Implementation:Mlflow Mlflow Log Metric

Knowledge Sources	MLflow MLflow API
Domains	ML_Ops, Experiment_Tracking
Last Updated	2026-02-13 20:00 GMT

Overview

Concrete tool for recording numeric performance metrics within an MLflow run, provided by the MLflow library.

Description

mlflow.log_metric and mlflow.log_metrics are fluent API functions for recording numeric performance measurements under the currently active MLflow run. log_metric records a single metric value at an optional step and timestamp, while log_metrics records a dictionary of metric key-value pairs in a single batch call. Both functions support optional association with a model ID and dataset for fine-grained tracking. If no run is active, a new run is automatically created.

Usage

Use log_metric inside a training loop to record per-step or per-epoch metrics such as loss, accuracy, or learning rate. Use log_metrics to log multiple evaluation metrics at once after a validation pass. Provide the step parameter to build training curves that can be visualized in the MLflow UI. Set synchronous=False in tight training loops where blocking on I/O is undesirable.

Code Reference

Source Location

Repository: mlflow
File: mlflow/tracking/fluent.py
Lines (log_metric): L1007-1100
Lines (log_metrics): L1163-1252

Signature

def log_metric(
    key: str,
    value: float,
    step: int | None = None,
    synchronous: bool | None = None,
    timestamp: int | None = None,
    run_id: str | None = None,
    model_id: str | None = None,
    dataset: Dataset | DatasetEntity | None = None,
) -> RunOperations | None: ...

def log_metrics(
    metrics: dict[str, float],
    step: int | None = None,
    synchronous: bool | None = None,
    run_id: str | None = None,
    timestamp: int | None = None,
    model_id: str | None = None,
    dataset: Dataset | DatasetEntity | None = None,
) -> RunOperations | None: ...

Import

import mlflow

I/O Contract

Inputs

Name	Type	Required	Description
key	str	Yes (log_metric)	Metric name. May contain alphanumerics, underscores, dashes, periods, spaces, and slashes. Maximum length 250 characters.
value	float	Yes (log_metric)	Metric value. Special values like +/- Infinity may be replaced by store-specific bounds (e.g., max/min float for SQL stores).
metrics	dict[str, float]	Yes (log_metrics)	Dictionary mapping metric names to float values for batch logging.
step	int or None	No	Training step or epoch number associated with the metric. Defaults to 0 if unspecified.
synchronous	bool or None	No	If True, blocks until logged. If False, returns a future. If None, reads from `MLFLOW_ENABLE_ASYNC_LOGGING`.
timestamp	int or None	No	Unix timestamp in milliseconds when the metric was computed. Defaults to current system time.
run_id	str or None	No	If specified, logs the metric to the given run instead of the currently active run.
model_id	str or None	No	ID of the model associated with this metric. Falls back to the active model ID if not specified.
dataset	Dataset or DatasetEntity or None	No	The dataset associated with the metric, enabling dataset-level performance tracking.

Outputs

Name	Type	Description
(log_metric return)	RunOperations or None	When `synchronous=True`, returns None. When `synchronous=False`, returns a `RunOperations` future representing the pending logging operation.
(log_metrics return)	RunOperations or None	When `synchronous=True`, returns None. When `synchronous=False`, returns a `RunOperations` future for the batch operation.

Usage Examples

Basic Usage

import mlflow

with mlflow.start_run():
    # Log a single metric
    mlflow.log_metric("mse", 2500.00)

    # Log a metric with a step for training curves
    for epoch in range(10):
        loss = 1.0 / (epoch + 1)
        mlflow.log_metric("train_loss", loss, step=epoch)

Batch Metric Logging

import mlflow

with mlflow.start_run():
    # Log multiple evaluation metrics at once
    eval_metrics = {
        "accuracy": 0.95,
        "precision": 0.93,
        "recall": 0.91,
        "f1_score": 0.92,
    }
    mlflow.log_metrics(eval_metrics)

Asynchronous Logging

import mlflow

with mlflow.start_run():
    # Log metrics asynchronously in a tight training loop
    for step in range(1000):
        loss = compute_loss(step)
        mlflow.log_metric("loss", loss, step=step, synchronous=False)

Related Pages

Implements Principle

Principle:Mlflow_Mlflow_Metric_Logging

Requires Environment

Environment:Mlflow_Mlflow_Python_Runtime_Environment

Uses Heuristic

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment