Filed from the IronCache pre-implementation coverage audit (2026-06-13): no existing issue adequately owned this.
Why this is needed
Pin the operational latency commitment the engine must hold under load, separate from the marketing-benchmark methodology: a per-command-class p99/p999 budget the engine must not exceed while a background activity runs (online defrag #43, mass eviction at the ceiling, forkless snapshot traversal #60, same-deadline TTL cascade #51, large-value free #51), plus a slow-operation guard for O(N) commands consistent with the SLOWLOG 10ms threshold, with acceptance tests asserting the budget holds during each named background activity. #7 fixes the headline metrics and bans closed-loop p99 from marketing; #86 ships SLOWLOG/LATENCY as a reporting surface, not an enforced budget. Several designs promise 'does not block the hot path' qualitatively but no issue pins or tests the numeric tail-latency target. This turns those qualitative promises into a falsifiable contract.
Context
Relates to / partially overlaps #7. Part of the vision EPIC #1.
Why this is needed
Pin the operational latency commitment the engine must hold under load, separate from the marketing-benchmark methodology: a per-command-class p99/p999 budget the engine must not exceed while a background activity runs (online defrag #43, mass eviction at the ceiling, forkless snapshot traversal #60, same-deadline TTL cascade #51, large-value free #51), plus a slow-operation guard for O(N) commands consistent with the SLOWLOG 10ms threshold, with acceptance tests asserting the budget holds during each named background activity. #7 fixes the headline metrics and bans closed-loop p99 from marketing; #86 ships SLOWLOG/LATENCY as a reporting surface, not an enforced budget. Several designs promise 'does not block the hot path' qualitatively but no issue pins or tests the numeric tail-latency target. This turns those qualitative promises into a falsifiable contract.
Context
Relates to / partially overlaps #7. Part of the vision EPIC #1.