I agree that an heuristic is needed to decide when a dict should be compacted.

> * When (dict size < dk_size/8), call insertion_resize()

In bpo-31179, I suggested to Yury to use 2/3 ratio... to avoid integer overflow :-) He first used 80%, but I dislike using the FPU in the dictobject.c. I'm not sure of the cost of switching from integers to floats, and more generally I hate rounding issues, so I prefer to use regular integers ;-)

+           (3) if 'mp' is non-compact ('del' operation does not resize dicts),
+               do fast-copy only if it has at most 1/3 non-used keys.
+
+           The last condition (3) is important to guard against a pathalogical
+           case when a large dict is almost emptied with multiple del/pop
+           operations and copied after that.  In cases like this, we defer to
+           PyDict_Merge, which produces a compacted copy.

By the way, if dict automatically compacts itself automatically, do we still need Yury's test "is the dict compact"?