clients_daily_scalar_aggregates_v1 contains a lot of redundant values

[edugfilho]

From mozilla/bigquery-etl#919

mozilla/bigquery-etl#919 (comment)

clients_daily_scalar_aggregates contains metric name and types for every key, process, agg_type combination in each row:

This can be improved by using a nested struct for each metric so that the name, type, and key are only stored once per row:

This would break the existing schema but that should be easy to deal with by modifying the existing table in place and updating the downstream query.

A further improvement is to not store metrics with only null values but I'm not sure if there's some downstream need for retaining those values.

From testing on a 1% sample, the above schema change would reduce the table size to ~29%, not storing nulls would reduce to ~31%, and doing both would reduce to ~12.3%. This reduce storage costs significantly. If we change to a 7 day retention period for clients_daily_scalar, this would save around $150-$180 a week on storage and some more on downstream etl data scanned (maybe another couple hundred). This isn't huge but worth considering.