Python adapter for Java code leveraging JPype.
| Requests/sec | Average (ms) | Slowest (ms) | Fastest (ms) | |
|---|---|---|---|---|
| NoJava (REST) | 547.3972 | 180.5 | 2778.4 | 8.4 |
| Baseline (REST) | 142.4239 | 701.4 | 1113.2 | 121.3 |
| ProtobufEncoding (REST) | 238.6726 | 418.6 | 651.1 | 63.1 |
| PayloadPassthrough (REST) | 537.7259 | 185.7 | 537.6 | 66.0 |
The NoJava wrapper is a special case which implements
the Java model in pure Python.
It doesn't use Java or JPype at all.
The Baseline wrapper hardcodes the conversion for a
particular input and output type between Python and Java.
That is, it explicitly reads / write data from the relevant fields and converts
it as required.
Note that this approach is not scalable as-is, since it only works when those
input and output types are used (e.g. won't work with strData).
The ProtobufEncoding wrapper encodes and
decodes the full input from / to bytes using the SeldonMessage protobuf
definition.
This encoding / decoding happens on both the Java and Python side.
The PayloadPassthrough wrapper passes the
input and output payloads as they are.
In other words, the Python side won't try to encode / decode the request /
response.
Instead, it will use the input / output formats as a communication protocol and
it will Java's responsibility to (de-)serialise it.
Note that this approach requires changes on the Python side.
You can check the java-wrapper-v2 branch in the
github.com/adriangonz/seldon-core
fork for a
prototype of the changes.