)
deriving anyclass (SOP.Generic, SOP.HasDatatypeInfo)The SOP.Generic and SOP.HasDatatypeInfo instances allow us to encode and decode BasicProducts.
We can execute raw statements:
cleanUp :: PQ DB DB IO ()
cleanUp =
execute_ teardown
where
teardown :: Statement db () ()
teardown = manipulation $ UnsafeManipulation "truncate warehouse, product_category, product, category"Manipulation represents update, insert, and delete statements.
We can specify the schema changes by using concrete PQ; when the schema doesn't change, we can use MonadPQ constraint (e.g., cleanUp :: (MonadPQ DB m) => m ()). In the end, we’ll turn either into IO:
withConnection Hardcoded.connectionString
$ cleanUpLet’s insert a product:
insertProduct :: Statement DB BasicProduct ()
insertProduct =
manipulation
$ insertInto_
#product
(Values_ (Default `as` #id :* Set (param @1) `as` #label :* Set (param @2) `as` #description))A Statement is either a Manipulation or a Query that can be run in a MonadPQ.
We use manipulation and insertInto_ to construct an insert. We pass a table and what to insert. Values_ describes a single n-ary product, where we must match all the columns. We can use Default value for id and set the rest using relevant parameters.
And then, we use executePrepared_ to run a statement that returns nothing. The function prepares the statement and runs it on each element.
insertStuff :: (MonadPQ DB m) => m ()
insertStuff = do
executePrepared_
insertProduct
[ BasicProduct "Wood Screw Kit 1" (Just "245-pieces")
, BasicProduct "Wood Screw Kit 2" Nothing
]insertInto_ is a specialized version of insertInto with OnConflictDoRaise (what to do in case of conflict) and no ReturningClause (what to return). ReturningClause returns a value based on each row; for example, we can use it to return the created id:
insertCategory :: Statement DB Category (Only Int32)
insertCategory =
manipulation
$ insertInto
#category
(Values_ (Default `as` #id :* Set (param @1) `as` #label))
OnConflictDoRaise
(Returning_ (#id `as` #fromOnly))Note that we have to use Only and #fromOnly, because we can’t use primitive types (because they don’t have named labels that the library relies on).
This time we have to use executePrepared, which returns a list of results:
insertStuff :: (MonadPQ DB m, MonadIO m) => m ()
insertStuff = do
result :: [Result (Only Int32)] <-
executePrepared insertCategory [Category "Screws", Category "Wood Screws", Category "Concrete Screws"]
rows <- traverse getRows result
liftIO $ putStrLn $ "Inserted categories: " <> show rowsWe use getRows to get all rows from a Result.
To retrieve data, we also write Statements, this time using query and select_:
query1 :: Statement DB () BasicProduct
query1 =
query
$ select_
(#product ! #label :* #product ! #description)
(from (table #product))The query returns all the products from the table.
💡 Note that we can use
printSQLto print statements and see what sql queries get executed.
💡 We can’t return tuples or primitive types because they don't have named fields. You must define a new datatype and derive Squeal typeclasses to return something new.
If you try using tuples, you get an error:
The type `(Text, Text)' is not a record type.
It has no labelled fields.
And then we execute the query:
insertStuff :: (MonadPQ DB m, MonadIO m) => m ()
insertStuff = do
result1 <- execute query1
rows1 <- getRows result1
liftIO $ putStrLn $ "Query 1: " <> show rows1We can select specific fields and narrow down the results:
query2 :: Statement DB (Only Text) BasicProduct
query2 =
query
$ select_
(#product ! #label :* #product ! #description)
(from (table #product) & where_ (#product ! #label .== (param @1)))We use where_ to filter the rows and .== to compare for equality.
This time, we use executeParams to pass the parameters into the statement:
queryData :: PQ DB DB IO ()
queryData = do
result2 <- executeParams query2 (Only "Wood Screw Kit 1") >>= getRows
liftIO $ putStrLn $ "Query 2: " <> show result2We can also use in_:
query3 labels =
query
$ select_
(#product ! #label :* #product ! #description)
(from (table #product) & where_ (#product ! #label `in_` labels))do
(result3 :: [BasicProduct]) <- execute (query3 ["Wood Screw Kit 2", "Wood Screw Kit 3"]) >>= getRows
liftIO $ putStrLn $ "Query 3: " <> show result3We can wrap computation in transactionally_:
insertWithTransaction :: PQ DB DB IO ()
insertWithTransaction =
transactionally_
( do
result1 <- executePrepared insertProduct [BasicProduct "Drywall Screws Set" (Just "8000pcs")]
productIds <- join <$> traverse getRows result1
result2 <- executePrepared insertCategory [Category "Drywall Screws"]
categoryIds <- join <$> traverse getRows result2
case (productIds, categoryIds) of
([Only productId], [Only categoryId]) -> do
executePrepared_ insertProductCategory [(productId, categoryId)]
executePrepared_ insertListing [(productId, 10)]
_ ->
throwM $ userError "Failed to insert product/category"
)
>> liftIO (putStrLn $ "Insert with transaction")In case of exception, it rollbacks the transaction and rethrows the exception.
