The bug we're hitting is in a different subsystem: the Usage Sanity Check (UsageSanityChecker.java). Specifically the "snapshot after removed" check, whose query is roughly:

SELECT count(*)
FROM   cloud_usage.cloud_usage cu
JOIN   cloud.snapshots s ON cu.usage_id = s.id
WHERE  cu.usage_type = 9
  AND  cu.start_date > s.removed;

i.e. usage records in cloud_usage.cloud_usage with a start_date AFTER the snapshot was already soft-deleted in cloud.snapshots. This drives the Usage Sanity Check failed mailer and feeds our Prometheus exporter.

On the running 4.22.1.0 (which includes 83ce006) we still see:

Originally posted by @PPisz in #13398 (comment)