NATS Blog (Sep 17) β Reliability-focused minor release that reworks stream/consumer scale, move, peer-add, peer-remove. Meta, streams, and consumers keep independent peer sets; previously shared peer-membership duties could disagree under degradation. In 2.15 the meta layer tracks desired peer membership; asset leaders drive safe single-step Raft membership changes with catchup/replication awareness so availability is preserved while converging (ADR-62). Also: stream backup/restore v2 (per-message format, no exclusive lock, CLI filter/obfuscate; ADR-63); stream-source recreate detection via persisted create-time + highest-sourced-sequence index; large sync: always throughput gains on replicated streams (Raft-layer batching + skip per-message stream sync β hundreds msg/s β tens of thousands, p99 cut >40%); default max 1000 consumers per stream. Upgrade: prefer β₯v2.14.7 β 2.15; defer scale/move during mixed-version windows.
π₯ What Changed
Netflix Technology Blog (Sep 26) β Stratum (serverless media on Titus; hundreds of thousands of containers/day) ships the opt-in Stratum Resource Tuner (SRT). Core construct: a recommendation state machine (CREATED β EVALUATING β APPLIED / REJECTED β MONITORED). Daily Conductor workflow reads a month of Iceberg high-res usage; compressible resources (CPU, network) target p99; incompressible (memory, disk) target p99.9 plus buffer. Safety stack: (1) per-container retry-with-more-resources watcher aborts at 85% non-cached memory / 95% disk and requeues at 1.5Γ; (2) synthetic 100%-replay then production 90/5/5 with fail-fast on >1% retry or 10β15% median slowdown; (3) 90-day soak with dual windows and auto-revert restoring last known-good. Hard lesson: encodeV2 50% memory cut passed experiments, then a traffic shift drove retries >60% β today SRT clamps memory/disk to 15% step reductions, enforces cooldown, and ships a global kill switch. Results: encodeV1 β50% CPU / β40% memory; re-tuned encodeV2 stable β20% memory.