We use innerJoin and leftOuterJoin to join the tables:
query1 :: Statement DB (Only Int32) Listing
query1 =
query
$ select_
(#w ! #quantity `as` #quantity :* #p ! #label `as` #label :* #p ! #description `as` #description :* #c ! #label `as` #category)
( from
( table (#warehouse `as` #w)
& innerJoin
(table (#product `as` #p))
(#w ! #product_id .== #p ! #id)
& leftOuterJoin
(table (#product_category `as` #pc))
(#pc ! #product_id .== #p ! #id)
& leftOuterJoin
(table (#category `as` #c))
(#c ! #id .== #pc ! #category_id)
)
& where_ (#w ! #quantity .> (param @1))
)Which generates:
SELECT "w"."quantity" AS "quantity",
"p"."label" AS "label",
"p"."description" AS "description",
"c"."label" AS "category"
FROM "warehouse" AS "w"
inner join "product" AS "p"
ON ( "w"."product_id" = "p"."id" )
left outer join "product_category" AS "pc"
ON ( "pc"."product_id" = "p"."id" )
left outer join "category" AS "c"
ON ( "c"."id" = "pc"."category_id" )
WHERE ( "w"."quantity" > ( $1 :: int4 ) )If you forget or mistype anything, most of the time, the error messages are rarely simple.
Sometimes, they overwhelm:
_ :: NP
(Aliased
(Optional
(Expression
'Ungrouped
'[]
'[]
'["public"
::: '["product" ::: 'Table (ProductConstraints :=> ProductColumns),
"category" ::: 'Table (CategoryConstraints :=> CategoryColumns),
"product_category"
::: 'Table (ProductCategoryConstraints :=> ProductCategoryColumns),
"warehouse"
::: 'Table (WarehouseConstraints :=> WarehouseColumns)]]
'[ 'NotNull 'PGtext, 'Null 'PGtext]
from0)))
'["description" ::: ('NoDef :=> 'Null 'PGtext)]
Where: ‘from0’ is an ambiguous type variableSometimes, they leak:
Couldn't match type: TupleOf (TupleCodeOf Text (SOP.Code Text))
with: null10 'PGtext : xs0Ambiguous type variable ‘y0’ arising from a use of ‘manipulation’
prevents the constraint ‘(SOP.Generic y0)’ from being solved.Couldn't match type: '["description"
::: ('NoDef :=> 'Null 'PGtext)]
with: '[]Sometimes, they really leak:
Couldn't match type: records-sop-0.1.1.1:Generics.SOP.Record.ExtractTypesFromRecordCode
(records-sop-0.1.1.1:Generics.SOP.Record.ToRecordCode_Datatype
y (SOP.DatatypeInfoOf y) (SOP.Code y))
with: records-sop-0.1.1.1:Generics.SOP.Record.GetSingleton
(SOP.Code y)
arising from a use of ‘manipulation’But when it comes to runtime SQL errors, the library provides a convenient SquealException for exceptions that Squeal can throw and a nice API for working with them built on top of exceptions. For example, we can use catchSqueal:
errors :: PQ DB DB IO ()
errors = do
insertDuplicateScrew
insertDuplicateScrew
`catchSqueal` (\err -> liftIO $ putStrLn $ "Caught Squeal/SQL Error: " <> displayException err)
where
insertDuplicateScrew = executePrepared_ insertProduct [BasicProduct "Duplicate screw" Nothing]
insertProduct =
manipulation
$ insertInto_
#product
(Values_ (Default `as` #id :* Set (param @1) `as` #label :* Set (param @2) `as` #description))The library comes with a quickstart and Core Concepts Handbook.
The library has a Migration module to change the database schema over time. They support linear, pure or impure, one-way or rewindable migrations.
Squeal is another type-safe postgres library not suitable for beginners. You should be comfortable working on the type level, reading generic-related errors, etc. The library uses generic encodings (generics-sop) of records/tuples, which keep getting into the error messages.
Okay, what if we did something quite similar but quite different?
Opaleye is “an SQL-generating DSL targeting PostgreSQL. Allows Postgres queries to be written within Haskell in a typesafe and composable fashion”.
Install opaleye (0.10.1.0 released in 2023) and product-profunctors, which the library uses under the hood.
opaleye is built on top of postgresql-simple, which is used for connection management, transaction support, serialization, and deserialization.
💡 We assume you’ve seen the part on
postgresql-simple.
We use postgresql-simple straight away. Reminder:
connectionInfo :: ConnectInfo
connectionInfo =
defaultConnectInfo
{ connectHost = Hardcoded.host
, connectDatabase = Hardcoded.database
, connectUser = Hardcoded.user
, connectPassword = Hardcoded.password
}Simple.withConnect connectionInfo $ \connection -> do
doFoo connection
doBar connectionWe define a table using the table function — specify the table name and the type of fields.
warehouseTable ::
Table
(Maybe (Field SqlInt4), ProductIdField, Field SqlInt4, Maybe (Field SqlTimestamp), Maybe (Field SqlTimestamp))
(Field SqlInt4, ProductIdField, Field SqlInt4, Field SqlTimestamp, Field SqlTimestamp)
warehouseTable =
table "warehouse"
$ p5
( tableField "id"
, pProductId (ProductId (tableField "product_id"))
, tableField "quantity"
, tableField "created"
, tableField "modified"
)Table takes write fields and view fields; for example, the first parameter (id) is auto-generated, so it’s optional on write (so we specify Maybe (Field SqlInt4)) but always there on view/read (so we specify Field SqlInt4).
p5 is a tiny glue function from product-profunctors; the number corresponds to the tuple arity (number of columns). We don’t need to know more than that.
tableField infers a required or an optional field depending on the write type.
(We’ll cover ProductId a bit later)
We can also use records instead of tuples; for example, for product:
data Product' a b c = Product {pId :: a, pLabel :: b, pDescription :: c}
type Product = Product' ProductId Text (Maybe Text)
deriving instance Show Product
type ProductFieldWrite = Product' (ProductId' (Maybe (Field SqlInt4))) (Field SqlText) (FieldNullable SqlText)
type ProductField = Product' (ProductIdField) (Field SqlText) (FieldNullable SqlText)Note that we prefix field names because we’ll have some derivable code that can’t handle duplicate record fields.
