GuidesUpgrade your client

Upgrade your client

Move your scripts and notebooks to the current welt-client release. Upgrading the SDK does not refit a model or change an existing remote predictor.

Install the current release in the environment used by your script or notebook:

python -m pip install --upgrade "welt-client[parquet]"

Connect to the intended workspace#

• Client() defaults to https://welt.ergodic.dev. An explicit base_url overrides WELT_BASE_URL; select http://localhost:8000 deliberately for a local server.
• Public catalogue reads need no key. Authenticated work needs a workspace credential: run welt login in a terminal or call Client().connect() in a notebook.
• Constructors never start authentication. Explicit keys and WELT_API_KEY keep precedence. A Python connection lasts only for the process unless save=True is explicit.

See Connect for private persistence and remote or headless use.

The current release supports multiple saved workspace profiles and explicit workspace selection. A key stays bound to its original tenant; a workspace option cannot retarget it. See CLI workspace selection.

Keep your existing supervised workflow#

fit(X, y) remains valid. You can also name the target in a DataFrame you pass directly. Give exactly one target source:

# Continue with the real train table and target name from Start.
model.fit(train, target=target)

• Select Classifier or Regressor explicitly.
• Missing or duplicate target names, conflicting y and target, and invalid schemas are rejected locally before upload.
• Query rows still contain features only.
• Predictors retain their exact model and configuration identity. Changing models requires an intentional new fit.

Use the released resource methods#

The current client includes:

• research metadata sidecars and one-step table loading;
• resumable file upload sessions;
• durable batch prediction;
• workspace inspection;
• hidden-model runs;
• cross-validated evaluations with DataFrame results and report links (0.12).

Both sync and async clients expose the corresponding resources. Estimator CSV and Parquet convenience reads still load into memory. For resumable file transfer, use Client.upload_file with a retained upload ID.

Job.result() returns a batch manifest for successful or partial batches. Retrieve its payload separately and inspect failed ranges. See capabilities, CLI recipes and hidden-model runs.

Preserve durable identities when recovering#

Job.result(timeout=...) budgets initial inspection, polling, GET retry waits and final result retrieval. It never cancels or resubmits work. Keep job_id and reconnect using the same endpoint and workspace credentials. See Wait and recover.

Usage now includes nullable credits and rate_card_version. HTTP 402 remains a WeltError whose code explains the balance, spend-cap or unpaid-plan condition. Check Workspace Billing before rerunning work; see usage accounting.

Next steps#