Skip to content

Release v2.0.0
Compare
Choose a tag to compare
@forgivedengkai forgivedengkai released this 24 Aug 08:18
· 157 commits to master since this release
v2.0.0
7fb7f81
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

  1. For single inference, the guest side and host side of version 2.0. * will be calculated in parallel, thus reducing the time consumption.

  2. Batch inference, which is a new feature introduced in version 2.0 *. To batch submit a batch of data to be predicted in one request, which greatly improves the throughput.

  3. Parallel computing: in version 1.3. * the inference of the guest side and the inference of the host side are serial. From version 2.0, the prediction of the guest side and the host side will adopt the method of parallel inference. The inference of each party can be divided into subtasks according to the number of features and then parallel computing.

  4. Introduce a new component serving-admin, which will provide the visual operation interface of cluster, including model management, traffic monitoring, configuration view, service management and so on.

  5. The new model persistence / recovery mode: when the service server is restarted, version 1.3. * uses the playback push model request to restore the model when the instance is restarted, and version 2.0. * uses the method of directly recovering the memory data to restore the model.

  6. Java SDK. With this SDK, you can use the service governance related functions of Fat-service, such as service automatic discovery and routing.

  7. In the new extension module, the user-defined development part (such as: host side feature acquisition adapter interface development) is put into this module, so as to separate from the core source code.

  8. Support a variety of caching methods. Fat-service strongly relies on redis in version 1.3 *. And no longer relies on redis since version 2.0 *. You can choose not to use cache, use local memory cache, and use redis.

  9. Change the internal prediction process, reconstruct the core code, remove the pre-processing and post-processing components, and use the unified exception handling. The algorithm component is no longer tightly coupled with the RPC interface.

  10. Provide command line tools to query configuration and model information

Assets 2
Loading