Duplicates after a merge are almost always a many-to-many key. Diagnose
before you merge — pandas even has a switch to make it fail loudly.

STEP 1 — CONFIRM WHICH SIDE
  left['customer_id'].duplicated().sum()
  right['customer_id'].duplicated().sum()
The side with a nonzero count is the culprit. Often it's the right
(e.g. one customer, many orders), which turns a 1-to-1 into a fan-out.

STEP 2 — LET PANDAS CATCH IT FOR YOU
  df.merge(right, on='customer_id', validate='one_to_one')
validate raises if the relationship isn't what you claim. Use
'one_to_many' / 'many_to_one' to assert the real shape — this turns a
silent duplicate bug into a loud error at merge time.

STEP 3 — FIX BY INTENT
  You want one row per customer, right has many → aggregate right first:
    right.groupby('customer_id').agg(total=('amount','sum')).reset_index()
    then merge. One-to-one, no fan-out.
  You genuinely want the fan-out (one row per order) → the "duplicates"
  are correct; the bug is downstream expecting uniqueness.

THE TRAP
Blindly .drop_duplicates() after the merge hides the real problem and
can drop legitimately distinct rows. Fix the join shape, don't paper
over it.

Share both dataframes' shapes and what one output row should represent,
and I'll write the exact merge.