Necessity is the mother of invention. I've been playing around with ideas for a persistence layer for a while now, but it wasn't until I got a hard deadline for a database-driven website that things really clicked. I want to share these thoughts here, both to get them down in writing, and to hopefully provoke some discussion.
There's some version of the code I discuss here available on github. It doesn't match the code here precisely, and is not what the final version will look like. But it does show how we can support multiple backends!
Pieces of the puzzle
There are going to be three separate components of persistent:
- A high-level interface for declaring entities. There will be a set of datatypes, and perhaps either a DSL, Quasi-Quoted syntax or external files for a nicer user interface.
- A set of type-classes for doing the actual operations on tables. The majority of this post will discuss this.
- Each backend will have its own Template Haskell code for generating instances of the type classes based on the higher-level interface. The SQL generators will probably share a lot of code, but that's mostly irrelevant to the outside user.
Type families
My proposal makes heavy use of type families. If you want a quick introduction to type families, they're basically a method to associate two different datatypes together.
Let's give a pertinent example. Every record in a table will have a unique key associated with it. If we're talking about a SQL backend, this will most likely be an auto_increment (MySQL) or serial (PostgreSQL) integer, but Amazon SimpleDB prefers to use UUIDs. So I want to associate a key value based on the datatype of the value and the backend. In type families, this might look like:
data family Key value backend
However, in general we'll be defining these inside typeclasses, so it will look more like:
class HasTable value backend where
data Key value backend
Then you would declare an instance like so:
instance HasTable MyValue MyBackend where
data Key MyValue MyBackend = MyKey Integer
No relation
I've mentioned a number of times in previous posts the idea of no relations. Let me clarify here: persistent will not automatically load related data from other tables. There are three good reasons for this approach:
- Not all backends natively support relations. Think of a set of CSV files, or Google BigTable for that matter.
- Relations will greatly complicate things, and I really want something simple.
- Arguably, there's a performance edge. But I really don't want to discuss performance right now.
Multiple type classes
So without relations, every record is really just a single row in a single table. We need to be able to do two things:
- Convert our value to a record.
- Convert a record to our value.
That's all well and good, but what's a record? If we're talking about CSV files, it's [String]
. But in a database (using HDBC) it's [SqlValue]
.
So we'll need to have different typeclasses for different backends. However, since we have multiple Template Haskell generators, all is fine: the CSV TH code knows to use the IsStringList
class, while the PostgreSQL TH uses IsSqlValueList
... or whatever it's called.
However, some database operations require more than just this. What if you want to update a record? What if you want to search? Or sort?
I propose a set of classes akin to the following:
class HasField a where
data Field a
applyField :: Field a -> a -> a
class HasFilter a where
data Filter a
applyFilter :: Filter a -> a -> Bool
class HasOrder a where
data Order a
applyOrder :: Order a -> a -> a -> Ordering
Let's see a sample set of instances.
data Person = Person String Int
instance HasField Person where
data Field Person = PersonName String | PersonAge Int -- type safe!
applyField (PersonName name) (Person _ age) = Person name age
applyField (PersonAge age) (Person name _) = Person name age
instance HasFilter Person where
data Filter Person = PersonNameEq String -- name equality
| PersonAgeLt Int -- age is less than parameter
-- we could add as many filters here as we like
applyFilter (PersonNameEq x) (Person y _) = x == y
applyFilter (PersonAgeLt x) (Person _ y) = y < x="" instance="" HasOrder="" a="" where="" data="" Order="" Person="NameAsc" --="" name,="" ascending="" |="" AgeDesc="" --="" age,="" descending="" applyOrder="" NameAsc="" (Person="" x="" _)="" (Person="" y="" _)="compare" x="" y="" applyOrder="" AgeDesc="" (Person="" _="" x)="" (Person="" _="" y)="compare" y="" x="" --="" reverse<="" /code="">
Notice how the names of the constructors in there don't clash. Once you have these data families, you get to do stuff like:
class HasField v => Updateable v backend where
update :: Key v backend -> [Field v] -> IO ()
class HasFilter v => Searchable v backend where
select :: [Filter v] -> IO [(Key v backend, v)]
In case it's not apparent, let me point out the huge win here: everything is completely type checked. We're not passing around SqlValue
s or something like that. Every single field requires a matching datatype, or it simply won't compile.
Of course, this won't be enough information for SQL. It would want something additional such as:
class HasFilter v => SqlFilter v where
toWhereClause :: Filter v -> String
toWhereData :: Filter v -> SqlValue
instance SqlFilter Person where
toWhereClause (PersonNameEq _) = "name=?"
toWhereClause (PersonAgeLt _) = "age<?"
toWhereData (PersonNameEq x) = SqlString x
toWhereData (PersonAgeLt x) = SqlInteger x
The beauty is that Template Haskell will address all these problems for us. You'll be left with being able to just use the select function, without writing any boilerplate.
Coming up
I'm actually actively using code very similar to this in developing a production website. There's a hard deadline for it of less than a week away. You might be wondering why I'm blogging when there's a deadline... good question.
Anyway, the actual type classes are slightly more complex than what I've laid out here, but I think this gives a good overview of the approach. I haven't started working on the high-level code or the Template Haskell yet. One idea I have in mind is putting enough information in the high-level design to be able to automatically generate web forms from them. In fact, I envision a Yesod subsite capable of letting you fully interact with a table in a database.
Questions for the audience
- MySQL support in HDBC does not seem too strong. HDBC also has some annoying tendencies: it doesn't let you manually finalize statements, for example. Is it worth coding the SQL backends straight against the C libraries?
- SimpleDB sounds like a good candidate for a non-SQL backend. CSV files good be a fun one as well. Any other thoughts?
- Where's Waldo?
- What would you like the high-level interface to look like?
- There will almost certainly be an enumerator/iteratee interface for this library. What style of enumerators are people interested in? I know the iteratee package is all the rage, but the new 0.4 version will be completely changing the internals.