Implementation:Tensorflow Serving ServableVersionPolicy

Knowledge Sources	TensorFlow Serving TF Serving Config
Domains	Version_Management, Configuration
Last Updated	2026-02-13 17:00 GMT

Overview

Concrete tool for configuring which model versions are loaded, defined as a protobuf message and enforced by the FileSystemStoragePathSource.

Description

ServableVersionPolicy is a protobuf message that specifies version selection rules. It is set on the model_version_policy field of ModelConfig and propagated through ServerCore::CreateStoragePathSourceConfig() to the filesystem source. The source implements three aspiration strategies: AspireLatestVersions(), AspireAllVersions(), and AspireSpecificVersions().

Usage

Set the model_version_policy field in your model configuration protobuf or config file. This is configured at server startup via --model_config_file or dynamically via HandleReloadConfigRequest.

Code Reference

Source Location

• Repository: tensorflow/serving
• File: tensorflow_serving/config/file_system_storage_path_source.proto (L8-36)
• Enforcement: tensorflow_serving/sources/storage_path/file_system_storage_path_source.cc (L97-321)

Signature

message ServableVersionPolicy {
  message Latest {
    uint32 num_versions = 1;  // default: 1
  }
  message All {}
  message Specific {
    repeated int64 versions = 1;
  }
  oneof policy_choice {
    Latest latest = 100;
    All all = 101;
    Specific specific = 102;
  }
}

Import

// In ModelConfig:
import "tensorflow_serving/config/file_system_storage_path_source.proto";

I/O Contract

Inputs

Name	Type	Required	Description
policy_choice	oneof	Yes	One of: latest, all, or specific
latest.num_versions	uint32	No	Number of latest versions to serve (default 1)
specific.versions	repeated int64	No	Explicit list of version numbers to serve

Outputs

Name	Type	Description
Aspired versions	vector<ServableData>	Filtered list of versions to load

Usage Examples

Model Config with Version Policies

# models.config - serve latest 2 versions
model_config_list {
  config {
    name: 'my_model'
    base_path: '/models/my_model/'
    model_platform: 'tensorflow'
    model_version_policy {
      latest {
        num_versions: 2
      }
    }
  }
}

# Serve all versions
model_config_list {
  config {
    name: 'my_model'
    base_path: '/models/my_model/'
    model_platform: 'tensorflow'
    model_version_policy {
      all {}
    }
  }
}

# Serve specific versions only
model_config_list {
  config {
    name: 'my_model'
    base_path: '/models/my_model/'
    model_platform: 'tensorflow'
    model_version_policy {
      specific {
        versions: 42
        versions: 43
      }
    }
  }
}

Related Pages

Implements Principle

• Principle:Tensorflow_Serving_Version_Policy_Configuration