Introduce KeySpace interface for API v2.

[iosmanthus]

Signed-off-by: iosmanthus [email protected]

This pull request is still working in progress, and the interfaces for request encoding are stable. However, lots of the request structs need to implement with these interfaces.

The core interfaces to encode the request are:

/// Abstracts any request sent to a TiKV server.
#[async_trait]
pub trait KvRequest<C>: Request + Sized + Clone + Sync + Send + 'static {
    /// The expected response to the request.
    type Response: HasKeyErrors + HasLocks + Clone + Send + 'static;
    fn encode_request(&self, _codec: &C) -> Cow<Self>;
    fn decode_response(&self, _codec: &C, _resp: Self::Response) -> crate::Result<Self::Response>;
}

When the C is bound to RequestCodec, the requests are able to be encoded:

pub trait RequestCodec: Sized + Clone + Sync + Send + 'static {
    fn encode_key(&self, key: Vec<u8>) -> Vec<u8> { key }
    fn decode_key(&self, key: Vec<u8>) -> Result<Vec<u8>> { Ok(key) }
    fn encode_range(&self, start: Vec<u8>, end: Vec<u8>) -> (Vec<u8>, Vec<u8>) { (start, end) }
    fn encode_pd_query(&self, key: Vec<u8>) -> Vec<u8> { key }
    fn decode_region(&self, region: Region) -> Result<Region> { Ok(region) }
}

TODO:

• Compile pass for API V1
• Fix test for API v1
• Impl V2 codec
• Write tests for V2 codec.

[sunxiaoguang]

Maybe it should return reference to Self::RequestCodec?

[iosmanthus]

Since the only use of RequestCodec is in the construction of Dispatch, and it requires a move, I think Self::RequestCodec is more direct.

[sunxiaoguang]

I see. Then we can probably rename it to create_request_codec or new_request_codec which means it's actually a factory method. It's not going to be called many times, and move is the right semantic.

[sunxiaoguang]

Client itself can tell we are using TxnKV already, can we hide things like TxnApiV1, TxnApiV2, RawApiV1, RawApiV2. Just let user chose the appropriate client and use ApiV1 or ApiV2.

[iosmanthus]

We may reexport them in their own module, like: raw::ApiV1 , txn::ApiV1

[iosmanthus]

This might break the original API, we should add a default implementation for API V1.

[iosmanthus]

These macro rules have the same patterns which are to take some field out and then encode or decode it. We should extract them into one rule.

[iosmanthus]

@sunxiaoguang PTAL. This is the logic for constructing keyspace from the name.

[adamf]

I am very excited for this!