Product' is polymorphic in all its fields. We’ll use Product in “normal” code and ProductField when interacting with a database. Because id is optional on write, we distinguish between ProductFieldWrite and ProductField.
We indicate nullable fields with FieldNullable, which will be converted into Maybe when executed.
We need some typeclass instances, which we can get with Template Haskell:
$(makeAdaptorAndInstance "pProduct" ''Product')💡 If you’d rather write these by hand, see Data.Profunctor.Product.TH.
And then, we define the table:
productTable :: Table ProductFieldWrite ProductField
productTable =
table "product"
$ pProduct
Product
{ pId = pProductId $ ProductId $ tableField "id"
, pLabel = tableField "label"
, pDescription = tableField "description"
}Note that instead of pN, we use pProduct, which we just generated with TH.
The library’s basic tutorial suggests using newtypes for ids. For example, we use one for the product id:
newtype ProductId' a = ProductId a
$(makeAdaptorAndInstance "pProductId" ''ProductId')
type ProductId = ProductId' Int
deriving instance Show ProductId
type ProductIdField = ProductId' (Field SqlInt4)See the repo for the rest of the boilerplate.
For raw queries, we can use postgresql-simple:
cleanUp :: Connection -> IO ()
cleanUp connection =
void $ Simple.execute_ connection "truncate warehouse, product_category, product, category"Otherwise, we create an insert by using Insert:
insert1 :: Insert [ProductId]
insert1 =
Insert
{ iTable = productTable
, iRows =
[ Product (ProductId Nothing) "Wood Screw Kit 1" (maybeToNullable Nothing)
, Product (ProductId Nothing) "Wood Screw Kit 2" (maybeToNullable (Just "245-pieces"))
]
, iReturning = rReturning (\(p :: ProductField) -> p.pId)
, iOnConflict = Nothing
}We specify the tables, rows to be inserted, conflict-handling strategy, and what to return. In this case, we return product id using rReturning.
And then we run the insert with runInsert:
insertStuff :: Connection -> IO ()
insertStuff connection = do
result1 <- runInsert connection insert1
putStrLn $ "Inserted products: " <> show result1If we want to return the number of affected rows, we can use rCount:
insert2 :: Insert Int64
insert2 =
Insert
{ iTable = categoryTable
, iRows = [Category (CategoryId 1) "Screws", Category (CategoryId 2) "Wood Screws", Category (CategoryId 3) "Concrete Screws"]
, iReturning = rCount
, iOnConflict = Nothing
}The basic select is simple:
selectProduct :: Select ProductField
selectProduct = selectTable productTable💡 Note that we can use
showSqlto printSelectand see what sql queries get executed.
We run a select with runSelect:
queryData :: Connection -> IO ()
queryData connection = do
result1 :: [Product] <- runSelect connection selectProduct
putStrLn $ "Query 1: " <> show result1runSelect “converts a "record" of Opaleye fields to a list of "records" of Haskell values.” We must specify the return type ([Product]) to help type inference.
We can select specific fields and narrow down the results:
select2 :: Select (Field SqlText, FieldNullable SqlText)
select2 = do
(Product _ aLabel description) <- selectProduct
where_ (aLabel .== "Wood Screw Kit 2")
pure (aLabel, description)result2 :: [(Text, Maybe Text)] <- runSelect connection select2We use where_ to filter the rows and .== to compare for equality. We can also use in_:
select3 :: Select (Field SqlText)
select3 = do
p <- selectProduct
where_ $ in_ ["Wood Screw Kit 2", "Wood Screw Kit 3"] p.pLabel
pure (p.pLabel)result3 :: [Text] <- runSelect connection select3We use postgresql-simple for transactions:
insertWithTransaction :: Connection -> IO ()
insertWithTransaction connection = Simple.withTransaction connection $ do
[newProduct] :: [ProductId] <-
runInsert connection
$ Insert
{ iTable = productTable
, iRows = [Product (ProductId Nothing) "Drywall Screws Set" (maybeToNullable $ Just "8000pcs")]
, iReturning = rReturning (\(p :: ProductField) -> p.pId)
, iOnConflict = Nothing
}
[newCategory] :: [CategoryId] <-
runInsert connection
$ Insert
{ iTable = categoryTable
, iRows = [Category (CategoryId 123) "Drywall Screws"]
, iReturning = rReturning (.cId)
, iOnConflict = Nothing
}
void
$ runInsert connection
$ Insert
{ iTable = productCategoryTable
, iRows = [Mapping (toFields newProduct) (toFields newCategory)]
, iReturning = rCount
, iOnConflict = Nothing
}
void
$ runInsert connection
$ Insert
{ iTable = warehouseTable
, iRows = [(Nothing, (toFields newProduct), 10, Nothing, Nothing)]
, iReturning = rCount
, iOnConflict = Nothing
}
putStrLn $ "Insert with transaction"Opaleye provides a couple of APIs for joins. They recommend using where_ directly for inner joins and optional for left/right joins. Which gives us something like this:
join :: Select (Field SqlInt4, Field SqlText, FieldNullable SqlText, MaybeFields (Field SqlText))
join = do
(_, wProductId, quantity, _, _) <- selectTable warehouseTable
p <- selectTable productTable
mpc <- optional $ do
pc <- selectTable productCategoryTable
where_ $ pc.productId .=== p.pId
pure pc
mc <- optional $ do
c <- selectTable categoryTable
where_ $ isJustAnd mpc $ \pc -> c.cId .=== pc.categoryId
pure c
where_ $ wProductId .=== p.pId
where_ $ quantity .> 3
let category = cLabel <$> mc
pure (quantity, p.pLabel, p.pDescription, category)💡 Note that we use
