werkator/docs/plan/09-system-metrics.md

Step 09: System Metrics

Prerequisites: step 07 (API), step 08 (layout). Read README.md and 00-legacy-analysis.md first.

Goal

Port the legacy system page: CPU, RAM, disk, and repository size with min/max/avg aggregation.

Design

Create package de.hoennig.werkator.metrics:

• SystemMetricsCollector sampling every 60s (server profile only): CPU used/idle from /proc/stat deltas, RAM from /proc/meminfo, disk from java.nio.file.FileStore, repo size via periodic du -sk (or a file walk) — throttle repo-size sampling (legacy ran du every cycle, which was expensive).
• Keep running min/max/avg per metric since server start; persist aggregation state in the artifact root so restarts continue the series (legacy system_state.dat, but as JSON).
• GET /api/system returning the current snapshot plus aggregates (legacy system.json fields are the reference).
• /system HTML view in the step 08 layout, polling /api/system every 60s with the same error-badge rules.

Out of Scope

• Alerting, historical time series, external monitoring integration.
• Windows/macOS support beyond graceful degradation (missing /proc → metric shows "n/a").

Tests

• Collector unit tests with fake /proc file content (read paths injectable).
• Aggregation math: min/max/avg over samples, persistence round-trip.
• Controller slice test for /api/system.
• Graceful degradation when a source is unreadable.

Acceptance Criteria

• ./gradlew ktlintFormat then ./gradlew build is green.
• /system renders live values on Linux (manual smoke test; document in this file).

Implementation Notes (2026-07-07)

Implemented as designed: SystemMetricsCollector in de.hoennig.werkator.metrics samples every 60s once ServerMetricsLifecycle (server profile only) calls start(), following the watcher's start/stop pattern. CPU comes from /proc/stat deltas, RAM from /proc/meminfo, disk from java.nio.file.FileStore (df semantics: used = total − unallocated, free = usable), and the repository size from a file walk. GET /api/system returns the snapshot plus aggregates, and /system renders the legacy system page in the step 08 layout, polling every 60s with the same timeout/error-badge rules. Since the metric rows are fixed, werkator.js only updates the cell texts in place — nothing is rebuilt.

Deviations and decisions:

• The JSON uses camelCase fields with nested {current, min, max, avg} aggregates instead of the flat snake_case legacy system.json; the value set matches legacy.
• The aggregation state persists as system-metrics-state.json in the artifact root and restarts continue the series, as this step requires. Legacy actually deleted system_state.dat on every start, so the footnote now reads "since the first server start" instead of "since script start". ArtifactStore gained rootDir() so the state can live next to the stored builds.
• CPU load needs a counter delta, so the first sample after process start reports no CPU metric yet (n/a); legacy aggregated a meaningless near-zero first delta instead.
• The repository size is re-probed only every 10th sample (10 minutes) and reused in between — the throttle this step requires; legacy ran du -sk every cycle. The file walk sums file sizes, not disk blocks like du, which is close enough for a trend metric.
• An unavailable source (no /proc outside Linux, unreadable file store) yields explicit null metrics over HTTP 200 and n/a cells; the failure is logged once, not every 60s.
• No new config keys: the 60s interval is fixed like legacy, so WerkatorConfig, the init templates, and docs/configuration.md are unchanged.
• 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.

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. Killing the server flipped the indicator to the red error badge and dimmed the table — zero spinners; after a restart the tab returned to live and the series continued from the persisted state (sampleCount 5, min/max from before the restart preserved). The 375px viewport stacked the rows as labeled cards, and SIGINT shut the server down cleanly (exit 130, no exceptions).