v2-tutorial

A tutorial to introduce users to Flyte v2 features through examples

Quick start

Prereqs: Python 3.10+.

Option A: uv (recommended)

If you don't have uv:

curl -fsSL https://astral.sh/uv/install.sh | sh

Then install dependencies declared in pyproject.toml:

uv sync

Option B: pip

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e .

Running examples:

• Use flyte run ... from your venv, or prefix with uv run if using uv.
• Ensure your Flyte CLI is configured for your project/domain. You can pass -p <project> -d <domain> to the command if needed.

Authentication

Create a flyte config file with:

flyte create config \
    --endpoint <orgName>.hosted.unionai.cloud \
    --builder remote \
    --domain development \
    --org <orgName> \
    --project <projectName> \
    --output ~/.union/config.yaml

Update this command to use the orgName and projectName for your org and project.

Exercises

1) Hello Polyglot

Command:

flyte run 1_hello_world/hello_polyglot.py main --letter e

What it does:

• Fans out dynamically across languages from polyglot-hello whose primary/alias code starts with the given letter.
• Runs hello_for_code in parallel via flyte.map, returning a mapping of language code → greeting.
• Validates letter (requires a single alphabetic character).

Tip: change --letter (e.g., p, s, f) to see different sets.

2) Failure handling and resource overrides

Command:

flyte run 2_failure_handling/oomer.py failure_recovery

What it does:

• Intentionally triggers an OOM in oomer (allocates a very large list) with small default memory.
• Catches flyte.errors.OOMError and retries the same task with higher memory via .override(resources=...).
• Always runs always_succeeds() in a finally block and returns its result to demonstrate cleanup and recovery.

3) Dependency-aware agents (toy sandwich planner)

Command:

flyte run 3_agents/smolagent.py main --goal "Make a peanut butter and jelly sandwich"

What it does:

• Builds a simple plan (steps with deps) in get_plan, then executes it in execute_plan respecting dependencies.
• Runs ready steps concurrently, grouping iterations with flyte.group, and aggregates results by step id.
• Uses a reusable task environment (single warm replica) and a preinstalled image dependency for faster iteration.

Try changing --goal to explore different prints/flow; the toy plan remains the same to showcase orchestration.

4) Reports bundle (interactive HTML reports)

Command:

flyte run 4_reports/run_all.py main

What it does:

• Orchestrates all report generators concurrently: globe visualization (Three.js), interactive dashboard (Chart.js/D3), 3D scatter (Plotly), satellite images gallery, protein 3D viewer, and a YouTube embed page.
• Each task emits a rich HTML report via flyte.report; run_all simply gathers them so you can open each artifact from the Union UI.

What's next?

Go check out the examples in the flyte-org repo: https://github.com/flyteorg/flyte-sdk/tree/main/examples Check out our latest blogs and resources to see what we're up to: https://www.union.ai/resources Reach out on slack: https://slack.flyte.org/