As seen in #1328 there is a substantial difference in performance in using a graph stored using C++ vectors (~2 seconds) vs when benchmarking end to end from python + pandas (~9 seconds). Benchmark script

A lot of data scientists use python and arrow ( a standardized columnar format) to read graphs into memory.

Polars (implemented in rust) uses arrow natively. Pandas optionally since 2.0:

df = pd.DataFrame({"a": [1, 2, 3]}, dtype_backend="pyarrow")

Optimizing for this use case could unlock even more performance for networkit users.