Node detail page overview

Problem

Two separate complaints, one page.

The page is a form, not a view. /admin/alerts/node/<id>/change/ renders the registry fields, an editable config, and three readonly extras stacked in single-column label/value rows: a 120x24 disk sparkline, a list of ten pipeline runs, and a one-line preflight summary. An operator who opens a node because the changelist flagged it red learns almost nothing about what is actually wrong with that machine. The severity chips that carry the signal live on the changelist and stop at its edge.

Latest preflight is broken for the hub itself. manage.py preflight persists instance_id=settings.INSTANCE_ID (apps/checkers/management/commands/preflight.py:73). That setting is "" on a hub that never set the env var, because only agents are told to set it (.env.sample:42). The hub’s own Node row, however, is keyed by apps.alerts.identity.local_instance_id(), which falls back to socket.gethostname(). NodeAdmin.latest_preflight filters PreflightRun.objects.filter(instance_id=obj.instance_id), matches nothing, and prints “No preflight recorded.” even immediately after a successful run. The run is not lost: the PreflightRun changelist shows it, because that view does not filter by node. Two writers, two spellings of the same machine.

The constraint that shapes the charts

CheckRun rows are written in exactly one place, apps/checkers/checkers/base.py:126, by the machine that ran the checker. There is no hub-push for them. So a hub holds CheckRun history for itself and for nobody else.

A peer reaches the hub as Alert rows instead: deduped on (fingerprint, source), updated in place, carrying metric_* labels and an annotations dict of metric values. One row per checker per node, overwritten on every push. Current state, never a series.

The consequence is blunt: today’s disk sparkline is permanently empty on every peer node page, and no amount of admin work changes that. Storing a metric history for pushed results is a real feature with its own model, migration and retention policy. It is deliberately out of scope here and left as a later project. This design instead makes the gap legible rather than pretending it does not exist.

Design

Page shape

A new templates/admin/alerts/node/change_form.html extends admin/change_form.html and overrides {% block content %} to render an overview above {{ block.super }}. NodeAdmin.render_change_form supplies the context, built by a new apps/alerts/node_overview.py, so panel logic is unit-testable without driving the admin.

disk_sparkline, recent_pipelines and latest_preflight move out of readonly_fields and fields into the overview. The form below is left as the identity fields plus the one operator-editable field, config.

Precedent: templates/admin/dashboard.html with config/dashboard.py.

Panel 1 — identity header

Role badge: This hub when node.instance_id == local_instance_id(), otherwise Peer. Then hostname, address, instance_id, last_source, and last_seen rendered as an age with a green/amber/red read.

The freshness thresholds are imported from config/dashboard.py rather than restated, so the dashboard nodes card and this page can never drift apart.

The unresolved-severity chips move out of NodeAdmin.incidents into a shared helper that both the changelist column and this header call, so one node reads the same on both pages.

Panel 2 — per-checker current state

One row per checker this node reports: checker name, status, latest metric value, age, and a link to the underlying record. Two sources, one shape:

• local node — the newest CheckRun per checker_name
• peer — the node’s Alert rows that carry a labels["checker"], newest per checker, value read from annotations

Both normalise to the same row dataclass, so the template runs one loop and has no idea which source it drew from. This is the only panel that works for every node, and it is the answer to “what is actually wrong with this box”.

Panel 3 — recent incidents

The ten newest incidents reached through alerts__node, each with severity, status, age and an admin link. The chips in the header give counts; this gives names.

Panel 4 — charts

Three time series from CheckRun, larger than today’s sparkline:

Chart	Checker	Metric
Disk usage	disk	worst_percent
CPU	cpu	cpu_percent
Memory	memory	memory_percent

render_sparkline (apps/checkers/admin_charts.py) grows optional y-axis min/max labels and a title, with defaults that keep its existing callers byte-identical.

Charts render only for the local node. A peer gets one honest sentence in their place: metric history is written by the machine that runs the checker and is not pushed to a hub, so there is nothing to plot here yet. No blank chart, no empty axes.

The preflight fix

Three parts:

1. preflight.py writes local_instance_id() instead of raw settings.INSTANCE_ID. One import, one line. Now both writers spell the machine the same way.
2. A data migration in apps/checkers stamps existing instance_id="" rows with this machine’s id, so history written before today becomes visible instead of staying orphaned.
3. The preflight panel renders only for the local node. Peers get the same node-local explanation as the charts: preflight is node-local and is not pushed to a hub, so a hub never holds a peer’s preflight run.

Read-time matching on a blank id was rejected: it leaves the bad data in place and obliges every future reader to know the trick.

Testing

• apps/alerts/_tests/test_node_overview.py — the four panel builders, each over local vs peer and empty vs populated, with non-numeric metric values skipped rather than raising.
• An admin test asserting the custom template renders and that the preflight panel resolves for the hub’s own node — the regression that started this.
• apps/checkers/_tests — the writer fix, and the migration backfill over a row with a blank instance_id.

100% branch coverage on changed code, per AGENTS.md.

Out of scope

• A metric-sample store for pushed results. Named here so the next person knows the empty-state text is a deliberate placeholder, not an oversight.
• Preflight retention pruning, already flagged as a follow-up in preflight.py.