Paddle Serving supports remote service access through RPC(remote procedure call) and HTTP. RPC access of remote service can be called through Client API of Paddle Serving. A user can define data preprocess function before calling Paddle Serving's client API. The example below explains how to define the input data of Paddle Serving Client. The servable model has two inputs with alias name of `sparse` and `dense`. `sparse` corresponds to sparse sequence ids such as `[1, 1001, 100001]` and `dense` corresponds to dense vector such as `[0.2, 0.5, 0.1, 0.4, 0.11, 0.22]`. For sparse sequence data, current design supports `lod_level=0` and `lod_level=1` of Paddle, that corresponds to `Tensor` and `LodTensor`. For dense vector, current design supports any `N-D Tensor`. Users do not need to assign the shape of inference model input. The Paddle Serving Client API will check the input data's shape with servable configurations.
The following code sample shows that Paddle Serving Client API connects to Server API with endpoint of the servers. To use the data parallelism ability during prediction, Paddle Serving Client allows users to define multiple server endpoints.
Paddle Serving adopts [baidu-rpc](https://github.com/apache/incubator-brpc) as underlying communication layer. baidu-rpc is an open-source RPC communication library with high concurrency and low latency advantages compared with other open source RPC library. Millions of instances and thousands of services are using baidu-rpc within Baidu.
The core execution engine of Paddle Serving is a Directed acyclic graph(DAG). In the DAG, each node represents a phase of inference service, such as paddle inference prediction, data preprocessing and data postprocessing. DAG can fully parallelize the computation efficiency and can fully utilize the computation resources. For example, when a user has input data that needs to be feed into two models, and combine the scores of the two models, the computation of model scoring is parallelized through DAG.
The underlying communication of Paddle Serving is implemented with C++ as well as the core framework, it is hard for users who do not familiar with C++ to implement new Paddle Serving Server Operators. Another approach is to use the light-weighted Web Service in Paddle Serving Server that can be viewed as a plugin. A user can implement complex data preprocessing and postprocessing logics to build a complex AI service. If access of the AI service has a large volumn, it is worth to implement the service with high performance Paddle Serving Server operators. The relationship between Web Service and RPC Service can be referenced in `User Type`.
Distributed Sparse Parameter Indexing is commonly seen in advertising and recommendation scenarios, and is often used coupled with distributed training. The figure below explains a commonly seen architecture for online recommendation. When the recommendation service receives a request from a user, the system will automatically collects training log for the offline distributed online training. Mean while, the request is sent to Paddle Serving Server. For sparse features, distributed sparse parameter index service is called so that sparse parameters can be looked up. The dense input features together with the looked up sparse model parameters are fed into the Paddle Inference Node of the DAG in Paddle Serving Server. Then the score can be responsed through RPC to product service for item ranking.
Why do we need to support distributed sparse parameter indexing in Paddle Serving? 1) In some recommendation scenarios, the number of features can be up to hundreds of billions that a single node can not hold the parameters within random access memory. 2) Paddle Serving supports distributed sparse parameter indexing that can couple with paddle inference. Users do not need to do extra work to have a low latency inference engine with hundreds of billions of parameters.
### 3.2 模型管理、在线A/B流量测试、模型热加载
### 3.2 Model Management, online A/B test, Model Online Reloading
Paddle Serving's C++ engine supports model management, online A/B test and model online reloading. Currently, python API is not released yet, please wait for the next release.
Paddle Serving provides RPC and HTTP protocol for users. For HTTP service, we recommend users with median or small traffic services to use, and the latency is not a strict requirement. For RPC protocol, we recommend high traffic services and low latency required services to use. For users who use distributed sparse parameter indexing built-in service, it is not necessary to care about the underlying details of communication. The following figure gives out several scenarios that user may want to use Paddle Serving.
For servable models saved from Paddle Serving IO API, users do not need to do extra coding work to startup a service, but may need some coding work on the client side. For development of Web Service plugin, a user needs to provide implementation of Web Service's preprocessing and postprocessing work if needed to get a HTTP service.
Web Service has lots of open sourced framework. Currently Paddle Serving uses Flask as built-in service framework, and users are not aware of this. More efficient web service will be integrated in the furture if needed.
`WebService` is a Base Class, providing inheritable interfaces such `preprocess` and `postprocess` for users to implement. In the inherited class of `WebService` class, users can define any functions they want and the startup function interface is the same as RPC service.
In recommendation and advertisement systems, it is commonly seen to use vector based index or tree based indexing service to do candidate retrievals. These retrieval tasks will be built-in services of Paddle Serving.
