Syvain Metrics API Client

Python client for the Syvain Metrics v2 API. Use it to find and organize folders and experiments and to read one experiment's record, catalog, annotations, and series from code.

Pick the right tool:

• Analysis across experiments: the syvain-metrics-duckdb extension attaches Metrics as a database with an on-disk cache. Use it for plots, tables, joins, and anything spanning a folder.
• Recording training or evaluation jobs: syvain-metrics-collector queues, batches, and delivers metrics in the background. This client's write session exists for API completeness, tests, and backfills.
• Programmatic control and one-experiment reads: this client.

Install

uv add syvain-metrics-api-client

Connect

from syvain_metrics_api_client import Metrics

metrics = Metrics()

The client resolves credentials the same way as the syvain-metrics CLI. The API key comes from api_key=, then SYVAIN_METRICS_API_KEY, then the login saved by syvain-metrics auth login. The host comes from host=, then SYVAIN_METRICS_HOST, then that login, then https://metrics.syvain.com. Metrics is also a context manager that closes its HTTP session.

Folders and experiments

Address a folder by absolute path or id; / is the root. Address an experiment by id or slug.

mamba = metrics.make_folder("/models/mamba")  # creates missing parents
mamba = metrics.folder("/models/mamba")       # MetricsNotFound when missing
mamba.folders()                               # direct children
mamba.experiments()                           # direct members
mamba.experiments("run-*", recursive=True)    # slug glob, whole subtree
metrics.experiments("syv-2039-*")             # every experiment matching a glob

exp = metrics.experiment("mamba-run-001")
exp.move("/archive")
mamba.rename("mamba-v2")
mamba.move("/")

Lookups use a snapshot of the folder tree, read on first use. Changes made through the client refresh it. A folder() or experiment() lookup that misses reads the tree again once before raising. Listings such as folders() and experiments() use the snapshot as is, so call metrics.refresh() to list what was created elsewhere. Folder and Experiment are immutable values. A rename or move returns a new handle.

Read one experiment

info = exp.info()          # status, description, meta, error, lifecycle datetimes
catalog = exp.catalog()    # series names with their metadata keys and values
notes = exp.annotations()  # every annotation, newest first

for series in exp.series("loss", where={"split": "valid"}):
    print(series.metadata, series.steps, series.values)

series() returns one Series per metadata partition. Its steps, timestamps_ms, and values are aligned. Pass one name, several names, or none to read every series in the catalog. where matches exact metadata values, has=("rank",) requires a key, and x="timestamp" orders points by time instead of step. The client returns a series only after the stream completes, so partial data is never returned.

Write session

with metrics.open_experiment(
    "backfill-001", description="Imported run", meta={"seed": 7}, folder="/imports"
) as run:
    run.start()
    run.metric("loss", 0.25, step=1, metadata={"split": "train"})
    run.done()

Each call is one acknowledged request. metrics() takes many MetricPoints and sends up to 1000 per request. The session refreshes its experiment token before it expires and revokes the token when the block exits. Reopening an existing slug replaces its description and meta. An experiment already in another folder stays there. A token left idle past its 24 hour expiry cannot be refreshed, so reopen the experiment instead. For real training jobs use syvain-metrics-collector.

Other events

metrics.send(event, data) sends any v2 organization event to /api/v2/events and returns the reply data. The API reference is at https://metrics.syvain.com/api/openapi.json.

Errors and retries

Every failure is a MetricsError with code, message, status, and retryable. MetricsNotFound (also a LookupError) reports an unresolved reference. MetricsAuthError reports missing credentials.

Reads and metric writes retry transient failures with backoff. A retried metric keeps its original message id, which the API deduplicates. Every other write never retries, because repeating it can have side effects: folder creation, moves, renames, open_experiment, lifecycle events (a replayed done would move done_at), annotations, and send().