Refactor write method in WalManager to improve performance

[ygf11]

Description

After #159, we pass dyn Payload to write method of WalManager, this may has performance issue.
It is better to decouple encoder from write method.

Proposal
We can define WalEncoder(maybe other name), like:

pub enum WalEncoder {
    Rocks(RocksLogEncoding),
    Obkv(ObkvLogEncoding)
}

impl WalEncoder {
   pub fn encode_to_write_batch<P:Payload>(&self, log_batch: &LogWriteBatch<P>) -> Result<(WriteBatch, u64)> {
       match self {
         WalEncoder::Rocks(encoder) => encoder.encode(payload),
         WalEncoder::Obkv(encoder) => encoder.encode(payload),
       }
   }
}

Encoder can get from walManager.

pub trait WalManager {
   fn encoder(&self) -> Encoder;
   fn write(&self, wb: WriteBatch) -> Result<()>,
}

So we need encode data before write to WalManager.

   let (wb, max_sequence_num) = self.wal.encoder().encode::<XXXPayload>(log_batch)?;
   self.wal.write(wb);
   ...

Additional context

cc: @waynexia @ShiKaiWi

[waynexia]

Thanks for writing this up. Here are some thoughts:

WAL backend like RocksDB or OBKV is only responsible for persisting things. I think we need not dispatch the data encoding method on the backend. I.e., WAL's write method should operate on bytes (Vec<u8> or dyn Payload) rather than a specific data type (what we used to do). However, Payload::encode_to() is writing bytes to a mutable buffer, this lazy materialization may reduce some copies compare to pre-encode the data.

So the trade offs are: (a) accept the dynamic dispatch overhead and achieve lazy materialization or (b) remove dynamic dispatch and do encode work in advance.

At a glance both are all not large consumption. But I cannot figure out which is more efficient if I have to rank them. Maybe we need a benchmark to help us to made decision. Personally I prefer the second approach (also the way you proposed). It looks more neat to me 👀

[ygf11]

However, Payload::encode_to() is writing bytes to a mutable buffer, this lazy materialization may reduce some copies compare to pre-encode the data.

Yes, aggree. pre-encode has this drawback.

Let me try to add a benchmark first.

[ShiKaiWi]

At a glance both are all not large consumption. But I cannot figure out which is more efficient if I have to rank them. Maybe we need a benchmark to help us to made decision. Personally I prefer the second approach (also the way you proposed). It looks more neat to me 👀

Provided a high throughout reducing some extra copies may make a difference to the performance. However, let the benchmark decide.

Thanks @ygf11 .

[ygf11]

It seems the two LogEncoding are almost the same, can we use a common LogEncoding?

https://github.com/CeresDB/ceresdb/blob/8cccedd01b2ac33768a6678f133e51a394468bda/wal/src/rocks_impl/encoding.rs#L18-L24

https://github.com/CeresDB/ceresdb/blob/8cccedd01b2ac33768a6678f133e51a394468bda/wal/src/table_kv_impl/encoding.rs#L100-L106

[ShiKaiWi]

It seems the two LogEncoding are almost the same, can we use a common LogEncoding?

https://github.com/CeresDB/ceresdb/blob/8cccedd01b2ac33768a6678f133e51a394468bda/wal/src/rocks_impl/encoding.rs#L18-L24

https://github.com/CeresDB/ceresdb/blob/8cccedd01b2ac33768a6678f133e51a394468bda/wal/src/table_kv_impl/encoding.rs#L100-L106

Sure. It seems the LogEncoding in the table_kv_impl/encoding.rs is a copy from rocks_impl/encoding.rs.