isJustAndthat will be added to Opaleye in the future.isJustAnd :: MaybeFields a -> (a -> Field SqlBool) -> Field SqlBool isJustAnd ma cond = matchMaybe ma $ \case Nothing -> sqlBool False Just a -> cond a
Which generates:
SELECT
"quantity2_1" as "result1_7",
"label1_2" as "result2_7",
"description2_2" as "result3_7",
NOT (("rebind0_6") IS NULL) as "result4_7",
"label1_5" as "result5_7"
FROM (SELECT
*
FROM (SELECT *
FROM
(SELECT *
FROM
(SELECT
*
FROM (SELECT
"id" as "id0_1",
"product_id" as "product_id1_1",
"quantity" as "quantity2_1",
"created" as "created3_1",
"modified" as "modified4_1"
FROM "warehouse" as "T1") as "T1",
LATERAL
(SELECT
"id" as "id0_2",
"label" as "label1_2",
"description" as "description2_2"
FROM "product" as "T1") as "T2") as "T1"
LEFT OUTER JOIN
LATERAL
(SELECT
TRUE as "rebind0_4",
*
FROM (SELECT
*
FROM (SELECT
"product_id" as "product_id0_3",
"category_id" as "category_id1_3"
FROM "product_category" as "T1") as "T1"
WHERE (("product_id0_3") = ("id0_2"))) as "T1") as "T2"
ON
TRUE) as "T1"
LEFT OUTER JOIN
LATERAL
(SELECT
TRUE as "rebind0_6",
*
FROM (SELECT
*
FROM (SELECT
"id" as "id0_5",
"label" as "label1_5"
FROM "category" as "T1") as "T1"
WHERE (CASE WHEN NOT (("rebind0_4") IS NULL) THEN ("id0_5") = ("category_id1_3") ELSE CAST(FALSE AS boolean) END)) as "T1") as "T2"
ON
TRUE) as "T1"
WHERE (("quantity2_1") > (CAST(3 AS integer))) AND (("product_id1_1") = ("id0_2"))) as "T1"Once again, type-safety and query validation equal compilation errors.
But because there isn’t much “type-level magic”, we only need to occasionally help the compiler with type inference. And it’s mainly about input and return types — not intermediate/internal library structures. For example, [Product] in this snippet:
result :: [Product] <- runSelect connection selectProduct
where
selectProduct = selectTable productTableSometimes, if you don’t specify enough types, profunctors show up:
Ambiguous type variable ‘haskells0’ arising from a use of ‘runSelect’
prevents the constraint ‘(Default
FromFields fields0 Product)’ from being solved.
Probable fix: use a type annotation to specify what ‘haskells0’ should be.Runtime sql errors are again from postgresql-simple. Review the relevant error section if you need a reminder.
errors :: Connection -> IO ()
errors connection = do
insertDuplicateScrew
insertDuplicateScrew
`catch` (\err@SqlError{} -> putStrLn $ "Caught SQL Error: " <> displayException err)
where
insertDuplicateScrew =
void
$ runInsert connection
$ Insert
{ iTable = productTable
, iRows = [Product (ProductId Nothing) "Duplicate screw" (maybeToNullable Nothing)]
, iReturning = rCount
, iOnConflict = Nothing
}There are a couple of basic tutorials in the repo and some external ones.
One thing to remember: sometimes, the library provides multiple ways of doing things (for example, left joins using optional vs. deprecated leftJoin or monadic vs. arrow syntax), and documentation/tutorials can do it one way or even deprecated way.
Opaleye assumes a database already exists — no support for migrations or creating tables and databases.
Opaleye allows us to define tables and write type-safe postgres queries using Haskell code.
The library uses product-profunctors and typeclasses. Both only come up in copy-pasteable boilerplate and when you under-specify the return types. No deep knowledge is required.
Okay, what if we did something quite similar but quite different?
Rel8 “is a Haskell library for interacting with PostgreSQL databases”, which aims to be concise, inferrable, and familiar.
For the database connection, instead of postgresql-simple, rel8 uses Hasql.
Install rel8 (1.4.1.0 released in 2023), hasql, and hasql-transaction.
We bring back the TypeFamilies extension and (in case you haven’t already) DuplicateRecordFields. The latter is required to disambiguate the record fields when working with inserts, updates, and deletes…
💡 We assume you’ve seen the parts on
postgresql-simple,hasql, andopaleye.
We use Hasql. Reminder:
Right connection <- getConnectiongetConnection :: IO (Either ConnectionError Connection)
getConnection =
acquire $ settings Hardcoded.host Hardcoded.portNumber Hardcoded.user Hardcoded.password Hardcoded.databaseFirst, we describe the structural mapping of the tables. Take for instance Product:
newtype ProductId = ProductId Int64
deriving newtype (DBEq, DBType, Eq, Show)
data Product f = Product
{ id :: Column f ProductId
, label :: Column f Text
, description :: Column f (Maybe Text)
}
deriving (Generic)
deriving anyclass (Rel8able)
deriving stock instance (f ~ Result) => Show (Product f)We define fields with Column and derive the Rel8able instance. We also declare a newtype for product id with a few instances.
Imagine that the last line is just deriving (Show).
Then, we describe a TableSchema for each table. The relevant table looks like this:
productSchema :: TableSchema (Product Name)
productSchema =
TableSchema
{ name = "product"
, schema = Nothing
, columns =
Product
{ id = "id"
, label = "label"
, description = "description"
}
}Note that defining columns looks repetitive — we can use some generics machinery to get that information from the Rel8able:
productSchema :: TableSchema (Product Name)
productSchema =
TableSchema
{ name = "product"
, schema = Nothing
, columns = namesFromLabels @(Product Name)
}💡
namesFromLabelsgenerates a table schema where every column name corresponds precisely to the field's name. Alternatively, we can usenamesFromLabelsWith.
