Trailing rolling average per key

Constraints

1. 1The average covers the current row and the window - 1 usable values before it, within the same key. It never reaches into another key.
2. 2Order within a key by order_field, ascending, compared as text.
3. 3Return the rows ordered by key then by the ordering field.
4. 4A value that is not a number does not enter the window. The row still appears, carrying the average of the values before it.
5. 5Before any usable value has been seen for a key, the average is None rather than zero.
6. 6A window larger than the data averages what exists. Do not pad.
7. 7A window of zero or less, or one that is not an integer, has no meaning: return an empty list.
8. 8Round each average to four decimal places, and return copies of the rows with rolling_avg added.

Worked example

The api service reports 100, 140 and 120 in its first three minutes, so the third row's three-minute average is 120.0. The db rows sit between them in the file; a single window shared across services would have averaged 8 and 12 into the api figures and produced a latency chart with no relationship to anything. At 09:04 the api collector reports no latency. The window does not advance — there is nothing to add — so the row carries the average of 100, 140 and 120 that preceded it. Treating the missing value as zero would drop the average to 90 and put a dip in the chart at the exact moment the collector was broken, which is the opposite of what the reader needs to see.

What this tests

Per-group windowed state, and the decision about what a missing value does to a window. Both mistakes — one window for everything, and missing-means-zero — produce a chart that looks smooth and says something false.