# Host and Disk Behavior Known Cases

Source note: adapted from the Feishu `Storage oncall handbook` <https://pingcap.feishu.cn/wiki/RNzIwCexFiA1uXkRCvzcsIocnVb> and `Raftstore handbook` <https://pingcap.feishu.cn/wiki/HaZTwLVC0iDxngk2hX3cbR33nqn>.

## Disk Issue Drives High IO Util Without Clear Business Impact

- Phenomenon:
  - TiKV disk IO util reaches very high values, sometimes near 100 percent, but user-facing impact remains limited or inconsistent.
- Trigger / Preconditions:
  - The dominant issue is on the disk device or host side rather than a TiKV logical bottleneck.
- Root cause:
  - Some device-level anomalies can push IO-util metrics high without matching end-to-end workload degradation.
- Workaround:
  - Treat this as a host or disk investigation, not as proof of TiKV saturation by itself.
- Related oncall:
  - v4.0.6 TiKV stress test showed uneven IO util distribution `TICKET-4367`

## RaftEngine Periodic Compaction Causes Transient Traffic Spikes

- Phenomenon:
  - Disk read and write traffic rises periodically and may briefly approach the device limit.
- Trigger / Preconditions:
  - RaftEngine periodic compaction is active.
- Key metrics:
  - `tikv-details -> raft engine -> flow -> rewrite-append`
- Root cause:
  - Periodic RaftEngine compaction rewrites data and can create short bursts of disk traffic.
- Workaround:
  - No universal workaround is documented in the source material; interpret the spike in the context of compaction scheduling and storage headroom.
- Related oncall:
  - Airbnb reported intermittent EBS read/write spikes on `v6.5.X` `GTOC-7369`

## irqbalance Should Usually Stay Enabled

- Phenomenon:
  - Some environments question whether disabling `irqbalance` will stabilize performance.
- Root cause:
  - `irqbalance` normally helps distribute interrupt load across CPU cores.
  - Disabling it can create core-level imbalance unless there is a very specific platform issue.
- Workaround:
  - Keep `irqbalance` enabled unless the environment has a proven kernel or NIC-specific problem.
  - If the platform is known to have `irqbalance`-related NIC instability, follow the kernel-upgrade path documented by the environment owner.
- Related oncall:
  - a Kylin vendor environment asked about disabling `irqbalance` because of NIC-related jitter `ONCALL-7086`