Hi Valentyn, Thank you for the proposal/summary. Leo Meyerovich and others have previously suggested adding Arrow support to TinkerPop; it just hasn't been prioritized. I like everything about your description apart from this phrase: "should replace the network layer with Arrow Flight". You are not suggesting that the WebSocket-based solution be removed, are you? If the two could exist in parallel, it definitely would be nice to have an Arrow option. WebSocket could perhaps be dropped later if it isn't being used much and/or the maintenance burden is too high. Just my $0.02.
Josh On Thu, Jun 30, 2022 at 4:36 PM Valentyn Kahamlyk <[email protected]> wrote: > Hello Everyone, > > I would like to propose exploring options to use Arrow Flight as a transport > for Gremlin Server. Currently Gremlin Server and Clients are based on > WebSockets with a custom sub-protocol and serialization to GraphSON and > GraphBinary. Developers for each driver must implement those protocols from > scratch and there is a limited amount of code which is being reused (only 3rd > party WebSocket libraries are currently reused in the client variants). The > protocol implementation is a complicated and error-prone process, so most > drivers only support some subset of Gremlin Server features. The maintenance > cost is also constantly increasing with the number of new client variants > being added to TinkerPop. > > ** Motivation ** > We would like to propose a solution to reduce maintenance and simplify the > development of the client drivers by using a standard protocol based on the > Apache Arrow Flight. As Arrow Flight is implemented in the most common > languages like C++, C#, Java and Python we anticipate a larger amount of > existing codebase can be reused which would help to reduce maintenance costs > in the future. Also, we can reuse some other Arrow Flight features like > authentication and error handling. > > ** Assumptions ** > Proof of Concept Development will be done with Java 8. > Need to reuse existing code as much as possible. > It is desirable, but not necessary, to maintain compatibility with existing > drivers. > To simplify development at the initial stage, we will reuse existing > serialization mechanisms. > > ** Requirements ** > Gremlin Server and drivers should replace the network layer with Arrow Flight. > No significant drop in performance. > Gremlin Arrow must pass the Gherkin test suite. > > ** Prototype Design Overview ** > We would like to explore solution below and create prototype to prove > approach is feasible. > The main idea is to replace the transport layer with FlightServer and > FlightClient. They support asynchronous data transfer, splitting data into > chunks, and authorization. While Arrow Flight typically requires schema, in a > short term we can proceed with implementation using existing serializers and > GraphBinary format. By using GraphBinary we will not have all capabilities > that Arrow Flight provides out of the box, like efficient compression. > However, in the future, we see the value of adding capabilities to generate a > schema from the server-side, and that can enable additional use cases. > > First stage: replace transport layer, but keep serializers > Pros: > Reduction of the code base to be developed and maintained > A relatively low number of modifications > > Cons: > We may observe reduced performance due to schema transfer and other overhead. > As part of the PoC we will assess performance overhead for small and large > responses and identify options to mitigate it. > Still need to support GraphBinary serialization. > > Second stage: replace transport layer, make dynamic schema generation and use > native Arrow structures for data transmission > Pros: > Greater reduction of the codebase to be developed and maintained > In addition, need to rework the serialization and add schema generation > Performance can be improved for large data sets due to Arrow Flight > optimizations and the ability to transfer data in parallel > No need to support GraphBinary and GraphSON serialization protocols > > Cons: > Reduced performance for small result sets > Can be complicated and expensive to generate a schema for each request > > Please find few more diagrams attached in the pdf file attached and please > share your thoughts. > > Regards, Valentyn > >
