For the record, the issue with Decimal context rounding is that double negation does not always round-trip:

>>> from decimal import Decimal
>>> from math import pi
>>> Decimal(pi)
Decimal('3.141592653589793115997963468544185161590576171875')
>>> - - Decimal(pi)
Decimal('3.141592653589793115997963469')
>>> Decimal(pi) == - - Decimal(pi)
False

So a value that goes into the negated lo heap may not exactly match that value that comes out. This only happens when the inputs have more precision than the current context precision. That arises when converting binary-floats to decimal-floats.

Possibly rename the iterable parameter to data, consistent with statistics.median? Or keep as-is to match more_itertools conventions and to emphasize the lazy evaluation which is the principal use case for running_median()?

+0 for as-is on this one.

Happy with the O(n) but mostly fast steps to maintain a sorted window? Or add more complex O(log n) code (IndexableSkiplist, blist, SortedContainers, etc)? Based on Grant Jenks' notes, the current list insort/bisect/del technique can be expected to win for window sizes up to several thousand.

+1 on the current implementation - it's reasonably clear what's going on, and the other options are probably overkill for this general purpose library.

Okay, I think we're good to go.

This is great, thanks - can't wait to find a place to use it.