Storage Tiering Service · View 06 of 31 · 2 · People and journeys
The worst moment
- Contain. The on-call engineer does not yet know how far the policy got. Rings, per-window caps and a single freeze flag mean the answer is bounded before anyone looks: at most the current ring, at most 2% of objects.
- Reverse. Objects moved more than 24 hours ago have already released their source. Those need a real promotion with a real retrieval cost, and that cost is the honest measure of the incident.
What the architecture owes this journey
- Every movement carries the decision id, and every decision carries the policy version, so 'what did v43 move' is one query (ADR-19).
- Reversal is itself a planned policy with guards, rate limits and a budget. An emergency is the worst time to bypass the rules that stop a tool relocating the corpus.
- The incident is closed with a dollar figure: retrieval, early-delete and movement charges attributed to v43.
Numbers
- Freeze to halted rings under 30 seconds. Pointer-flip reversal of the last 24 hours of movements at 5,000 commits per second per shard group.