Agent Dashboard — Disk Write Latency
What this chart shows
Disk Write Latency (ms) is the weighted-average write service time (w_await) across every block device on the agent host, summed across drives. The figure is identical to the w_await column in iostat -x.
The agent polls /proc/diskstats every 5 seconds, computes the delta, and reports the aggregated value on each heartbeat.
This is a host-wide measure of how long each individual write I/O takes. It is not a probe file write time, and it is not specific to the agent's working directory. Any process on the host that issues writes contributes to the sample.
Write latency is collected the same way as Disk Read Latency. The only difference is that it uses the write counters (writes_completed and write_ms) instead of the read counters.
Platform support
| Platform | What is reported |
|---|---|
| Linux | Aggregated w_await across all disks (ms). This is a real value. |
| Other / unsupported | -1, displayed as "N/A". |
Idle-disk behavior
When no read or write IOPS were observed in the last 5-second sample window, the Live Agents disk-I/O pill shows idle in gray instead of "0 ms". The chart point for that sample is still 0.
Read the chart together with the IOPS pill (R/s, W/s) on the agent card. This tells you whether the disk is fast or no I/O happened.
On SSD and NVMe storage, the kernel can report 0 ms because the real latency is below the integer millisecond resolution of /proc/diskstats. In that case the pill shows < 1 ms instead of zero.
Why it matters
During every copy job, the copy engine writes per-page checkpoints, ClickHouse spool entries, and step logs to disk. High write latency slows these writes and can cause copy jobs to stall between pages.
If the agent runs on network-attached or slow storage, this chart shows the I/O pressure on the write path specifically. Pair it with Disk Read Latency to see whether the backing device is slow for reads, writes, or both.
How to read it
| Value range | Interpretation |
|---|---|
| 0 / "< 1 ms" | Sub-millisecond. Typical of local SSD or NVMe. |
| 1–5 ms | Healthy local disk or well-provisioned volume. |
| 5–20 ms | Normal for network-backed or shared storage. |
| 20–100 ms | Elevated. Storage contention or a slow backing store. |
| > 100 ms | Problematic. Checkpoint writes lag. Consider moving to faster storage. |
Threshold coloring
The Live Agents disk-I/O pill uses these colors.
| Pill color | w_await range |
|---|---|
| Green | < 5 ms |
| Amber | 5 – 19.99 ms |
| Red | ≥ 20 ms |
| Gray | N/A (unsupported or idle) |
Polling interval
The agent samples every 5 seconds and reports on the heartbeat (about 30 seconds plus 0–10 seconds of jitter).
A spike that lasts only one sample is usually a transient I/O burst. Sustained elevation indicates a persistent storage performance problem.
Tips
- Correlate write-latency spikes with copy job throughput drops. If objects per second falls when latency spikes, the underlying device is the bottleneck.
- A large gap between read and write latency can point to a write-back cache, a full disk, or fsync pressure on the checkpoint path.
- Check the storage type where the agent writes its
data/directory. Local SSD or NVMe typically delivers< 1 ms. NFS-backed or network storage can spike to 50–200 ms under load. - On non-Linux hosts the value is intentionally "N/A".