schemas
schemas is a Haskell-centric serialization library written with versioning in mind. Since a schema is a first-class citizen, it can be serialized, reasoned about, and transmitted together with the data. Serialization and deserialization work better when the source schema is provided, and versioning is accomplished by checking that the two schemas are related by a subtyping relation. This alleviates the need to keep old versions of datatypes around.
Consider a schema modification that adds a field. To support upgrading old documents to the new schema, the only requirement is that the new field is optional. Downgrading is easy too, simply omit the new field. Conversely, a schema modifcation that removes a field supports trivial upgrades but the removed field must be optional to support downgrading. Changing the type of a field is supported in as much as the old and new types are relatable. Field renaming is not supported. More importantly, all these changes are defined by a schema relation, and the library provides a predicate to check whether the relation holds.
schemas can also be used in a GraphQL-like fashion, allowing clients to request a subset of the schema. This comes up specially when working with recursive schemas involving cyclic data.
Features
- schemas are first-class citizens and can be serialized,
- schema construction is statically typed,
- versioning is driven by a subtyping relation, no need for version numbers,
- Serialization to JSON only currently
Why schemas
A quick seach in Hackage reveals a large number of libraries about schemas, including json-schema, hjsonschema, aeson-schema, aeson-schemas and hschema, amongst others.
There's undoubtedly a large amount of overlapping amongst all these libraries, so the immediate question is, why introduce another one ?
This library is a re-implementation of an encoding library found in the Strats codebase at Standard Chartered Bank, the origins of which is go back a few years in time.
It predates other libraries that accomplish a similar task, including most of the ones mentioned before.
The approach has worked well but the codebase is showing its age and limitations, notably the lack of decoding capabilities.
This library extends the original approach with decoding and alternatives, hopefully keeping the good parts like the subtyping relation, intact.
Subtyping relation
schemas relies on a simple subtyping relation between schemas to perform value conversions. The basic idea is that these conversions are fully guided by the source and target schemas and involve only simple projections and injections:
- Projecting a subset of the source record fields.
- Turning a source field of type
A
into a target field of type Array A
.
For more concrete details on the subtyping relation check the definition of isSubtypeOf
. This function returns a witness, i.e. a conversion function, whenever the relation holds.
Versioning makes use of this subtyping relation as follows. Downgrading a value v_2 :: T_2
into a previous version T_1
is accomplished via the witness of schema(T_1) > schema(T_2)
. Similarly, upgrading a v_1 :: T_1
message into a newer version T_2
can be accomplished via the witness of schema(T_1) < schema(T_2)
. Therefore, a type T_1
can only be replaced by a type T_2
in an downgrade-compatible way if schema(T_1) > schema(T_2)
; if upgrades are required, then schema(T_1) < schema(T_2)
is required too.
The <
relation is reflexive and transitive, but importantly not asymmetric or antisymmetric: it can be that both T_1 < T_2
and T_2 < T_1
and yet they are not the same type. For example, given a S_2
schema that adds a required field to S_1
, we would have that S_1 > S_2
but not S_1 < S_2
. However, if new the field was optional, then we would have S_1 < S_2
too. In such case, we say that S_1 ~ S_2
because they only differ on optional fields.
For example, given a S_3
schema that removes a field from S_2
, we have:
S_2 < S_3
therefore we can upgrade S_2
values to S_3
S_2 ~ S_3
if the removed field is optional, in which case we can also downgrade S_3
values to S_2
The ~
relation is an equivalence class, i.e. it is reflexive, symmetric and transitive.
Alternative encodings
Sometimes there is more than one way to encode a value. A field can be renamed or change its type, an optional field become mandatory, several fields can be merged into one, etc. Alternative encodings allow for backwards compatible schema evolution.
This library support alternative encodings via the Monoid
instance for typed schemas and the Alternative
instance for RecordFields
.
The schema A|B
encodes a value in two alternative ways A
and B
. A message created with this schema may use encodings A, B or both. 'encode' will always create messages with all the possible encodings. While messages with multiple alternative encodings are not desirable for serialization, the desired message can be carved out using the subtyping relation. All the following hold:
A < A|B (the coercion A -> A|B will produce a message with an A encoding)
B < A|B (the coercion B -> A|B will produce a message with a B encoding)
A|B < A (the coercion A|B -> A will succeed only if the message contains an A encoding)
A|B < B (the coercoin A|B -> B will succeed only if the message contains a B encoding)
Typed schemas implement a limited form of alternative encodings via the Alternative
instance for record fields. In the future a similar 'Alternative' instance for union constructors could be added.
Example