The Arnold projectFour ways to copy a table, and the three that do not
A census table of household sizes, one tidy-up, and a check on what each copy still remembers.
The data is Census 2021 table TS017, household size, for the 128 output areas of Gedling that the ONS places inside the Arnold built-up area. Each output area becomes a row with eight numbers: households of one person, two people, and so on up to eight or more. That makes 1,024 cells. The learner builds the table in Python as a list of lists, which is the first structure most people reach for.
The first attempt creates the empty grid with [[0] * 8] * 128 and fills it cell by cell from the census file. It runs without complaint. Then the learner compares the grid with the file and finds that 789 of the 1,024 cells are wrong, and only one row, the last, is right. Asking Python for the identity of each row with id() explains it: there is one row object, and the outer list holds 128 labels pointing at it. Every row written simply overwrote the same list. Building the grid with a comprehension, [[0] * 8 for _ in range(128)], gives 128 separate rows and no wrong cells.
The second experiment is about backups. Before tidying the table, the learner makes a copy, then folds the "8 or more" column into the "7" column to make "7 or more", drops the last column and appends a blank row for notes. Seventeen of the output areas had at least one household of eight or more people, so seventeen cells genuinely change. What does each backup look like afterwards?
What the backup shows after tidying the original, our Python run| How the backup was made | Changed cells | Rows now one column short | Rows in the backup |
|---|
| backup = table | 17 | 128 | 129 |
| backup = table.copy() | 17 | 128 | 128 |
| backup = [row[:] for row in table] | 0 | 0 | 128 |
| backup = copy.deepcopy(table) | 0 | 0 | 128 |
Plain assignment copies nothing; the backup is the table under a second name, so it shows every change including the extra row. table.copy() is a shallow copy: it makes a new outer list, so the appended row does not appear, but the inner rows are shared, so every edit inside them leaks through. Copying each row, or calling copy.deepcopy, gives a true backup. The two safe methods differ in speed on this laptop: deepcopy took 0.38 milliseconds and the row-by-row slice 0.011 milliseconds, because deepcopy has to inspect every object it meets. For a table of plain numbers, slicing each row is enough; for rows that themselves contain lists or dictionaries, it is not.
Ages 8 to 11
Sticky labels on boxes: what happens if two labels go on the same box and you change what is inside?
Ages 11 to 15
Build the grid both ways in Python and count the wrong cells yourself.
Ages 15 and up
Test all four backups, explain each result with id() and is, and time the two safe copies.
Sources and limits
Household sizes are Census 2021 TS017 from Nomis, published by the ONS under the Open Government Licence. Which output areas count as Arnold comes from the ONS 2021 output area to built-up area lookup, restricted to Gedling. The grid, the tidy-up and every count and timing above come from our own Python run; timings depend on the machine.
