NULL value for metrics

Is it practical to support NULL values for metrics? Currently NULL values are saved as 0 by default.

This means we need to do messy workarounds, such as creating separate datasets for null and non-null values, or defining a magic number that does not exist in the actual data.

Using a magic number creates complexity for others trying to consume the dataset.
For example with NULL you can use SUM() or AVG() on a field and the NULL values will be ignored, whereas using a magic number would return an incorrect result unless the user knew the magic number needed to be filtered first.

The issue names no files, tests, or entry points. Start by tracing how metric NULL values are stored and how SUM() and AVG() consume them; done means defining and implementing NULL semantics without requiring magic values or separate datasets, with coverage for the affected query behavior.

Written by the indexing model from the issue text.