Expman Core

Core storage and logging engine for expman-rs.

Overview

The expman module provides the foundational components for tracking, storing, and managing experiment data with high performance. Its central design principle ensures that logging operations (e.g., log_vector()) are extremely fast (~100ns channel sends) and never block the main experiment process. All I/O operations are handled asynchronously via a background tokio task.

Key Features

• Asynchronous Logging: Uses a tokio background task and mpsc channels to offload I/O operations, ensuring the main application thread is never blocked.
• Efficient Storage: Writes metrics using Apache Arrow and Parquet formats in batches, rather than inefficient per-step read-append-write cycles.
• Strong Typing: Provides robust data models for experiment configurations, run states, and metric values.

Usage Example

Below is a basic example of how to use expman directly in a Rust application to track an experiment:

use expman::{ExperimentConfig, LoggingEngine, RunStatus};

fn main() -> anyhow::Result<()> {
    // 1. Define the experiment configuration
    // This sets up an experiment named "my_rust_exp" in the "./experiments" directory
    let config = ExperimentConfig::new("my_rust_exp", "./experiments");

    // 2. Initialize the logging engine
    // This spawns the background tokio task for handling I/O
    let engine = LoggingEngine::new(config)?;

    // 3. Log some parameters (hyperparameters, configuration, etc.)
    engine.log_params([
        ("learning_rate".to_string(), 0.001.into()),
        ("batch_size".to_string(), 32.into()),
        ("optimizer".to_string(), "Adam".into())
    ].into());

    // 4. Log metrics during your training or simulation loop
    for step in 0..100 {
        let loss = 1.0 / (step as f64 + 1.0);
        let accuracy = step as f64 / 100.0;

        // log_vector takes a HashMap of metrics and an optional step counter.
        // This operation takes ~100ns and does not block!
        engine.log_vector([
            ("loss".to_string(), loss.into()),
            ("accuracy".to_string(), accuracy.into())
        ].into(), Some(step));
    }

    // 5. Gracefully close the engine, ensuring all pending metrics are flushed to disk
    engine.close(RunStatus::Finished);

    Ok(())
}

Architecture Map

• ExperimentConfig: Configures the base directory, experiment name, flush intervals, and environment metadata.
• LoggingEngine: The main entry point. It manages the background writer task and exposes non-blocking methods to send data.
• models: Contains definitions for RunStatus, MetricValue (which supports Floats, Ints, Bools, and Strings), and MetricRow.
• storage: Internal module that handles the actual file I/O, writing config.yaml, run.yaml, run.log, and vectors.parquet.