pulseon 0.1.0a5

Local-first training metrics tracker backed by Rust, PyO3, DuckDB, and DuckLake.

PulseOn

PulseOn is a local-first training metrics tracker backed by Rust, PyO3, DuckDB, and DuckLake.

Upcoming alpha: 0.1.0a5 / v5 headless read surface:

• discover projects, runs, metrics, and persisted metric points
• query running runs after reports have reached persistent storage
• return Python objects or Arrow PyCapsule-compatible tables
• inspect existing stores through a dependency-free, read-only CLI
• use DuckDB or SQLite catalogs with local or S3-compatible Parquet data
• keep the Parquet schema as the long-term compatibility boundary

Quickstart:

import pulseon

client = pulseon.init()
project = client.create_project("local training")
run = client.create_run(project.project_id, "baseline")
run.log("train/loss", 0, 0.25)
client.finish_run(run.run_id)

projects = client.list_projects()
runs = client.list_runs(project.project_id, status="finished", limit=20)
metrics = client.list_metrics(run.run_id)
points = client.query_metric(
    run.run_id,
    "train/loss",
    start_step=0,
    end_step=100,  # Exclusive: the query range is [0, 100).
)
table = client.query_metric_table(run.run_id, "train/loss")
client.shutdown()

ArrowTable does not require PyArrow, pandas, or Polars. Consumers that support the Arrow PyCapsule protocol can import it through __arrow_c_stream__.

The pulseon command opens an existing store and never creates a missing one:

pulseon --path runs projects list
pulseon --path runs runs list <project-id> --status finished --limit 20
pulseon --path runs metrics list <run-id>
pulseon --path runs --format json metrics query <run-id> train/loss --all

By default, PulseOn stores local state under ./.pulseon. Pass an explicit root path when a project should use a different local store:

client = pulseon.init("runs")

The existing storage keywords remain available: data_path, catalog_backend, catalog_path, and metric_queue_capacity. catalog_path must be a local filesystem path. data_path may be local, or it may use an S3-compatible URI such as s3://bucket/prefix.

Project-local storage settings can live in ./.pulseon/config.toml. Relative data_path and catalog_path values in this file are resolved from the project root passed to pulseon.init(...) or pulseon --path:

data_path = "s3://example-bucket/pulseon/demo"

[s3]
endpoint = "https://s3.example.com"
region = "us-east-1"
access_key_id = "<access-key-id>"
secret_access_key = "<secret-access-key>"
path_style = true
use_ssl = true

Do not commit real S3 credentials. Explicit pulseon.init(...) keyword arguments override values from config.toml:

client = pulseon.init(
    data_path="s3://example-bucket/pulseon/demo",
    s3_endpoint="https://s3.example.com",
    s3_access_key_id="<access-key-id>",
    s3_secret_access_key="<secret-access-key>",
    s3_path_style=True,
)

For bounded teardown, stop active logging threads before calling client.shutdown(timeout=...); PulseOn keeps admission open while bounded shutdown is draining, so concurrent run.log(...) calls can prevent that drain from completing before the timeout.

Architecture entry points:

• Docs index
• Native storage boundary
• Glossary
• Roadmap
• ADRs

Runtime extensions:

• DuckLake is installed and loaded by the native engine because it is required for native storage.
• DuckDB LTTB is optional and is not bundled into PulseOn wheels. Downsampling first uses an already installed lttb extension. Set PULSEON_LTTB_AUTO_INSTALL=1 to let the first downsampled query run the official INSTALL lttb FROM community; LOAD lttb; flow and cache the extension in DuckDB's user extension directory.
• PulseOn embeds DuckDB 1.5.4. The release gate verifies the LTTB 0.1.0 community build for DuckDB 1.5.4. DuckDB extension binaries are specific to a DuckDB version and platform; for offline deployment, set PULSEON_LTTB_EXTENSION_PATH=/path/to/lttb.duckdb_extension to a compatible, signed binary rather than reusing one built for another DuckDB version.

See DuckDB's LTTB extension page and extension installation guide for the upstream commands and compatibility rules.