safe-json

Automatic JSON format versioning

This library aims to make the updating of JSON formats or contents, while keeping backward compatibility, as painless as possible. The way this is achieved is through versioning and defined migration functions to migrate older (or newer) versions to the one used.

---

• Why?
  • Some effects of changing a message format
• How does it work
  • SafeJSON
    • Version
    • Kind
    • Type name
    • safeFrom and safeTo
  • Migrate
    • Reverse migration
• Keep in mind
  • Testing
  • Using noVersion
  • Using setVersion and removeVersion
  • Non-object versioning
• Examples
  • Fresh start
  • Production migration
    • No down-time cheatsheet
• Acknowledgments

---

Why?

An obvious example would probably be JSON messages used in a production environment for communication between (micro-)services.

In a long running setting, there will most likely come a moment that the JSON message in question will need to be restructured, updated, or otherwise changed. This can result in some mild or more serious headaches, depending on the architecture of the communicating services.

Some effects of changing a message format

• If the messages are being logged (e.g. to a database) and are used in any way in the normal operation of the services, this results in “old format” messages either needing to be excluded from queries after the change to the new format, or all messages needing to be migrated over to the new format manually.

• If downtime is undesirable (which is often the case), then an in-place update of services will cause “new format” messages to be received by the “old format” expecting services, and vice-versa. This might be countered with creating new endpoints for the new format and making sure the messages are routed correctly, etc. etc. But this still includes overhead when the semantics of the endpoints didn’t actually change.

Why can’t you just update the message format and leave everything as is? That is the goal of this library.

How does it work

The library mainly consists of two classes:

• SafeJSON a: Defines the version of a and if (and how) it is migratable.
• Migrate a: Defines the data type (MigrateFrom a) that can be migrated to a and how to migrate from that type.

SafeJSON

The SafeJSON class defines the following:

• version: a unique identifier
• kind: if and how to migrate to this type
• typeName: string to identify the type
• safeFrom: how to parse from a JSON value
• safeTo: how to parse to a JSON value

The default implementations of those last two use parseJSON and toJSON, so if you already have FromJSON and ToJSON instances for your type, you don’t have to define them. (There are some exceptions: read Version / safeFrom/-To)

safeFrom and safeTo can not be used directly! The functions that implement the versioning and migrating are safeFromJSON and safeToJSON, which use safeFrom and safeTo internally.

So given that your type already has FromJSON and ToJSON instances (and is Typeable), the most basic definition of a SafeJSON instance would be the following:

instance SafeJSON MyType

This will set the version tag as 0 and assumes this type can only be migrated forward in the migration chain.

---

Version

The version can be set using integer literals. The only requirement is that no two types in a chain share the same version number. (It is only used as a unique identifier.)

The implication is that, when using safeToJSON, the resulting JSON will have an additional version field. How it’s added depends on the format.

If the resulting JSON is an object (which is the most likely), an extra field will be added called "!v" with the number as the value.

{
  "type": "my custom type",
  "someValues": [{},{"testing":true}],
  "!v": 1
}

If the resulting JSON is not an object, it will be wrapped in the following 2-field object:

{
  "~v": 2,
  "~d": "my non-object type"
}

The fields ("!v", "~v" and "~d") are chosen to be the least likely used in any conventional setting, and as such are least likely to clash with any existing JSON formats. safeFromJSON depends on these fields to recover the version number, so any potential clashes should be avoided. (If required, this can be accomplished by adjusting the safeFrom and safeTo methods)

It is possible to omit a version tag, this is not advised, but might be needed for integrating types that have been used before using SafeJSON and therefore have no version field.

Kind

There are four different kinds a type can be:

• base: This type will not be migrated to when parsing. This can be seen as the bottom of the migration chain.
• extension: This type has at least one older version it can migrate from: the type defined as the MigrateFrom in the Migrate instance.
• extended_base: This type has at least one newer version it can reverse migrate from, and none it can regularly migrate from: this newer type is defined as MigrateFrom (Reverse a). (cf. Reverse migration)
• extended_extension: This type has at least one newer and one older version it can migrate from. (cf. Migrate and Reverse migration)

A chain of extensions makes the backward compatibility work. If a type is trying to be parsed using safeFromJSON, all older version numbers will be able to be parsed, and subsequently migrated to the current version.

