Currently, virtual row groups (see https://github.com/hyparam/hyperparam-cli/blob/master/src/lib/tableProvider.ts) are needed to avoid decoding unneeded rows in case of big parquet files with only one row group.

We favor low memory usage, while we do avoidable network requests, because the same row group will be fetched again and again, but only the relevant rows will be decoded.

Ideally, or at least optionally, we could cache the undecoded row groups, to avoid fetching them again, but decode the cells on demand to save memory.
It would also reduce the risk of being rate-limited or blocked by the parquet publishing platform (hi Hugging Face!)