Pull request overview

Adds new documentation/examples demonstrating datacenter pinning for endpoints, and updates the network volume example to explicitly set datacenter on both the NetworkVolume and Endpoint now that the SDK default is no longer EU-RO-1.

Changes:

• Add a new 04_scaling_performance/02_datacenters/ example (GPU single-DC + multi-DC, and CPU with CPU-supported DC restriction).
• Update the 05_data_workflows/01_network_volumes/ GPU/CPU workers to explicitly set datacenter for the shared volume and the endpoints.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

1. Blocker: companion PR #266 must merge first

All new DataCenter values used in this PR (US_GA_1, EU_NL_1, etc.) don't exist in the current SDK — DataCenter on main has one member: EU_RO_1. The examples fail on import against flash v1.9.1:

DataCenter.US_GA_1  # AttributeError — doesn't exist yet

PR #266 also adds the datacenter= alias on NetworkVolume and the list-accepting signature for Endpoint(datacenter=[...]). Without it:

• NetworkVolume(datacenter=...) — silently dropped (no such field on the model; BaseResource uses Pydantic v2's default extra='ignore')
• Endpoint(datacenter=[DataCenter.US_GA_1, DataCenter.EU_RO_1]) — validation error; current type is DataCenter, not Union[DataCenter, list[DataCenter]]

Hold this PR until #266 merges.

2. cpu_worker.py uses LB-style API; gpu_worker.py uses QB-style

Both workers demonstrate datacenter selection, but they use different API patterns: cpu_worker.py creates an api = Endpoint(...) instance and registers routes with @api.post()/@api.get(), while gpu_worker.py uses the @Endpoint(...) function decorator. A user reading the example to learn about datacenters ends up learning two different deployment patterns simultaneously, which blurs what the example is actually teaching.

If the LB-style is intentional for CPU workers in this section, the README should call it out. If not, it would be cleaner to use QB-style for both:

@Endpoint(name="04_02_cpu_eu", cpu="cpu3c-2-4", workers=(0, 2), datacenter=DataCenter.EU_RO_1)
async def process(data: dict) -> dict:
    return {"datacenter": "EU-RO-1", "result": data}

3. Multi-DC response is missing the datacenter indicator

The single-DC endpoint returns {"datacenter": "US-GA-1", "result": payload}, making it easy to verify which DC handled the request. The multi-DC endpoint only returns {"result": payload} — you can't tell which DC ran the job. Worth including for the example to be demonstrable:

# multi-DC endpoint
return {"datacenters": ["US-GA-1", "EU-RO-1"], "result": payload}

Or note in the README how to verify DC placement another way.

4. Network volume update

Good change — explicitly setting datacenter=DataCenter.EU_RO_1 on both NetworkVolume and Endpoint is the right fix now that the default changes to None. The inline comment # same volume as gpu_worker.py -- must match name and datacenter is helpful.

Nits

• README.md Quick Start says pip install -r requirements.txt but no requirements.txt is included in the diff. If there's a shared one at a parent level, the path should be explicit.
• The hardcoded string "datacenter": "EU-RO-1" in the cpu_worker.py response will drift if the DC is changed. Could use DataCenter.EU_RO_1.value instead.

Verdict

Hold until PR #266 merges. The examples are well-structured and the datacenter coverage (single DC, multi-DC, CPU restriction) is the right scope for this section. The LB vs QB inconsistency is worth addressing before merge, but isn't a blocker on its own.