TiKV
v8.5.5Storage & Data2026년 1월 15일
이번 릴리스는 TiKV의 정상 종료와 관련 설정 동작을 다듬고, 성능과 자원 동작을 개선했습니다. TiKV의 일부 정확성 문제도 수정했으며, 보안 권고나 명시된 취약점은 포함되지 않습니다.
그 외 기록된 변경 16건 전체fixes 6 · value changes 6 · additions 4
fixes (6)
- - Fix the issue that the KV Cursor Operations metric for analyze requests is always 0 #19206 @ glorv
- - Fix the issue that Region heartbeats might report incorrect Region size or key statistics to PD after a leader change #19180 @ glorv
- - Fix the issue that unsafe recovery gets stuck by removing tombstone TiFlash learners from the unsafe recovery demotion list #18458 @ v01dstar
- - Fix the issue that snapshots might be repeatedly canceled during continuous writes, which blocks replica recovery #18872 @ exit-code-1
- - Fix the issue that compaction slows down due to increased flow-control thresholds #18708 @ hhwyt
- - Fix the issue that Raft peers might enter hibernation prematurely in a corner case, causing them to remain busy and block leader transfers after a TiKV restart #19203 @ LykxSassinator
value changes (6)
- TiKV rocksdb. ( defaultcf | writecf | lockcf | raftcf ) .level0-slowdown-writes-trigger Modified Starting from v8.5.5, when the flow control mechanism is enabled ( storage.flow-control.enable is set to true ), this configuration item is overridden by storage.flow-control.l0-files-threshold only if its value is greater than the storage.flow-control.l0-files-threshold . This behavior prevents weakening RocksDB's compaction acceleration mechanism when you increase the flow control threshold. In v8.5.4 and earlier versions, when the flow control mechanism is enabled, this configuration item is directly overridden by storage.flow-control.l0-files-threshold .
- TiKV rocksdb. ( defaultcf | writecf | lockcf | raftcf ) .soft-pending-compaction-bytes-limit Modified Starting from v8.5.5, when the flow control mechanism is enabled ( storage.flow-control.enable is set to true ), this configuration item is overridden by storage.flow-control.soft-pending-compaction-bytes-limit only if its value is greater than storage.flow-control.soft-pending-compaction-bytes-limit . This behavior prevents weakening RocksDB’s compaction acceleration mechanism when you increase the flow control threshold. In v8.5.4 and earlier versions, when the flow control mechanism is enabled, this configuration item is directly overridden by storage.flow-control.soft-pending-compaction-bytes-limit .
- - Introduce CPU-aware scaling for the unified read pool to avoid CPU starvation under hot read workloads #18464 @ mittalrishabh
- - Add network latency awareness to slow score to avoid scheduling leaders to TiKV nodes with unstable network conditions #18797 @ okJiang
- - Optimize hibernate Region behavior by allowing leaders to enter the hibernation state immediately after receiving a majority of votes, without waiting for offline non-voter peers #19070 @ jiadebin
- - Throttle BR log restore requests when TiKV memory usage is high to prevent TiKV OOM #18124 @ 3pointer
additions (4)
- - Support gracefully shutting down TiKV #17221 @ hujiatao0 When shutting down a TiKV server, TiKV attempts to transfer the Leader replicas on the node to other TiKV nodes within a configurable timeout duration before the shutdown. The default timeout duration is 20 seconds, and you can adjust it using the server.graceful-shutdown-timeout configuration item. If the timeout is reached and some Leaders have not been successfully transferred, TiKV skips the remaining Leader transfers and proceeds with the shutdown. For more information, see documentation .
- TiKV readpool.cpu-threshold Newly added Specifies the CPU utilization threshold for the unified read pool. The default value is 0.0 , which means that there is no limit on the CPU usage of the unified read pool. The size of the thread pool is determined solely by the busy thread scaling algorithm, which adjusts the size dynamically based on the number of threads handling current tasks.
- TiKV server.graceful-shutdown-timeout Newly added Controls the timeout duration for graceful shutdown of TiKV. The default value is 20s .
- TiKV server.inspect-network-interval Newly added Controls the interval at which the TiKV HealthChecker actively performs network detection to PD and other TiKV nodes. The default value is 100ms .
TiKV 스택에 추가
조치가 필요한 변경이 나오면 주간 메일로 받아볼 수 있습니다. 이번 릴리스처럼 조치가 없는 주에는 메일을 보내지 않습니다.