mi/werkator
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Implement quota-aware disk metrics (PR#16)

Browse Source 
The system page showed the disk of the host volume, not the budget the
instance can actually fill — on a Hostsharing Managed Webspace that is a
group quota, tighter than the volume by an order of magnitude, so the
warn/critical highlighting could never fire before a build failed with
"Disk quota exceeded".

DiskQuota parses `quota -u -g --no-wrap --raw-grace` and picks the
tightest of the user quota, the group quota, and the volume itself;
SystemMetricsCollector reports whichever binds, resets the disk
min/max/avg when the binding source changes, and the system page names
the source in its info line. A host without a binding quota (Docker
hosts, developer machines) renders exactly as before.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
mhoennig
2026-09-03 17:23:39 +02:00
co-authored by Claude Sonnet 5
parent be965617cd
commit 61c235313f
12 changed files with 574 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/plan/09-system-metrics.md
+2
View File
@@ -55,6 +55,8 @@ Deviations and decisions:
- The legacy `generation` field was not ported; it only guarded the legacy JS against monitor restarts.
- The CPU count comes from `Runtime.availableProcessors()` instead of `nproc`.
**Implementation note (PR#16, 2026-09-03):** the disk metric is now the tightest of the user quota, the group quota, and the volume, not the volume alone — `DiskQuota` parses `quota -u -g --no-wrap --raw-grace` and `SystemMetricsCollector.readDisk()` picks whichever candidate has the smallest headroom. On hosts without a binding quota (Docker hosts, developer machines) nothing changes; on a Hostsharing Managed Webspace the group quota is usually the real limit, so `diskTotalGib` there is the quota's soft limit instead of the host volume's size, and the info line names the source. A source change (volume → quota, or one quota subject to another) resets `diskUsedGib`/`diskFreeGib`'s min/max/avg so a stale ceiling from the previous source never survives.
Manual smoke test (2026-07-07): scratch repository with a bare origin, server on port 18986, observed through a real browser tab (via a TCP proxy, so the tab outlived backend restarts).
The first sample rendered RAM/disk/repo values immediately with CPU `n/a` and the totals line (`8 cores`, RAM/disk GiB, updated time).
After the next 60s poll the open tab updated in place without reload: the updated time ticked, CPU used appeared (1.58 cores, idle 6.42 = 8 total), and min/max diverged.