A file reader needs a path, a format and a schema strategy. CSV stores text fields and usually needs a header option plus explicit types or carefully reviewed inference. JSON supports nested fields. Parquet stores a schema with its columnar data, so it does not use a CSV header option.
The training workspace provides three representations of the same original order extract under /data. Those paths are browser fixtures, not your computer's file system. You can practice how a reader produces a DataFrame and how transformations follow the read. Production ingestion also needs credentials, governed storage locations and decisions about malformed records. Inferring a schema from an early sample can miss a later type change. For a stable pipeline, define and review the contract rather than assuming each file will contain the same clean values.