See the repo for the rest of the boilerplate.
For raw queries, we can use Hasql:
cleanUp :: Connection -> IO (Either QueryError ())
cleanUp connection = run cleanUpSession connection
where
cleanUpSession = statement () $ Statement rawSql E.noParams D.noResult True
rawSql = "truncate warehouse, product_category, product, category"Otherwise, we create Insert:
insert1 :: Statement () [ProductId]
insert1 =
insert
$ Insert
{ into = productSchema
, rows =
values
[ Product unsafeDefault "Wood Screw Kit 1" null
, Product unsafeDefault "Wood Screw Kit 2" (lit $ Just "245-pieces")
]
, returning = Projection (.id)
, onConflict = Abort
}We’ve seen this in Opaleye’s insert: the table, rows to insert, conflict-handling strategy, and what to return.
We use unsafeDefault for sql DEFAULT, lit to turn Haskell values into expressions, and values to construct a query out of the given rows.
💡 Note that unsafeDefault is named unsafe for a reason; see the docs.
And run this like any other Hasql statement:
result1 <- run (statement () insert1) connectionIf we want to return the number of affected rows, we can use NumberOfRowsAffected:
Insert
{ into = categorySchema
, rows =
values
[ Category unsafeDefault "Screws"
, Category unsafeDefault "Wood Screws"
, Category unsafeDefault "Concrete Screws"
]
, returning = NumberOfRowsAffected
, onConflict = Abort
}We build select statements using Query. We select all rows from a table using each and turn (run) the query into Statement using select:
select1 :: Statement () [Product Result]
select1 = select $ each productSchema💡 Note that we can use
showQueryto print sql queries that will be executed.
And once again we run the statement:
result1 <- run (statement () select1) connectionWe can select specific fields and narrow down the results:
select2 :: Statement () [(Text, Maybe Text)]
select2 = select $ do
p <- each productSchema
where_ $ p.label ==. "Wood Screw Kit 2"
pure (p.label, p.description)We use where_ to filter the rows and ==. to compare for equality. We can also use in_:
select3 :: Statement () [Text]
select3 = select $ do
p <- each productSchema
where_ $ p.label `in_` ["Wood Screw Kit 2", "Wood Screw Kit 3"]
pure p.labelNote that the order of parameters is different from Opaleye.
We use Hasql for transactions:
insertWithTransaction :: Connection -> IO ()
insertWithTransaction connection = do
result <- run (transaction Serializable Write insertAll) connection
putStrLn $ "Insert with transaction: " <> show result
where
insertAll = do
productIds <-
Transaction.statement ()
$ insert
$ Insert
{ into = productSchema
, rows = values [Product unsafeDefault "Drywall Screws Set" (lit $ Just "8000pcs")]
, returning = Projection (.id)
, onConflict = Abort
}
-- insert category
-- insert mapping
-- insert warehouse listingRel8 doesn’t have a specific join operation — we use where_ (or filter) to filter the results and optional to do what outer joins do.
queryWithJoins :: Connection -> IO ()
queryWithJoins connection = do
result1 <- run (statement () join) connection
putStrLn $ "Query with join: " <> show result1
where
join :: Statement () [(Int32, Text, Maybe Text, Maybe Text)]
join = select joinQuery
joinQuery = do
w <- each warehouseSchema
p <- productsInWarehouse w
pc <- optional $ mappingsForProduct p
c <- traverseMaybeTable categoriesForMapping pc
where_ $ w.quantity >. 3
let category = maybeTable null (nullify . (.label)) c
pure (w.quantity, p.label, p.description, category)
productsInWarehouse :: Warehouse Expr -> Query (Product Expr)
productsInWarehouse w =
each productSchema >>= filter (\p -> p.id ==. w.product_id)
mappingsForProduct :: Product Expr -> Query (ProductCategory Expr)
mappingsForProduct p = do
each productCategorySchema >>= filter (\pc -> pc.product_id ==. p.id)
categoriesForMapping :: ProductCategory Expr -> Query (Category Expr)
categoriesForMapping pc =
each categorySchema >>= filter (\c -> c.id ==. pc.category_id)We extract “each join” into a specialized function to make the code cleaner (according to the Rel8 tutorials). We use optional and traverseMaybeTable to account for the partiality of queries. MaybeTable results from an outer join, which we unwrap with maybeTable.
filter is an alternative way to write where clauses.
