Network
Traffic intelligence
The panel's Traffic page is a control room for who visits the apps: where they come from (country, network), which server and app they reach, on what (browser, platform, device), and whether they're people or bots, down to the addresses and botnet campaigns behind the threats.
What's recorded
Every request to the apps is classified and summed per minute across all
servers. The page runs about a minute behind. The same data is at
GET /api/intel.
Only these request headers are kept: User-Agent, X-Real-Ip, Referer,
Accept-Language, Sec-Fetch-Mode and CF-IPCountry. Query strings are dropped
from paths, and long numeric, UUID and hex segments become :id.
GeoIP
Countries and networks (ASN) come from the DB-IP Lite
databases (CC BY 4.0, no key). Each server downloads them to
/var/lib/stackd/geoip/ (set with STACKD_GEOIP_DIR) and refreshes them
every 30 days; a failed download keeps the old copy. STACKD_GEOIP=false
turns lookups off. Behind Cloudflare, its CF-IPCountry header is used too,
but only on connections from Cloudflare's ranges.
Classes
Every request gets one class:
| Class | What |
|---|---|
| human | A browser with nothing suspicious about it. |
| crawler | Search and social crawlers. Google, Bing, Apple, Meta, Yandex and Baidu must come from their own networks. |
| ai | AI crawlers and assistants (GPTBot, ClaudeBot, PerplexityBot, CCBot, Bytespider, …). |
| monitor | Uptime monitors and probes. |
| tool | Scripts and HTTP libraries (curl, python-requests, Go, okhttp, …). |
| scanner | Scanners by user agent, and probes for well-known exploits (/.env, /.git, traversal, /wp-login.php on a site that isn't WordPress, …). |
| suspicious | Headless browsers, fake crawlers, an empty user agent, floods (over 300 requests a minute from one address on a browser user agent, 600 otherwise), 10 or more failed logins a minute, or two weak signals. |
| botnet | Part of an active campaign (below). |
Weak signals only count for browser user agents:
datacenter: a hosting network.no_fetch_meta: a modern browser over HTTPS withoutSec-Fetch-Mode.no_accept_language.
One weak signal is recorded but leaves the request human. Crawler, ai and monitor requests keep their class, though their signals are still recorded.
Campaigns
A campaign is the same request (app, method, normalized path, user agent) seen over the last 10 minutes from at least 25 addresses on at least 8 networks, where either:
- at least 60% of its requests carry a bot signal (one that isn't just a script user agent), or
- it is a POST to a login-like path with at least half of the requests answered 4xx (credential stuffing).
While a campaign is active, its requests are classed botnet.
The firewall acts on what the intel finds (Firewall): it bans floods, brute-forced logins and path enumeration, challenges suspicious and botnet visitors, and challenges the requests of active campaigns.
Retention
| What | Kept |
|---|---|
| Requests, errors, bytes and humans per app, server, country and the other dimensions | 6 hours by minute, 31 days by hour |
| Distinct addresses per app, server and minute | 6 hours |
| Per app, server and minute: the busiest addresses, then any non-human ones, up to 100 | 48 hours |
| Active campaigns | 7 days |
Data can stay up to an hour or a day longer than listed.
Busy dimensions are capped per app and minute (networks 50, paths 40,
referrers 25, hosts 20); the rest shows as (other). An IPv6 /64 gets at
most 8 addresses of its own in a minute; any more from it are listed as one
visitor.