Function analysis_rows 

pub fn analysis_rows(
    lf: &LazyFrame,
    sample_rows: Option<usize>,
    known_total: Option<usize>,
    seed: u64,
    polars_streaming: bool,
) -> Result<AnalysisRows>

Read the rows an analysis works on: all of them when the table has no more than sample_rows (or sample_rows is None), and otherwise a seeded sample of that many, spread across the whole table rather than taken from its head.

Two ways to spread it, chosen by what the plan can do cheaply:

• A plan whose slices reach into a single Parquet or IPC scan reads [SAMPLE_BLOCKS] short runs at seeded places across the table. Each run is a row group or two, so a sample of a 400-million-row hive table reads a few dozen row groups, not the table. known_total saves the count; the footers give it cheaply otherwise.
• Anything else — a filter, a query, a union of files, a CSV — is read once as a stream, keeping the rows whose seeded rank is lowest. That is a uniform sample in bounded memory, and the same pass counts the rows, so a filtered view is read once rather than counted and then read.