TopolearnGetting data in with the right types, and saving it in a fast typed format, is the first step of every real job.
Polars has eager readers that return a DataFrame: pl.read_csv, pl.read_parquet, pl.read_json, pl.read_ndjson, pl.read_excel and pl.read_database. Writers mirror them: df.write_csv, df.write_parquet, df.write_json.
Useful read_csv options:
separator=";",has_header=Falsecolumns=["a", "b"]to read only some columnsschema_overrides={"zip": pl.String}to stop codes with leading zeros becoming integerstry_parse_dates=Trueto turn ISO date strings intoDate/Datetimenull_values=["", "NA", "-"]for extra missing markers
In Polars 2, read_csv is literally scan_csv(...).collect(), so it runs on the same multi-threaded engine as lazy queries, and it accepts a list of files or a glob such as "data/*.csv".
Parquet is the format to prefer: it is columnar, compressed and stores the schema, so reads are fast, exact and need no type guessing. pl.read_parquet(path, columns=[...]) reads only the columns you ask for. When you will reread data, convert CSV to Parquet once.
For files bigger than memory, or when you only need part of a file, use the lazy scan_* functions instead; they come up in the lazy lessons.
Example
import polars as pl
with open("orders.csv", "w") as f:
f.write("id;day;zip;qty\n1;2026-03-01;02134;2\n2;2026-03-02;10001;NA\n")
df = pl.read_csv(
"orders.csv",
separator=";",
schema_overrides={"zip": pl.String}, # keep the leading zero
null_values=["NA"],
try_parse_dates=True,
)
print(df.schema)
# Schema({'id': Int64, 'day': Date, 'zip': String, 'qty': Int64})
df.write_parquet("orders.parquet")
back = pl.read_parquet("orders.parquet", columns=["zip", "qty"])
print(back.rows()) # [('02134', 2), ('10001', None)]Remember
schema_overrides,null_valuesandtry_parse_datesfix most CSV typing issues- Parquet keeps the schema exactly and reads only requested columns
- Eager
read_*loads everything; lazyscan_*is for big or partial reads
Watch out for
- Letting ID or postal-code columns be inferred as integers and losing leading zeros.