Simulating Clients: IID Partitions

Intuition

In federated learning (FL), data lives on many devices (phones, hospitals). To study FL algorithms on one machine we split a dataset into "client" shards. The easiest split is IID: every client gets a random sample, so all clients look statistically alike.

Precise explanation

Shuffle indices with a seeded generator: rng.permutation(n).
Split into n_clients nearly equal parts with np.array_split (sizes differ by at most 1).
Shards must be disjoint and cover all indices.
Real federated data is rarely IID (FL-04) and client sizes are often very unequal — which is why FedAvg weights clients by size (FL-02). McMahan et al. (2016) evaluate both IID and pathological non-IID partitions.

flowchart LR D[(dataset)] --> P[shuffle + split] --> C1[client 1] & C2[client 2] & C3[client 3]

Worked example

n = 10, 3 clients → sizes [4, 3, 3].

Common mistakes

Overlapping shards (sampling with replacement).
Forgetting the seed → irreproducible experiments.
Assuming equal client sizes later in the pipeline.

Why this matters

Every FL experiment starts with a partition; its realism decides whether conclusions transfer to real deployments.

References

Communication-Efficient Learning of Deep Networks from Decentralized Data — H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, Blaise Agüera y Arcas (S028)
Advances and Open Problems in Federated Learning — Peter Kairouz, H. Brendan McMahan et al. (S032)