Implementation:Tensorflow Serving Simple Servers

Knowledge Sources	Tensorflow_Serving
Domains	Model Serving, Server Bootstrap
Last Updated	2026-02-13 00:00 GMT

Overview

Provides a simple bootstrapping function that creates a complete serving Manager for a single TensorFlow SavedModel with automatic version loading from a filesystem path.

Description

The Simple Servers module provides CreateSingleTFModelManagerFromBasePath, a convenience function that constructs a complete model serving pipeline with default configurations. It wires together three components:

1. FileSystemStoragePathSource: Monitors a base path on the filesystem for new model version directories, polling every second. New directories are detected and reported as aspired versions.

2. SavedModelBundleSourceAdapter: Converts storage paths into loadable SavedModelBundles with default configuration, acting as the bridge between the path source and the model manager.

3. AspiredVersionsManagerBuilder: Creates an AspiredVersionsManager with an AvailabilityPreservingPolicy, which ensures that the old version is unloaded before the new version is loaded. The builder connects the source chain (path source feeding into bundle source adapter) and produces the final Manager.

The created Manager serves SavedModelBundle servables under the fixed name "default". The AvailabilityPreservingPolicy trades availability for reduced resource usage by not keeping two versions loaded simultaneously.

Usage

Use this module for simple, single-model deployments where minimal configuration is needed. It is intended for testing, prototyping, and basic experimental deployments. For production deployments, use ServerCore with full configuration options instead.

Code Reference

Source Location

• Repository: Tensorflow_Serving
• Files:
  • tensorflow_serving/servables/tensorflow/simple_servers.h (lines 1-56)
  • tensorflow_serving/servables/tensorflow/simple_servers.cc (lines 1-101)

Signature

namespace simple_servers {

Status CreateSingleTFModelManagerFromBasePath(
    const string& base_path, std::unique_ptr<Manager>* manager);

}  // namespace simple_servers

Import

#include "tensorflow_serving/servables/tensorflow/simple_servers.h"

I/O Contract

Inputs

Name	Type	Required	Description
base_path	string	Yes	Filesystem path where model version directories are located

Outputs

Name	Type	Description
manager	std::unique_ptr<Manager>*	Created Manager that automatically loads/unloads model versions from base_path
return	Status	OK on success; error if source or adapter creation fails

Usage Examples

Creating a Simple Server

std::unique_ptr<Manager> manager;
TF_RETURN_IF_ERROR(
    simple_servers::CreateSingleTFModelManagerFromBasePath(
        "/models/my_model", &manager));

// Get a handle to the loaded model
ServableHandle<SavedModelBundle> bundle;
TF_RETURN_IF_ERROR(manager->GetServableHandle(
    ServableRequest::Latest("default"), &bundle));

Related Pages

• Principle:Tensorflow_Serving_Simple_Server_Bootstrap