The generated query:
SELECT
CAST("quantity2_1" AS int4) as "_1",
CAST("label1_3" AS text) as "_2",
CAST("description2_3" AS text) as "_3",
CAST(CASE WHEN ("rebind0_8") IS NULL THEN CAST(NULL AS text) ELSE "label1_12" END AS text) as "_4"
FROM (SELECT
*
FROM (SELECT *
FROM
(SELECT *
FROM
(SELECT
*
FROM (SELECT
"id" as "id0_1",
"product_id" as "product_id1_1",
"quantity" as "quantity2_1",
"created" as "created3_1",
"modified" as "modified4_1"
FROM "warehouse" as "T1") as "T1",
LATERAL
(SELECT
"id" as "id0_3",
"label" as "label1_3",
"description" as "description2_3"
FROM "product" as "T1") as "T2"
WHERE (("id0_3") = ("product_id1_1"))) as "T1"
LEFT OUTER JOIN
LATERAL
(SELECT
TRUE as "rebind0_8",
*
FROM (SELECT
*
FROM (SELECT
"product_id" as "product_id0_6",
"category_id" as "category_id1_6"
FROM "product_category" as "T1") as "T1"
WHERE (("product_id0_6") = ("id0_3"))) as "T1") as "T2"
ON
TRUE) as "T1"
LEFT OUTER JOIN
LATERAL
(SELECT
TRUE as "rebind0_14",
*
FROM (SELECT
*
FROM (SELECT
0) as "T1",
LATERAL
(SELECT
"id" as "id0_12",
"label" as "label1_12"
FROM "category" as "T1") as "T2"
WHERE (("id0_12") = ("category_id1_6")) AND (("rebind0_8") IS NOT NULL)) as "T1") as "T2"
ON
TRUE) as "T1"
WHERE (("quantity2_1") > (CAST(3 AS int4))) AND (((("rebind0_14") IS NULL) AND (("rebind0_8") IS NULL)) OR ((("rebind0_14") = ("rebind0_8")) AND (COALESCE(("rebind0_14") = ("rebind0_8"),FALSE))))) as "T1"Which looks similar to the relevant Opaleye query in the previous section.
On top of type-safety, according to the docs, “Rel8 aims to have excellent and predictable type inference”. And they deliver — type inference rarely needs any guidance, and the compilation errors are pretty good.
Although it’s possible to introduce runtime errors using unsafe operations like unsafeDefault, the name is explicit, well documented, and has proper alternatives.
Runtime errors come from Hasql — all error-reporting is explicit and is presented using Either. As a reminder, violating the constraint returns a familiar error:
errors :: Connection -> IO ()
errors connection = do
Left failure <-
run insertDuplicateScrew connection
>> run insertDuplicateScrew connection
putStrLn $ "Constraint violation (Left): " <> show failure
where
insertDuplicateScrew =
statement ()
$ insert
$ Insert
{ into = productSchema
, rows = values [Product unsafeDefault "Duplicate screw" null]
, returning = NumberOfRowsAffected
, onConflict = Abort
}
Constraint violation (Left): QueryError "INSERT INTO \"product\" (\"id\",\n \"label\",\n \"description\")\nVALUES\n(DEFAULT,CAST(E'Duplicate screw' AS text),CAST(NULL AS text))" [] (ResultError (ServerError "23505" "duplicate key value violates unique constraint \"product_label_key\"" (Just "Key (label)=(Duplicate screw) already exists.") Nothing Nothing))
Rel8 has the Getting Started tutorial, the Concepts documentation, the cookbook, and good API docs. This would have been one of the best coverages, but unfortunately, some basic snippets (like running selects or constructing inserts) aren’t valid anymore.
Also, you have to keep in mind hasql.
Rel8 assumes a database already exists — no support for migrations or creating tables and databases.
Rel8 also allows us to write type-safe postgres queries using concise, inferrable, and familiar Haskell code. It builds on top of opaleye and hasql, and you must be somewhat familiar with the latter.
Okay, what if we did something quite similar but quite different?
Selda “is a Haskell library for interacting with SQL-based relational databases” (PostgreSQL or SQLite). “The library was inspired by LINQ and Opaleye.”
Install selda (0.5.2.0 released in 2022) and selda-postgresql.
Enable OverloadedLabels.
Create connection info:
connectionInfo :: PGConnectInfo
connectionInfo =
PGConnectInfo
{ pgHost = Hardcoded.host
, pgPort = Hardcoded.portNumber
, pgDatabase = Hardcoded.database
, pgUsername = Just Hardcoded.user
, pgPassword = Just Hardcoded.password
, pgSchema = Nothing
}And use it with withPostgreSQL:
withPostgreSQL connectionInfo $ do
doFoo
doBarFirst, we declare normal types and derive SqlRow, for example, for product:
data Product = Product
{ id :: ID Product
, label :: Text
, description :: Maybe Text
}
deriving (Generic, Show)
deriving anyclass (SqlRow)Then, we use table to declare a table:
productTable :: Table Product
productTable = table "product" [#id :- autoPrimary]We specify constraints by linking selectors of the table to the definitions. We use autoPrimary for auto-incrementing primary keys, primary for regular primary keys, and foreignKey for foreign keys:
mappingTable :: Table ProductCategory
mappingTable =
table
"product_category"
[ #product_id :- foreignKey productTable #id
, #category_id :- foreignKey categoryTable #id
]See the repo for the rest of the boilerplate.
We can use rawStm from Database.Selda.Unsafe to execute raw queries:
cleanUp :: SeldaM PG ()
cleanUp =
rawStm "truncate warehouse, product_category, product, category"
SeldaM is an alias for SeldaT IO, SeldaT is a Selda computation — a concrete implementation (of MonadSelda) with Selda SQL capabilities.
At the end we’ll turn it into IO:
withPostgreSQL connectionInfo $ do
cleanUpTo insert data, we can use insert_ that doesn’t return anything, insert that returns the number of inserted rows, and insertWithPK that returns the primary key of the last inserted row.
insertStuff :: SeldaM PG ()
insertStuff = do
productId <-
insertWithPK
productTable
[ Product def "Wood Screw Kit 1" (Just "245-pieces")
, Product def "Wood Screw Kit 2" Nothing
]
liftIO $ putStrLn $ "Inserted product with id: " <> show productId rows <-
insert
categoryTable
[Category def "Screws", Category def "Wood Screws", Category def "Concrete Screws"]
liftIO $ putStrLn $ "Inserted categories: " <> show rowsWe use def when we want to use the default value, which is the case with ids.
We can get all the rows from the given table using select:
selectProduct :: Query t (Row t Product)
selectProduct = select productTable💡 Note that we can use
compilefromDatabase.Selda.Debugto print sql queries that will be executed.
And execute the query with query:
queryData :: SeldaT PG IO ()
queryData = do
result1 <- query selectProduct
liftIO $ putStrLn $ "Query 1: " <> show result1🤷 Somehow, here, Selda didn’t want to read/parse back the ids it just inserted:
elephants-exe: [SELDA BUG] fromSql: RowID column with non-int value: SqlInt32...
If we change the type from ID Foo to Int32, the select works, but then insert with auto-incremental primary keys and other functionality doesn’t 🤷
So let’s ignore this for now because other queries work fine.
We can select specific fields and narrow down the results:
select2 :: Query t (Col t Text :*: Col t (Maybe Text))
select2 = do
p <- selectProduct
restrict (p ! #label .== "Wood Screw Kit 2")
pure (p ! #label :*: p ! #description)Query is parameterized over a scope parameter t, ensuring that queries are always well-scoped, but we don’t have to worry about it now.
We use ! with selectors to extract a column, restrict to filter the rows, and .== to compare for equality. A result is an inductive tuple — one or more values separated by the :*: data constructor.
We can also use isIn:
select3 = do
p <- selectProduct
restrict (p ! #label `isIn` ["Wood Screw Kit 2", "Wood Screw Kit 3"])
pure (p ! #label)We use transaction:
insertWithTransaction :: SeldaT PG IO ()
insertWithTransaction = transaction $ do
productId <- insertWithPK productTable [Product def "Drywall Screws Set" (Just "8000pcs")]
categoryId <- insertWithPK categoryTable [Category def "Drywall Screws"]
insert_ mappingTable [ProductCategory productId categoryId]
insert_ warehouseTable [Warehouse def productId 10 def def]
liftIO $ putStrLn $ "Insert with transaction"We use restrict and leftJoin to query with joins:
join :: Query s (Col s Int32 :*: (Col s Text :*: (Col s (Maybe Text) :*: Col s (Coalesce (Maybe Text)))))
join = do
w <- select warehouseTable
p <- select productTable
restrict (w ! #product_id .== p ! #id)
pc <- leftJoin (\pc -> pc ! #product_id .== p ! #id) (select mappingTable)
c <- leftJoin (\c -> just (c ! #id) .== pc ? #category_id) (select categoryTable)
pure (w ! #quantity :*: p ! #label :*: p ! #description :*: c ? #label)We use ? to extract a column from the nullable row.
The generated query:
SELECT
"quantity_2",
"label_6",
"description_7",
"label_13_15"
FROM
(
SELECT
"id_12_14",
"label_13_15",
"category_id_9_11",
"label_6",
"description_7",
"quantity_2"
FROM
(
SELECT
"product_id_8_10",
"category_id_9_11",
"id_5",
"label_6",
"description_7",
"quantity_2"
FROM
(
SELECT
"id_5",
"label_6",
"description_7",
"product_id_1",
"quantity_2"
FROM
(
SELECT
"product_id" AS "product_id_1",
"quantity" AS "quantity_2"
FROM
"warehouse"
) AS q0,
(
SELECT
"id" AS "id_5",
"label" AS "label_6",
"description" AS "description_7"
FROM
"product"
) AS q1
WHERE
("product_id_1" = "id_5")
) AS q3
LEFT JOIN (
SELECT
"product_id_8" AS "product_id_8_10",
"category_id_9" AS "category_id_9_11"
FROM
(
SELECT
"product_id" AS "product_id_8",
"category_id" AS "category_id_9"
FROM
"product_category"
) AS q2
) AS q4 ON "product_id_8_10" = "id_5"
) AS q6
LEFT JOIN (
SELECT
"id_12" AS "id_12_14",
"label_13" AS "label_13_15"
FROM
(
SELECT
"id" AS "id_12",
"label" AS "label_13"
FROM
"product"
) AS q5
) AS q7 ON (
Cast("id_12_14" AS INT)
) = "category_id_9_11"
) AS q8From Selda’s tutorial: “While the types keep queries nice and safe, Haskell's type errors can be a bit daunting even under the best circumstances.” In practice, type inference rarely needed guidance, and the compilation errors were relatively clear.
The only problem we’ve encountered was the mismatch of ID and SqlInt32.
All Selda functions may throw SeldaError:
errors :: SeldaM PG ()
errors = do
insertDuplicateScrew
insertDuplicateScrew
`catch` (\(err :: SeldaError) -> liftIO $ putStrLn $ "Caught Selda Error: " <> displayException err)
where
insertDuplicateScrew = insert_ productTable [Product def "Duplicate screw" Nothing]
elephants-exe: SqlError "error executing query INSERT INTO \"product_category\" (\"product_id\", \"category_id\") VALUES ($1, $2)': ERROR: insert or update on table \"product_category\" violates foreign key constraint \"product_category_category_id_fkey\"\nDETAIL: Key (category_id)=(748) is not present in table \"category\".\n"
Selda comes with a simple overview and example. There is also a tutorial.
The library has a Migrations module for upgrading a table from one schema to another.
Selda allows us to write type-safe queries in a linear, natural style.
Depending on your experience and situation, you can use SeldaM straight, or you may need to get familiar with mtl, exceptions, lifting/unlifting IO, etc.
GenericPersistence "is a small Haskell persistence layer for relational databases", which uses GHC.Generics.
For the database connection, the library uses the HDBC library.
Install generic-persistence (0.6.0 released in 2023) and HDBC-postgresql.
Connect to the database using HDBC's connectPostgreSQL and connect with a transaction mode:
getConnection :: IO Conn
getConnection = connect ExplicitCommit <$> connectPostgreSQL Hardcoded.connectionStringNote that you can use createConnPool to work with a connection pool.
