Background
Protocol
We've been writing a new client-server protocol as a public API for the creation of new Geode clients. We settled on using Protobuf as the encoding for the protocol as it makes writing clients easier by abstracting away a lot of the encoding details.
Encoding
One of the big challenges in designing a new protocol has been how to encode values. Like the old binary client protocol, the PDX encoding is complicated, underdocumented, and stateful. However, we need a way for users to send values that are more complex than mere primitives or maps of primitives..
The first approach was to use the JSON-PDX conversion that is already used for the REST API. Many languages have libraries to encode objects as JSON, and it's a familiar format to many developers. However, using JSON for encoding has downsides, in that it's large and slow.
Selecting an Encoding Mechanism
Adding this encoding mechanism to the protocol is not exclusive with allowing JSON or custom encodings – the object encoding can be pluggable, and the user's desired response encoding should be selected during the handshake process.
Regardless of proposal, we should allow users to have a pluggable object encoding that they can register a handler with on the server. This encoder will receive a byte array and return an Object
. This allows users to do custom serialization if desired.
Protobuf Struct Encoding and Extension
This section is intended as useful background on the start of some thoughts about encoding a PDX-like type with Protobuf; for the proposed encoding, see "The Proposed Encoding", below.
Protobuf ships with a Struct
message, defined in struct.proto
, that can be recursively nested to encode JSON.
By extending the same sort of structure, we can encode almost any value with a type that is supported, including adding support for more complex types like dates or UUIDs. This should make it possible to write serializers and should enable driver developers to write auto-serializers that will serialize these objects via reflection to Protobuf.
The typeName
field can be used for other clients to recognize the same type. Internally, it will be stored in the PDXInstance that this is converted to, but that detail shouldn't need to be exposed to the driver developer.
So for example, the following class and value:
class User { String name; int age; } value = new User("Amy", 64);
would encode as (using a pseudo-static initializer syntax):
Struct{ typeName: "User", entries: [ StructEntry{ fieldName: "name", value: stringValue{"Amy"} }, StructEntry{ fieldName: "age", value: intValue{44} } ] }
This all gets compiled down to binary for an encoding that is more efficient than JSON.
Ideally, a driver developer would provide annotations or registration for application developers to specify the manner in which a type should be serialized. In languages that use setters and getters by convention, it would probably be more idiomatic to refer to setters getters for reflection rather than the member variables of the object.
The Proposed Encoding
As an optimization to the "Struct" encoding, we can caching metadata using type registration. This encoding
Type registration
As an optimization to avoid sending field names with every message, allow clients to register types to communicate the metadata for data they are about to send. The server will give back an ID for that datatype, and the ID can be used in future messages to refer to the metadata without retransmitting that metadata. This encoding will not actually be smaller for single values, but if multiple values of the same type are sent the savings can be significant.
Type registration will be per-connection (meaning IDs cannot be cached between connections). This eliminates the need to keep synchronization on the server, as well as decoupling type registrations from the internal details of PDX. It also means that the drivers only have to keep track of a relatively small amount of data.
The outline of type registration for the client is this:
- Send a type definition
- Get back a type ID that references the type description
- Use that type ID when encoding values of that type
So for example, using the same User from above:
Suppose the client chose ID 42 for this type. Then the first put message using such a value would have a value like so:
PutRequest{ key: EncodedValue{intValue: 12}, EncodedValue{ newStructType: NewStructType{ typename: "User", typeID: 42, fieldNames: ["name", "age"], fieldValues: ["Amy", "42"] } } }
A later PutRequest would encode the value like this (enclosing Request omitted for succinctness):
PutRequest{ key: EncodedValue{intValue: 111}, value: EncodedValue{ structById: StructById{ id: 42, fields: [ ValueField{stringField: "Amy"}, ValueField{intField: 64} ] } } }
Message Definitions
This is the proposed EncodedValue message that will contain values a client sends to the server or the server sends to the client:
Under this EncodedValue scheme, types defined by the server and types defined by the client will use different sets of IDs (though these can refer to the same cached values if they are the same). This is because we intend to add support for asynchronous messages and/or multiplexing of multiple channels of communication over one socket, and this avoids having the server and client race to assign IDs. If IDs were shared, the server would need to send back new IDs when it sent back types the client had not seen before.
The Object field is for fields that may be an Integer, String or Array type but have a broader type on the server side. Structs are viewed as Object type – more complex typing is not present. This is, in significant part, a leaky abstraction due to the way PDX saves values.
If a client is sending mutually recursive types or types that contain instances of themselves, it should send the type definition the first time one is seen (or in the parent instance) and send the type with ID in each later instance.
Whether a client must send all following values by ID or the values can be sent with a full ID each time
Considerations
In order to avoid arbitrary object serialization (which can lead to gadget chain exploits), we will probably need to constrain valid types to those registered as DataSerializable, or possibly even only those registered with the ReflectionBasedAutoSerializer
. This may also mean that we need a special class of typenames for those types that are put first by a client.
A driver developer may wish to provide a way for users to register types before sending values.
Driver developers will have to make sure that types they want to use in different language clients can be correlated. So package names may or may not make sense. The naming convention is not entirely decided, nor is whether we can register nameless types. It may be wise to reserve a set of names with special meaning ("JSON" perhaps?) and perhaps a set of names that would correspond to classes that have no domain class in Java (leading underscore, or just those with no package name?)
If a server sends back a value of a type a client has not registered, the client can send a TypeDefinitionLookupRequest.
The use of NumericArray for all the integral types is because they all have the same varint encoding and will encode the same way on the wire. It may be advisable to use more restricted types and separate messages to get better typing in the generated Protobuf code.