Getting data in

pd.read_csv takes a path, a URL, or anything file-like, and hands back a DataFrame. It also guesses a great deal — the delimiter, the header row, and every column's type — which is why the next paragraph matters more than this one.

Look at it before you trust it

Four calls, every time, before writing any analysis:

df.head()      # do the values look like what you expected?
df.shape       # did you get all the rows?
df.dtypes      # is anything text that should be a number or a date?
df.info()      # dtypes, non-null counts and memory, all at once

Then df.describe() for ranges — a -999 hiding in a column of ages shows up here — and df.isna().sum() for how much is missing, per column.

Five seconds of looking prevents the failure mode where you build an entire analysis on a date column that's actually strings.

Tell it what you meant

Fixing types on the way in is cheaper than fixing them afterwards:

pd.read_csv(
    "orders.csv",
    parse_dates=["date"],           # real timestamps, not strings
    dtype={"cups": "Int64"},        # nullable integer, not float
    index_col="date",               # use it as the index straight away
)

Files that fight back

Real CSVs are rarely tidy, and there's a keyword for each way they misbehave:

sep=";"                    # a different delimiter
skiprows=3                 # junk above the header
header=None, names=[...]   # no header row at all
na_values=["N/A", "-"]     # what counts as missing in this file
decimal=","                # European decimal commas
thousands=","              # 1,234 as a number
encoding="latin-1"         # when UnicodeDecodeError appears

na_values is the one people miss. If a file writes missing data as "N/A" or "-", pandas reads it as text — and your numeric column silently becomes strings.

Big files

usecols=["date", "cups"]   # only the columns you need
nrows=1000                 # sample while developing
chunksize=100_000          # stream it in pieces

usecols is the biggest easy win — most analyses touch a handful of columns out of dozens.

Other sources, and writing back

pd.read_excel / read_json / read_parquet / read_sql / read_html

df.to_csv("out.csv", index=False)      # index=False, nearly always
df.to_parquet("out.parquet")           # smaller, faster, keeps dtypes

Two habits worth forming. Pass index=False to to_csv unless the index means something, or you'll grow a stray Unnamed: 0 column every round trip. And for anything you'll read more than once, use Parquet — it's columnar, compressed, and preserves dtypes exactly, so none of the guessing above has to happen again.