jcoffey-dev
e10289ee36
Twelve hosts running six services read as somebody's side project. Fifty hosts across twenty-one services read as an estate, which is what a visitor is trying to see themselves in. Thirty-one Linux, eighteen Windows, one Linux host whose agent is gone. The proportions are the point: Windows now carries Active Directory, IIS, SQL Server, Exchange, file shares, Remote Desktop, print, WSUS and SCCM rather than appearing as a Security channel on one box. Linux gains a load-balancer tier, an outbound proxy, MySQL beside Postgres, RabbitMQ, Elasticsearch, three Kubernetes nodes, CI, Vault, OpenLDAP, BIND and a backup server. Fifteen new generators, each writing what the real daemon writes -- HAProxy's timing quintuple, MySQL slow-query blocks, W3C extended format for IIS, kubelet PLEG lines, BIND query logging with NXDOMAIN, Squid's TCP_DENIED, SQL Server deadlock and I/O-stall messages -- with the structured fields carried in attributes so both halves of the query language have something to work on. Three things found while doing it, each of which would have shipped as a quiet wrongness: linuxHosts() decided Windows by `service == "eventlog"`. That held while eventlog was the only Windows role; with IIS and SQL Server on Windows it would have given every one of them a journald system stream -- sshd and UFW lines on a Windows box. It now decides from `os`. worker-02 lost its filling disk when the fleet was rewritten, which is the story worker-disk-filling thresholds on. Restored at its original rate, with a comment saying why it cannot move. Five dashboard panels were written as `timechart`, which this query language does not have -- its stages are where/stats/sort/fields/head/ tail. They are raw ClickHouse SQL now, which the reference recommends for exactly this, and a `pie` panel was dropped before it shipped because the web's VizType union has no such member even though the API accepts one. Five new dashboards: platform/Kubernetes, directory and DNS, messaging and search, the Windows server estate, and edge/proxy. Every panel was checked against the language's real stage list and the web's real viz types. Volume roughly quadruples: about 316 records/minute at rate-scale 1, and 1.9M per nightly reset at the demo's own settings against about 0.5M before. ClickHouse will not notice; reset time and disk on the demo box might, so the README says so and names RATE_SCALE as the lever.
114 lines
5.8 KiB
Markdown
114 lines
5.8 KiB
Markdown
# demo-seed
|
||
|
||
Everything the public demo deployment (`demo.cairnobs.org`) is built out
|
||
of, kept in the repo rather than only on the box so a demo can be rebuilt
|
||
from scratch and so its dashboards and rules are reviewable like any
|
||
other code.
|
||
|
||
- `reset-demo.sh` — the nightly reset. Wipes every volume, brings the
|
||
stack back up, re-seeds users, notification targets, a week of
|
||
synthetic history, dashboards, and alert rules. Run from cron at 04:00
|
||
on the demo host.
|
||
- `dashboards/*.json` — one file per dashboard, in the shape
|
||
`POST /dashboards/import` consumes (identical to what
|
||
`GET /dashboards/{id}/export` and the web UI's Export JSON button
|
||
produce, so a dashboard edited in the UI can be exported straight back
|
||
into this directory).
|
||
- `alerts/*.json.template` — one file per alert rule, in the shape
|
||
`POST /rules` consumes. `__TARGET_OPS__` / `__TARGET_SECURITY__` /
|
||
`__TARGET_PLATFORM__` are substituted at apply time with the IDs of the
|
||
three notification targets `reset-demo.sh` creates: every reset starts
|
||
from an empty database, so the IDs can't be baked in.
|
||
- `cairnobs-demo-simulator.service` — systemd unit for the live half of
|
||
the demo, `/hack/demo-simulator`. Installed at `/etc/systemd/system/`
|
||
on the demo box.
|
||
|
||
## The fleet
|
||
|
||
Fifty hosts, shaped like an estate rather than a stack: thirty-one Linux,
|
||
eighteen Windows, and one Linux host whose agent is gone so the Agents
|
||
page has something stale to show.
|
||
|
||
| Tier | Hosts |
|
||
|---|---|
|
||
| Edge and proxy | `lb-01/02` (HAProxy), `edge-01/02` (nginx), `proxy-01` (Squid) |
|
||
| Application | `api-01`–`04`, `worker-01`–`03`, `arm-build-01` (aarch64) |
|
||
| Data | `db-01/02` (Postgres), `mysql-01`, `cache-01/02` (Redis), `mq-01/02` (RabbitMQ), `search-01/02` (Elasticsearch) |
|
||
| Platform | `k8s-node-01`–`03` (kubelet), `ci-01` (Jenkins), `vault-01`, `ldap-01` (OpenLDAP), `dns-01` (BIND), `backup-01`, `mail-01` |
|
||
| Windows | `DC-01/02`, `IIS-01`–`03`, `WIN-SQL-01/02`, `EXCH-01/02`, `FS-01/02`, `RDS-01/02`, `WIN-APP-01/02`, `PRINT-01`, `WSUS-01`, `SCCM-01` |
|
||
|
||
The Windows share is the point of the proportions. An enterprise looking
|
||
at this should recognise its own estate, which means Windows carrying
|
||
real services -- Active Directory, IIS, SQL Server, Exchange, file
|
||
shares, Remote Desktop, print, WSUS and SCCM -- rather than appearing
|
||
only as a Security channel on one box.
|
||
|
||
Two hosts carry stories the alert rules fire on and must not be moved:
|
||
`worker-02`'s disk fills at 0.04 of the volume per day, which is what
|
||
`worker-disk-filling` thresholds against, and `legacy-01` checks in once
|
||
and goes quiet, which is what `agent-legacy-01-unavailable` catches.
|
||
`api-02` is the host the outage window hits.
|
||
|
||
**Volume.** Fifty hosts generate about 316 records/minute at
|
||
`-rate-scale 1`, and the nightly reset runs at `RATE_SCALE=0.5` over a
|
||
168-hour backfill -- roughly **1.9M records per reset**, against about
|
||
0.5M when the fleet was twelve hosts. ClickHouse is untroubled by that;
|
||
what it costs is reset time and disk on the demo box. `RATE_SCALE` is the
|
||
lever if either becomes a problem, and lowering it keeps every host and
|
||
service present rather than dropping any of them.
|
||
|
||
## Prefilled login
|
||
|
||
The demo's login page comes up with the read-only `demo` account already
|
||
in both fields, so a visitor doesn't need credentials handed to them.
|
||
That's a build-time opt-in, off everywhere else: the web image is built
|
||
with `VITE_DEMO_USERNAME`/`VITE_DEMO_PASSWORD` (set in the demo host's
|
||
`docker-compose.override.yml`), and the login page prefills only when it
|
||
has both. Any deployment that doesn't set them gets the ordinary empty
|
||
form -- see `web/src/lib/api.ts`'s `demoUsername`.
|
||
|
||
The password is baked into the static bundle, which is fine for exactly
|
||
this case and nothing else: a Viewer-role account on a deployment whose
|
||
database is wiped and reseeded nightly. It has to match `DEMO_PASSWORD`
|
||
in `reset-demo.sh`, and changing that means rebuilding the web image.
|
||
|
||
## Why the demo needs a long-running process
|
||
|
||
Three things the demo has to show are only true if data keeps arriving,
|
||
and no amount of one-shot seeding fixes any of them:
|
||
|
||
- **Agents.** The Agents page is populated by the `AgentControl.CheckIn`
|
||
RPC, and marks a host stale once it stops calling in. A fleet seeded
|
||
once at 04:00 is entirely stale by 04:10.
|
||
- **Alerts.** Rules evaluate over trailing windows (`earliest=-5m`).
|
||
Against a frozen dataset every rule settles into a permanent state
|
||
within minutes and the Alerts page never moves again.
|
||
- **Recent views.** A "last 15 minutes" dashboard, or a query for what
|
||
just happened, is empty on a dataset that stopped growing overnight.
|
||
|
||
So the demo runs `demo-simulator` continuously, and `reset-demo.sh` only
|
||
handles the parts that genuinely are one-shot: the history behind the
|
||
present, and the dashboards and rules themselves.
|
||
|
||
## Editing a dashboard
|
||
|
||
Change it in the web UI, hit Export JSON, and drop the file in
|
||
`dashboards/` — the export shape and the import shape are the same one.
|
||
The next reset picks it up. (Note that panel IDs and the dashboard ID are
|
||
not part of that shape: every reset creates them fresh.)
|
||
|
||
## What the queries can't do yet
|
||
|
||
Every panel here uses `table`, `bar`, `top_n`, `single_stat`, or
|
||
`heatmap`. None uses `line`, because a line chart needs a time axis and
|
||
the query language has no time-bucketing function — `stats count by
|
||
<field>` groups by literal column values, so there's no equivalent of
|
||
Splunk's `bin`/`timechart` to group by hour or minute. Raw SQL could
|
||
express it (`toStartOfHour(timestamp)`), but dashboard panels reject the
|
||
SQL escape hatch by design, since the time-range picker works by
|
||
prepending `earliest=`/`latest=` terms to a pipe-syntax query
|
||
(see `api/dashboards/types.go`'s `validatePanel`).
|
||
|
||
That gap is the one real thing standing between these dashboards and a
|
||
conventional observability overview screen.
|