Version 2.5.0

memtier_benchmark 2.5.0

First release of the 2.5 minor series (cut from master @ ec3d55f). Same-source performance verified equal-or-better vs 2.4.4 on Intel Ice Lake (pipeline 1: +0.04%, pipeline 16: +0.29%, p99 unchanged).

Features

• Built-in Prometheus /metrics exporter via --prometheus-port (opt-in), with per-run labels and configurable latency buckets (#468); arbitrary-command hit/miss counter series (#484)
• --read-preference read-endpoint routing for cluster and standalone (#456)
• Per-thread and whole-process CPU tracking with saturation warning (#471)
• --transaction EXEC-abort breakdown and live miss/abort rates (#486); single-key reuse across a --command rotation in cluster mode (#485)
• Cluster-mode endpoint count and connection fan-out reporting (#489, #491)

Performance

• fill_pipeline: make bufferevent_enable conditional, recovering the ~4% ops/sec regression introduced in 2.4 (#448)
• Alloc-free per-position miss tracking; drop reply materialization for empty-collection arbitrary commands (#472, #469)

Fixes

• Enable arbitrary-command miss tracking on cluster shard connections (#477)
• Normalize latency stats through int64 microseconds (#463)
• Monitor-input hardening: skip HELLO while loading, bound fuzz payload (#498)

Packaging / CI

• RPM release objects uploaded with public-read ACL; one-time ACL maintenance workflow hard-scoped to memtier objects only (#501, #506, #509)