If a type is also extended_*, that means the next/future version will be tried before going down the chain. This is convenient when old programs might receive the new type before being phased out, while still being able to handle the data coming in.

Type name

The name of the type is used in the reporting of errors and when making the objectProfile. A set of pre-made functions can be used to easily define this method. (typeName0-5)

safeFrom and safeTo

If the type already has FromJSON and ToJSON instances, the default definition will just use those. But if you’re only going to use the SafeJSON variants, the parsing from and to JSON can be defined in the safeFrom and safeTo methods. There might also be times when it is preferable to have the SafeJSON parsing be different from the From-/ToJSON parsing.

While using safeFromJSON in a parseJSON definition is completely valid, it can be desirable to only have versioned sub-parsing (parsing of versioned values inside other values) happen when using the SafeJSON interface. In those cases, you would have to define the safeFrom and safeTo methods. (and using safeFromJSON and safeToJSON in those definitions where appropriate)

When defining safeFrom and safeTo, you need to use the contain function.

Since version 1.0.0, a few convenience functions have been added to make defining safeFrom and safeTo methods a lot easier, and to make the experience more similar to defining parseJSON and toJSON. (e.g. containWithObject, .:$, .=$, etc.)

---

Migrate

The Migrate class is where the magic happens; here you define which type can be converted into which other type.

When defining a migration from an older type to a newer type, it’s as easy as defining:

data OldType = OldType Text
data NewType = NewType [Text]

instance Migrate NewType where
  -- Read this as:
  -- > type MigrateFrom a = b
  -- "The type you migrate from to the 'a' is 'b'"
  --   or
  -- "'b' is the type you migrate from to get 'a'
  type MigrateFrom NewType = OldType
  -- Here we define how to change the structure
  -- from one type to the other.
  migrate (OldType txt) = NewType [txt]

Now, whenever JSON is encountered that should be parsed as an OldType, we can parse it as such, and then immediately migrate it to NewType, which is the one the program actually uses.

Do not forget to set the kind of NewType to either extension or extended_extension to make use of this type of migration.

Reverse migration

There is also the option to support a migration from one version higher up in the chain to the current version. This is done by defining the kind of the type in the SafeJSON instance as one of the extended_* kinds and defining a Migrate (Reverse a) instance for the current type. In that case, a definition might look something like this:

-- (using the above data definitions of OldType and NewType)
instance Migrate (Reverse OldType) where
   type MigrateFrom (Reverse OldType) = NewType
   migrate (NewType [])    = Reverse $ OldType ""
   migrate (NewType (t:_)) = Reverse $ OldType t

N.B.: At the moment there is no support for extended migrating from newer versions further than the one directly defined in the type’s reverse migrate instance, i.e. if the parsing of the type defined in type a’s MigrateFrom (Reverse a) fails, the other attempts will go down the chain, not further up.

Keep in mind

Here are some points to take note of when using this library.

Testing

The module Data.SafeJSON.Test contains very useful testing functions to ensure your types and instances are consistent. It is advised to at least testConsistency of any type you create a SafeJSON instance for. This makes sure you don’t have any inconsistencies in your migration chain, which would result in failed parsing of your type(s).

I would also advise to use migrateRoundTripProp or migrateReverseRoundTripProp whenever you add new migrations, to be sure they are parsable, and that you didn’t forget to adjust the kind definition in the SafeJSON instances.

Using noVersion

There is a way to omit the version tag. It’s by using noVersion instead of an integer literal when defining the version method. This is used to not add a version tag to ‘primitive’ values (Int, Text, [], etc.), and also to give the possibility of adding a version-less format to a migration chain.

When switching to SafeJSON, you might already have a JSON format in use. Adding this type as the bottom of a chain is as easy as:

instance SafeJSON VersionLessType where
  version = noVersion
  kind = extended_base -- or 'base', if you don't need the forward migration

If you include a ‘noVersion’ (vNil) in your chain, it is advised to remove the need to include it as soon as possible; since, if the JSON being parsed has no version (because it might be a completely different message), having a noVersion type in your chain alone will make it try to parse it as such. In some cases this might lead to a succesful parse, even though it’s a completely different JSON message. For that reason, it is advised to remove the vNil from your chain as soon as possible.

As long as there is a version number in the JSON, though, vNil will not be attempted to be parsed, since “a version” doesn’t match “no version”.

Using setVersion and removeVersion

These functions are new in safe-json-1.0.0, since the FromJSON and ToJSON constraints have been dropped from SafeJSON’s definition and thus it is not guaranteed you can use Data.Aeson functions to handle versionless JSON Values.

CAUTION: Use these functions at your own risk! It is always best to use versioning if possible!

• setVersion will insert/override the given type’s version in the given JSON Value.
• removeVersion will remove all the SafeJSON versioning from the JSON Value

setVersion

Of course, you’d always like to have the correct version present in your Values, but sometimes this is not appropriate or desireable. One example would be when parsing incoming JSON from third parties (e.g. customers), which you don’t want to impose the SafeJSON versioning onto.

Be warned this does only set the version at the top-level! setVersion does not recursively set versions! (i.e. if your safeFrom definition, for example, uses safeFromJSON when parsing certain fields, these fields will not get the correct version from using setVersion on the overall JSON Value)

In these cases, it is recommended to use a FromJSON instance (which doesn’t use safeFromJSON in its definition) for your incoming type, instead of SafeJSON. You can still use SafeJSON internally, because if the type has a FromJSON instance, the SafeJSON instance can just use that implementation for the safeFrom definition.

This way you are guaranteed the only difference between safeFromJSON and parseJSON is the requirement of a version field in safeFromJSON’s case and it removes the need to use setVersion (which is preferable).

GIVEN:

data MyType
instance SafeJSON MyType where
  version = 0

-- toJSON adds no version fields
incomingJSON = toJSON [MyType, MyType]

# WRONG: this will not parse using SafeJSON functions.

λ> encode $ setVersion @MyType incomingJSON
{
  "~v": 0,
  "~d": [
    array_of_unversioned_MyTypes
  ]
}

# RIGHT: This will parse using SafeJSON functions.

λ> Just vals = parseMaybe safeFromJSON/parseJSON incomingJSON :: Maybe [Value]
λ> encode $ setVersion @MyType <$> vals
[
  {
    "my_type_field1": xxx,
    "my_type_field2": xxx,
    "!v": 0
  },
  {
    "my_type_field1": xxx,
    "my_type_field2": xxx,
    "!v": 0
  },
]

removeVersion

Conversely, removeVersion should, ideally, only be used when the JSON is leaving your application or platform and you don’t want the SafeJSON versioning to be visible to the outside world.

removeVersion, on the other hand, does remove all version fields recursively, so removeVersion . safeToJSON will produce a JSON Value with all "!v", "~v" and "~d" fields removed.

Non-object versioning

As described in Version, the version tag added to a non-object Value has more relative overhead than the tag added to a JSON object. (min. 14 bytes and min. 7 bytes, respectively)

To keep general overhead low, it is advised to version your entire message, and only version individual fields if really necessary.

Examples

I want to give two example use cases for SafeJSON. The first is a fresh start with SafeJSON and how you can migrate different versions at once. The second is starting from a JSON format without versioning and then migrating into SafeJSON in a production setting.

Fresh start

This is an arbitrary example; some things might seem contrived.

The data types we’re working with:

data FirstType = FirstType Text
  deriving (Eq, Show)

data SecondType = SecondType (Text, Maybe Int)
  deriving (Eq, Show)

data ThirdType = ThirdType {
  ttFirstName :: Text
  ttLastName :: Text,
  ttAge :: Int,
} deriving (Eq, Show)

FromJSON and ToJSON instances are included at the bottom of this example.

