Bug report

Bug description:

The binary plist writer builds its struct formats for the object references of every array and dict, and for the offset table, as one format code per item ('>' + code * n). The struct module caches compiled formats, and a format with one code per item compiles to about 33 bytes per item, so after plistlib.dumps() returns, memory proportional to the largest containers written stays allocated in the cache (up to 100 such formats).

import plistlib, struct, tracemalloc

tracemalloc.start()
for i in range(40):
    plistlib.dumps([True] * (250_000 - i), fmt=plistlib.FMT_BINARY)
print(tracemalloc.get_traced_memory()[0])  # ~330 MB
struct._clearcache()
print(tracemalloc.get_traced_memory()[0])  # < 1 kB

The reader already uses repeat-count formats (f'>{n}{code}', gh-86269), which compile to a constant size; the writer should do the same. The output bytes are unchanged.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-161106: Use repeat counts in plistlib binary writer struct formats #161107
• [3.14] gh-161106: Use repeat counts in plistlib binary writer struct formats (GH-161107) #161137
• [3.15] gh-161106: Use repeat counts in plistlib binary writer struct formats (GH-161107) #161138