Implementation:Tensorflow Serving PrometheusExporter GeneratePage

Knowledge Sources	TensorFlow Serving
Domains	Monitoring, Operations
Last Updated	2026-02-13 17:00 GMT

Overview

Concrete tool for exporting TensorFlow Serving metrics in Prometheus text format via HTTP endpoint, provided by the prometheus_exporter module.

Description

PrometheusExporter::GeneratePage() collects all registered metrics from CollectionRegistry::Default() and serializes them in Prometheus text format. It uses three serialization helpers:

• SerializeHistogram() (L65-115): Converts histogram/sampler metrics to Prometheus bucket format
• SerializeScalar() (L117-130): Converts counter/gauge metrics
• SerializeMetric() (L132-152): Dispatches to histogram or scalar serializer based on metric type

The exporter is exposed at kPrometheusPath ("/monitoring/prometheus/metrics") via the HTTP server when monitoring is enabled.

Usage

Accessible at http://<host>:<rest_api_port>/monitoring/prometheus/metrics. Configure Prometheus to scrape this endpoint.

Code Reference

Source Location

• Repository: tensorflow/serving
• File: tensorflow_serving/util/prometheus_exporter.cc L156-193
• Header: tensorflow_serving/util/prometheus_exporter.h L29-44

Signature

class PrometheusExporter {
 public:
  static const char* const kPrometheusPath;  // "/monitoring/prometheus/metrics"

  PrometheusExporter();

  Status GeneratePage(string* http_page);

 private:
  monitoring::CollectionRegistry* collection_registry_;
};

Import

#include "tensorflow_serving/util/prometheus_exporter.h"

I/O Contract

Inputs

Name	Type	Required	Description
http_page	string*	Yes	Output buffer for the metrics page

Outputs

Name	Type	Description
http_page	string*	Prometheus text-format metrics page

Usage Examples

Scrape Metrics

# Get all metrics
curl http://localhost:8501/monitoring/prometheus/metrics

# Example output:
# HELP :tensorflow:serving:batching_session:queuing_latency Queuing latency
# TYPE :tensorflow:serving:batching_session:queuing_latency histogram
# :tensorflow:serving:batching_session:queuing_latency_bucket{le="100"} 42
# :tensorflow:serving:batching_session:queuing_latency_count 100
# :tensorflow:serving:batching_session:queuing_latency_sum 5432.1

Prometheus Config

# prometheus.yml
scrape_configs:
  - job_name: 'tf_serving'
    scrape_interval: 15s
    static_configs:
      - targets: ['localhost:8501']
    metrics_path: '/monitoring/prometheus/metrics'

MonitoringConfig Proto

# monitoring_config.txt
prometheus_config {
  enable: true
  path: "/monitoring/prometheus/metrics"
}

Related Pages

Implements Principle

• Principle:Tensorflow_Serving_Performance_Monitoring