Backstory

We’ve started using our program with FirstType, and that went well for a couple of weeks. Then we wanted to add a field to maybe include the age of whatever the data was (if it was something that has an age), this is the SecondType. A while after, we noticed we only had data on people, so why not represent it as such. (ThirdType)

JSON formats with versions added:

// FirstType
{
  "type": "myType",
  "data": "Johnny Doe",
  "!v": 0
}

// SecondType
{
  "type": "myType",
  "name": "Johnny Doe",
  "age": 27,
  "!v": 1
}

// ThirdType
{
  "type": "myType",
  "firstName": "Johnny",
  "lastName": "Doe",
  "age": 27,
  "!v": 2
}

---

SafeJSON instances:

-- This sets 'version' to '0', and 'kind' to 'base'
instance SafeJSON FirstType

instance SafeJSON SecondType where
  version = 1
  kind = extension

instance Migrate SecondType where
  type MigrateFrom SecondType = FirstType
  migrate (FirstType name) = SecondType (name,Nothing)

import qualified Data.Char as C
import qualified Data.Text as T

instance SafeJSON Thirdtype where
  version = 2
  kind = extension

instance Migrate ThirdType where
  type MigrateFrom ThirdType = SecondType
  migrate (SecondType (name, mAge)) = ThirdType {
      ttFirstName = firstName,
      ttLastName  = lastName,
      ttAge       = fromMaybe (-1) mAge
    }
    where (firstName,rest) = T.break C.isSpace name
          lastName = T.dropWhile C.isSpace rest

---

Our database in which we saved our data in JSON now has three different JSON formats, depending on how far back in time we go. But that’s no problem for us! We just request all the JSON from the database, which result in a [Value], and if we want to use it immediately, just use something like the following:

parseValues :: [Value] -> Either String [ThirdType]
parseValues = mapM $ parseEither safeFromJSON

But we can also use the [Value] in the response body of a HTTP request when requested by a different program, and that program can just use a function from the Data.Aeson.Safe module, like eitherDecode, to parse the ByteString body:

{-# LANGUAGE TypeApplications #-}
import Data.Aeson.Safe as Safe

foo = do
  res <- httpLbs theRequest
  -- Since we've defined the instance for `SafeJSON` and `Migrate`
  -- `FirstType` and `SecondType` can now migrate to `ThirdType`
  -- Resulting in `Right [ThirdType]` in this transformation.
  case Safe.eitherDecode @[ThirdType] $ responseBody res of
    Left err  -> putStrLn $ "bad value in response: " ++ err
    Right tts -> thirdTypeHandler tts

thirdTypeHandler :: [ThirdType] -> IO ()
thirdTypeHandler = ...

The HTTP response would maybe look something like this:

[
  {
    "type": "myType",
    "data": "Johnny Doe",
    "!v": 0
  },
  {
    "type": "myType",
    "name": "Jonathan Doe",
    "age": null,
    "!v": 1
  },
  {
    "type": "myType",
    "name": "Shelley Doegan",
    "age": 27,
    "!v": 1
  },
  {
    "type": "myType",
    "firstName": "Anita",
    "lastName": "McDoe",
    "age": 26,
    "!v": 2
  }
]

Which would result in the following Haskell data:

[ ThirdType {ttFirstName = "Johnny"  , ttLastName = "Doe"   , ttAge = -1}
, Thirdtype {ttFirstName = "Jonathan", ttLastName = "Doe"   , ttAge = -1}
, Thirdtype {ttFirstName = "Shelley" , ttLastName = "Doegan", ttAge = 27}
, Thirdtype {ttFirstName = "Anita"   , ttLastName = "McDoe" , ttAge = 26}
]

FromJSON and ToJSON instances

instance ToJSON FirstType where
  toJSON (FirstType txt) = object
      [ "type" .= String "myType"
      , "data" .= txt
      ]

instance FromJSON FirstType where
  parseJSON = withObject "FirstType" $ \o -> do
      typ  <- o .: "type"
      guard $ typ == String "myType"
      val <- o .: "data"
      return $ FirstType val

instance ToJSON SecondType where
  toJSON (SecondType (name, age)) = object
      [ "type" .= String "myType"
      , "name" .= name
      , "age"  .= age
      ]

instance FromJSON SecondType where
  parseJSON = withObject "SecondType" $ \o -> do
      typ  <- o .: "type"
      guard $ typ == String "myType"
      name <- o .: "name"
      age  <- o .:? "age"
      return $ SecondType (name, age)

