The honest daily record of what the vaked inference node did — PII-free usage, published daily at coder.vaked.dev/blog. 🜂
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Peter Lodri dc0b48456f feat: finish the almanac — render status/route mix; deploy live to coder.vaked.dev/blog
publish.py now renders the status mix (by class) and the route breakdown the
collector already stored but never showed — closing the gap between the README's
promises and the page. Adds all-time stats (requests / bytes served / days) and
humanised byte figures; empty-state stays honest.

Deployed on the node 2026-07-25: the Caddy access log is enabled with a `format
filter` that DELETES remote_ip/remote_port/client_ip/headers/resp_headers at the
log layer — so "no IPs, no tokens" is structural, not just collector behaviour.
Scripts installed under /opt/vaked/almanac; timer fires daily at 18:00
Europe/Budapest. /blog now serves real metadata — the placeholder is retired.

README honest-state flipped to LIVE and deploy/README corrected to the config that
was actually shipped (incl. the caddy-must-own-the-logfile EACCES gotcha that
`caddy validate` can't catch).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-25 05:06:05 +02:00
deploy feat: finish the almanac — render status/route mix; deploy live to coder.vaked.dev/blog 2026-07-25 05:06:05 +02:00
.gitignore feat: vaked-almanac — daily public usage record (SQLite timeseries, PII-free) 2026-07-24 09:14:20 +02:00
collect.py feat: vaked-almanac — daily public usage record (SQLite timeseries, PII-free) 2026-07-24 09:14:20 +02:00
LICENSE feat: vaked-almanac — daily public usage record (SQLite timeseries, PII-free) 2026-07-24 09:14:20 +02:00
publish.py feat: finish the almanac — render status/route mix; deploy live to coder.vaked.dev/blog 2026-07-25 05:06:05 +02:00
README.md feat: finish the almanac — render status/route mix; deploy live to coder.vaked.dev/blog 2026-07-25 05:06:05 +02:00

vaked-almanac 🜂

The honest daily record of what the node did.

Every day at 18:00 Europe/Budapest, the vaked inference node publishes a PII-scrubbed summary of its own usage to coder.vaked.dev/blog — request counts, models served, status mix, latency, bytes. Not marketing numbers. The actual metered reality, in the open, because entropy cannot lie — and neither will the usage page.

An almanac is a periodical that publishes tables of observed data. That's exactly what this is: the node observing itself, out loud, on a schedule.


What it publishes (and what it never does)

Publishes — metadata only:

  • daily request count, and the running history (the timeseries)
  • breakdown by model and by HTTP status
  • latency p50 / p95, total bytes served
  • a small sparkline of daily volume over time

Never touches — by construction, not by promise:

  • ❌ prompt or completion bodies (they're never read by the collector)
  • ❌ IP addresses (aggregated to counts; never stored per-address)
  • ❌ any per-user identity

The collector reads only Caddy's JSON access log for /v1/*, which contains request metadata — method, path, status, size, duration — never payloads. So the "PII scrubbing" is structural: there is no PII in the input to begin with.


The "best timeseries DB" question — answered honestly

You asked for the best timeseries database. The honest answer for one box at this volume is SQLite — and a dedicated TSDB would be the wrong call:

why
SQLite (chosen) one file, zero ops, atomic writes, ships with Python. Daily aggregates = a handful of rows a year. It is the timeseries.
InfluxDB / Timescale / QuestDB built for millions of points/sec and high-cardinality tag queries — none of which one CPU box publishing one row a day has. Pure ops overhead here.

The store is almanac.db: one row per day of aggregates. publish.py regenerates the whole /blog page from that history. The day we shard across many nodes and need high-cardinality, cross-node queries, the honest upgrade is TimescaleDB (Postgres extension) or QuestDB — not before. Right tool for the reality that exists, nothing more.


Pieces

file does
collect.py reads the last 24h of Caddy /v1 access logs → aggregates → upserts one row into almanac.db
publish.py renders /var/www/blog/index.html from the full almanac.db history (pocoo-styled)
deploy/almanac.service + .timer systemd timer firing daily at 18:00 Europe/Budapest
deploy/README.md how to enable the Caddy /v1 access log + install the timer

Run it by hand

ALMANAC_DB=./almanac.db ALMANAC_LOG='/var/log/caddy/v1-access.log*' python3 collect.py
ALMANAC_DB=./almanac.db ALMANAC_OUT=./index.html python3 publish.py

Honest state (2026-07-25) — LIVE

Deployed on the node and collecting. As of 2026-07-25 the Caddy access log is enabled, collect.py + publish.py are installed under /opt/vaked/almanac, and the almanac.timer fires daily at 18:00 Europe/Budapest. coder.vaked.dev/blog now shows real metadata — the placeholder is gone. History accrues from today (2026-07-25 day one); the collector does not reconstruct traffic from before the log was enabled.

Two honesty/privacy details that matter, made structural at the log layer:

  • No IPs — by construction. The Caddy log uses a format filter that deletes remote_ip, remote_port, client_ip, headers, and resp_headers before anything is written to disk. So "no IPs, no tokens" is a property of the log file itself, not just of the collector's good behaviour. (Deleting headers also means a bearer API key in an Authorization header is never written down.)
  • The log file is caddy-owned. Caddy runs as the caddy user (uid 996); the log at /var/log/caddy/v1-access.log must be owned by caddy or the daemon gets EACCES on reload. The vaked collector reads it via membership in the caddy group (mode 0640). See deploy/README.md for the exact, deployed config.

🜂 AHOGY A DOLGOK VANNAK — as things are. Nothing more, nothing less.