Send a representative sample of logs to Grafana Loki, including high-volume services and a known incident. Choose labels deliberately: enough to narrow searches, but not unique per request or user. Measure ingestion delay, query speed, storage cost, and whether an on-call engineer can find the event they need.
Set retention and access rules before broad ingestion, and remove secrets at the source. Test a backup or export path and a failed ingestion scenario. A log system should make diagnosis faster without silently dropping evidence or creating excessive label cardinality.