{-# LANGUAGE RecordWildCards #-}

instance ToJSON ThirdType where
  toJSON ThirdType{..} = object
      [ "type" .= String "myType"
      , "firstName" .= ttFirstName
      , "lastName"  .= ttLastName
      , "age"       .= ttAge
      ]

instance FromJSON ThirdType where
  parseJSON = withObject "ThirdType" $ \o -> do
      typ  <- o .: "type"
      guard $ typ == String "myType"
      ttFirstName <- o .: "firstName"
      ttLastName  <- o .: "lastName"
      ttAge       <- o .: "age"
      return Thirdtype{..}

Production migration

This is an arbitrary example; some things might seem contrived.

We’ve started using JSON (without versioning) as a messaging format between live services. Adding to it has been easy, but we’ve hit a point where we need to change the format in a way that current services would not be able to parse them.

The data type already in production:

{
  "id": "00000000-0000-0000-0000-000000000000",
  "command": "add_user",
  "person": {
    "firstName": "John",
    "middleName": null,
    "lastName": "Doe"
  },
  "age": 45,
  "address": {
    "street": "Steenstraat",
    "number": "25",
    "addition": "A",
    "city": "Koekel",
    "country": "Friesland"
  },
  "phoneNumber": null
}

data Message = Message {
  mId :: UUID,
  mCommand :: Text,
  mPerson :: Person,
  mAge :: Int,
  mAddress :: Address,
  mPhoneNumber :: Maybe PhoneNumber
} deriving (Eq, Show)

FromJSON and ToJSON instances are included at the bottom of this example.

---

The format we want to change to:

{
  "!v": 0,
  "id": "00000000-0000-0000-0000-000000000000",
  "command": "add_user",
  "data": {
    "person": {
      "firstName": "John",
      "middleName": null,
      "lastName": "Doe"
    },
    "age": 45,
    "address": {
      "street": "Steenstraat",
      "number": "25",
      "addition": "A",
      "city": "Koekel",
      "country": "Friesland"
    },
    "phoneNumber": null
  }
}

We’ll represent this as the first versioned type of this message.

data Message_v0 = Message_v0 {
  msgId :: UUID,
  msgCommand :: Text,
  msgData :: PersonalInfo
} deriving (Eq, Show)

data PersonalInfo = PersonalInfo {
  piPerson :: Person,
  piAge :: Int,
  piAddress :: Address,
  piPhoneNumber :: Maybe PhoneNumber
} deriving (Eq, Show)

We’ll then create the SafeJSON and corresponding Migrate instances:

{-# LANGUAGE RecordWildCards #-}

instance SafeJSON Message where
  -- | This is important, since our old type has no version tag
  version = noVersion
  -- | extended_* makes sure we can migrate from the newer version
  --   back to this one, since the newer formats will start going
  --   through the system the moment the new services are deployed
  --   and we want the older services to keep functioning.
  kind = extended_base

instance Migrate (Reverse Message) where
  type MigrateFrom (Reverse Message) = Message_v0
  migrate Message_v0{..} = Reverse $
    Message
      msgId
      msgCommand
      piPerson
      piAge
      piAddress
      piPhoneNumber
    where PersonalInfo{..} = msgData

instance SafeJSON Message_v0 where
  version = 0
  -- | 'extension' ensures the new services will be able to
  --   handle any old formats still floating around.
  kind = extension

instance Migrate Message_v0 where
  type MigrateFrom Message_v0 = Message
  migrate Message{..} =
      Message_v0 mId mCommand person
    where person = PersonalInfo
              mPerson
              mAge
              mAddress
              mPhoneNumber

---

Assuming we’ve added Message_v0 on a new development branch, and modified the business logic to use Message_v0, at this point we create a temporary branch from the production branch (probably master) and let’s call it something like master-message-migration.

On this branch, we only add Message_v0 with the SafeJSON and Migrate instances. And replace the JSON functions with SafeJSON ones for anywhere Message is received by or sent to the current services. (i.e. use decode/encode from Data.Aeson.Safe instead of Data.Aeson, or safeFromJSON/safeToJSON instead of parseJSON/toJSON)

Then we update the current services so they’re ready to use SafeJSON for migrating the new JSON formats we’re expecting. After the services that have SafeJSON implemented are the only ones running, we can roll out the new format without fear of anything throwing parsing errors.

After that, we rename Message to Message_old and all usages of Message_v0 to Message and we’re back to using Message in our code, but now with a new structure. All while not having to worry about rerouting messages or down-time.

FromJSON and ToJSON instances

{-# LANGUAGE RecordWildCards #-}

instance ToJSON Message where
  toJSON Message{..} = object
      [ "id"          .= mId
      , "command"     .= mCommand
      , "person"      .= mPerson
      , "age"         .= mAge
      , "address"     .= mAddress
      , "phoneNumber" .= mPhoneNumber
      ]

instance FromJSON Message where
  parseJSON = withObject "Message" $ \o -> do
    mId      <- o .: "id"
    mCommand <- o .: "command"
    mPerson  <- o .: "person"
    mAge     <- o .: "age"
    mAddress <- o .: "address"
    mPhoneNumber <- o .:? "phoneNumber"
    return Message{..}

instance ToJSON Message_v0 where
  toJSON Message_v0{..} = object
      [ "id"      .= msgId
      , "command" .= msgCommand
      , "data"    .= msgData
      ]

instance FromJSON Message_v0 where
  parseJSON = withObject "Message_v0" $ \o ->
      msgId      <- o .: "id"
      msgCommand <- o .: "command"
      msgData    <- o .: "data"
      return Message_v0{..}

instance ToJSON PersonalInfo where
  toJSON PersonalInfo{..} = object
      [ "person"      .= piPerson
      , "age"         .= piAge
      , "address"     .= piAddress
      , "phoneNumber" .= piPhoneNumber
      ]

instance FromJSON PersonalInfo where
  parseJSON = withObject "PersonalInfo" $ \o ->
      piPerson      <- o .: "person"
      piAge         <- o .: "age"
      piAddress     <- o .: "address"
      piPhoneNumber <- o .: "phoneNumber"
      return PersonalInfo{..}

---

No down-time cheatsheet

From not using SafeJSON to using SafeJSON:

• Add SafeJSON instance to MyType.
  • OPTIONAL: use noVersion when previous JSON is already used in production
• Switch Data.Aeson functions for the Data.Aeson.Safe ones
  • Preferably everywhere in the codebase, unless explicitly needed for other purposes.
• At this point everything should still work like before.

Create a new development branch to keep the following changes separate from what’s running on the servers. Do the following on the new branch:

• Rename MyType to MyType_old:
  • type definition
  • FromJSON/ToJSON instances
  • SafeJSON instance
• Add the type with the new JSON representation
  • name it MyType
  • And it’s FromJSON/ToJSON/SafeJSON instances
• At this point you can change your business logic to use the new type.
• change kind methods of both SafeJSON instances to make both types migrate from eachother.
  • kind of MyType_old: extended_base
  • kind of MyType: extension
• Define Migrate instances for both types:
  • Migrate (Reverse MyType_old)
  • Migrate MyType
• At this point you have your new updated code ready for use.

Copy the definition of the MyTypes with their *JSON and Migrate instances to the original branch and overwrite the type definition and instances of the original MyType. Do the following only on the code you just copied.

• Rename MyType to MyType_v0 (or MyType_new or something)
• Rename MyType_old to MyType.
• At this point everything should still work like before.

Now you should have two branches. One with the code that’s still running on your servers (to which we added the new type, its instances and the migration instances to migrate between the two), and one branch with the new code that will use the new type.

• Use the original branch to update your running services to make them ready to migrate from the new JSON formats to the ones currently used.
• After all services are running using the new SafeJSON code, update your services with the code on the new branch.

Enjoy a migration where all services keep parsing all JSON they receive.

---

Acknowledgments

The core of this library is inspired by the safecopy library by David Himmelstrup and Felipe Lessa, found on GitHub, Hackage and Stackage