First, we declare normal types and derive Generic, for example, for a product:
data Product = Product {id :: Int64, label :: Text, description :: Maybe Text}
deriving (Show, Generic)We can derive an Entity instance (the default implementations use Generics), but we want to modify a few things. For example, specify the name of the primary key:
instance Entity Product where
idField = "id"Or override table name and default primary-key behavior:
instance Entity ProductCategory where
tableName :: String
tableName = "product_category"
autoIncrement = FalseSee the repo for the rest of the boilerplate.
We can use runRaw to execute raw queries:
cleanUp :: Conn -> IO ()
cleanUp connection = runRaw connection "truncate warehouse, product_category, product, category"
To insert data, we can use insert, which returns an inserted entity:
product1 <- insert connection $ Product{label = "Wood Screw Kit 1", description = Just "245-pieces"}
putStrLn $ "Insert 1: " <> show product1Note that we create a Product and omit the id because we want it generated. We can see that it works as expected from the log:
Insert 1: Product {id = 1139, label = "Wood Screw Kit 1", description = Just "245-pieces"}
💡 Compiler isn't particularly happy about missing fields, so we have
{-# OPTIONS_GHC -Wno-missing-fields #-}at the top of the module.
We can insert multiple categories with insertMany:
insertMany connection [Category{label = "Screws"}, Category{label = "Wood Screws"}, Category{label = "Concrete Screws"}]The function returns IO ().
We can get all the rows from the given table using select and allEntries:
products1 <- select @Product connection allEntries
putStrLn $ "Query 1: " <> show products1allEntries is a where-clause expression. We can also select specific fields and narrow down the results:
let labelWsk2 = "Wood Screw Kit 2" :: Text
products2 <- select @Product connection (field "label" =. labelWsk2)
putStrLn $ "Query 2: " <> show products2We use field and =. to compare the fields. Note that we have to specify the type to help type inference.
We can also use in':
products3 <- select @Product connection (field "label" `in'` [labelWsk2, labelWsk3])
putStrLn $ "Query 3: " <> show products3We use transaction:
insertWithTransaction :: Conn -> IO ()
insertWithTransaction conn = withTransaction conn $ \connection -> do
(Product productId _ _) <- insert connection (Product{label = "Drywall Screws Set", description = Just "8000pcs"})
(Category catId _) <- insert connection (Category{label = "Drywall Screws"})
now <- getCurrentTime
_ <- insert connection (Warehouse productId 10 now now)
_ <- insert connection (ProductCategory catId productId)
putStrLn $ "Inserted with transaction"We use getCurrentTime to explicitly get the current time.
We perform a custom HDBC quickQuery and convert the resulting rows into a list of Listing:
queryWithJoins :: Conn -> IO ()
queryWithJoins connection = do
let stmt =
[sql|
select w.quantity, p.label, p.description, c.label
from warehouse as w
inner join product as p on w.product_id = p.id
left outer join product_category as pc on p.id = pc.product_id
left outer join category as c on c.id = pc.category_id
where w.quantity > (?)|]
listings <- entitiesFromRows @Listing connection =<< quickQuery connection stmt [toSql (3 :: Int)]
putStrLn $ "Query with join: " <> show listingsThe library provides two different APIs:
- The default API, which uses exceptions to signal errors (as demonstrated).
- The safe API, which uses Either to signal errors.
For more information, see Deal with runtime exceptions or use total functions? Your choice!
For example, if we violate the uniqueness constraint, we get PersistenceException:
errors :: Conn -> IO ()
errors connection = do
insertDuplicateScrew
handle @PersistenceException (\err -> putStrLn $ "Caught SQL Error: " <> displayException err) insertDuplicateScrew
where
insertDuplicateScrew =
void $ insert connection $ Product{label = "Duplicate screw", description = Nothing}
Caught SQL Error: DuplicateInsert "Entity already exists in DB, use update instead"
GenericPersistence comes with a Getting Started tutorial, explanations of the internals, a bunch of how-tos, and API documentation – pretty good coverage.
Seems like database migrations aren't in scope yet.
GenericPersistence is a small Haskell persistence layer for relational databases with the design goal of minimizing the boilerplate. The library uses plain IO and doesn't require upper-intermediate knowledge, making it suitable for beginners.
Keep in mind that the library is younger compared to the rest.
At some point, I ran out of steam and didn’t have the energy to make these things work.
I didn’t manage to connect to the database using postgresql-typed. And I didn’t manage to build Haskell Relational Record and groundhog with ghc 9.4.5. Neither had releases this year. Also, there are hdbc and postgresql-libpq (you might have noticed the libpq connection string we used here and there).
All of these sound fun. But we had enough for now. Maybe we’ll revisit and extend it later.
I can’t tell you which library to use. I don’t know myself. But I can tell you the questions you (and your team) should consider:
- Do you want to write raw sql queries? or query builder? or use some sort of ORM? Do you want to learn a new DSL?
- How type-safe do you want to be? How readable/optimizable the generated queries should be?
- Do you need bells and whistles, like built-in connection pools or migrations?
- How comfortable are you with Haskell and type-level? Is this something you can afford?
- Do you need to be database agnostic?
Also, this is just the tip of the iceberg; we haven’t discussed specific features, performance, compilation speed… And spoiler alert: generics, template haskell, and type-families don’t come for free.
We aren’t really at the point where people care to compare the performance. But you don’t know unless you measure. Feel free to explore and get some community karma.
Another excellent contribution opportunity is documentation and tutorials, especially if you have a favorite library and want to convince others to consider it.
Wait, is it actually bad that we have so many libraries?
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent