Development With Client SDK

Client SDK is the most efficient approach to develop pipelines, run quality assurance tests to validate the pipeline behavior when models are under development, and to develop custom sources and sinks.

Client SDK together with DevServer and OpenTelemetry provide a complete set of technologies required to develop efficiently. Let us discuss why:

  • Client SDK allows sending a single frame from Python and retrieving the corresponding result. Because it is a Python code developers can use OpenCV to display the resulting frame, dumping meta to JSON, etc.

  • The Client SDK source sends every frame with OpenTelemetry propagated through the pipeline to the Client SDK sink, so developers can trace the code end-to-end and match frames sent with frames retrieved by Trace ID.

  • Client SDK can retrieve and display logs from Jaeger by a trace ID.

  • You can run the code directly from your IDE.

  • You can analyze code instrumenting in Jaeger Web GUI for every single query, which is beneficial for code optimization and debugging.

OpenTelemetry in Client SDK

Savant supports OpenTelemetry for collecting traces and logs. Frames, delivered to a module, can optionally include OpenTelemetry propagated information, which allows instrumenting pipelines precisely.


In practice, we recommend using sampled OpenTelemetry traces; however, Client SDK creates a trace for every single frame sent to the module.

Client SDK integrates with OpenTelemetry, which means that every time a frame is sent to a module, it contains a new trace, thus the trace causes the module to instrument its internal parts, creating necessary OpenTelemetry spans in the pipeline, including pyfuncs. The developer can create auxiliary spans within her code to find out what happens. Please, refer to OpenTelemetry Support for details.

During the frame processing ongoing logs are attached to the currently active span and thus collected and associated with a specific trace. The developer also can attach additional information to the trace by calling corresponding span methods.

When the frame is sent, the savant.client.runner.source.SourceResult is returned. The developer can retrieve the trace_id from it for matching sent and delivered frames.


The trace_id can also be used to observe the trace in the Opentelemetry management system like Jaeger.

When the frame processing result is retrieved from the module, the developer can request the frame, metadata and logs collected by OpenTelemetry.

Currently we support only Jaeger OpenTelemetry collector. Logs are fetched from Jaeger REST API.

Remote Development

Client SDK enables remote development, allowing running programmatic sources and sinks locally, while processing the data remotely. It can be convenient in the following cases:

  • develop on a host machine without GPU;

  • develop on remote hardware with no KVM access (e.g. Jetson or dGPU in a datacenter).

To utilize the full power of Client SDK it must be paired with:

To find out more, explore Module Development Guide.

Source Example

Sources ingest frames and their metadata to a running module.


Currently, Client SDK supports only JPEG or PNG sources, but you can implement your own source based on savant.client.JpegSource or savant.client.PngSource.

import time
from savant_rs import init_jaeger_tracer
from savant.client import JaegerLogProvider, JpegSource, SourceBuilder

# Initialize Jaeger tracer to send metrics and logs to Jaeger.
# Note: the Jaeger tracer also should be configured in the module.
init_jaeger_tracer('savant-client', 'localhost:6831')

# Build the source
source = (
    # Note: healthcheck port should be configured in the module.

# Send a JPEG image from a file to the module
result = source(JpegSource('cam-1', 'data/AVG-TownCentre.jpeg'))
time.sleep(1)  # Wait for the module to process the frame

Sink Example

Sinks retrieve results from a module.

from savant.client import JaegerLogProvider, SinkBuilder

# Build the sink
sink = (
    # Note: healthcheck port should be configured in the module.

# Receive results from the module and print them
for result in sink: