---

name: runtime-data description: "Guides SecretFlow local runtime setup, device objects, and federated data containers such as FedNdarray and federated dataframes." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Runtime and data

Use this sub-skill when the task is about getting SecretFlow running locally, creating devices, moving objects between devices, or shaping data into the federated containers used by later model or component workflows.

Owns

• sf.init, sf.shutdown, and the basic local simulation flow
• PYU, SPU, HEU, TEEU
• reveal, to, wait, and device-object transfer rules
• FedNdarray, HDataFrame, VDataFrame, MixDataFrame, partition
• CSV/ORC IO and small data-loading helpers
• common direct-device and partition-alignment failures

Does not own

• component CLI / evaluation / export workflows — use component-cli
• preprocessing/statistics/model training — use analytics
• PSI, Kuscia, TEEU deployment, or production orchestration — use privacy-orchestration

Trigger phrases

Use this route when a user asks things like:

• how to start SecretFlow locally
• how to reveal or transfer objects between devices
• how to build a vertical or horizontal dataframe
• how to read CSV or ORC into a federated structure
• why to(SPU) or to(HEU) failed
• how to align partitions or select columns across parties

Reading order

1. Read references/runtime-data.md for the API map and flow choices.
2. Read references/troubleshooting.md in the root skill when install/runtime symptoms are not clearly local to this workflow.
3. Use scripts/local_quickstart.py as the tiny smoke helper when you need to confirm that a local runtime works from the target environment.

Workflow

1. Confirm the party/device layout first. Decide whether the workflow is a simple local simulation or an explicit multi-party cluster config.
2. Put data on a plain device first, then move it onward. Direct placement on SPU or HEU is intentionally rejected by the runtime.
3. For federated dataframes, verify which party owns each partition before you start selecting, joining, or assigning columns.
4. If a task only needs a tiny end-to-end proof, run the bundled quickstart smoke helper and inspect its output before moving on to a larger workflow.
5. When the task later needs component evaluation or model training, hand off to the sibling sub-skill rather than expanding this one.

Common decisions

• Use address='local' for the fastest local simulation proof.
• Use cluster_config only when the task truly depends on explicit party addresses and ports.
• Prefer PYU as the staging device for plain Python objects and small data.
• Treat reveal as a last-mile inspection tool, not as the default data path.

Bundled files

• references/runtime-data.md — API table, quick-start flow, and runtime caveats.
• scripts/local_quickstart.py — minimal local hello-world smoke helper.