Limitations¶
FlareHound states its limits up front rather than letting you discover them.
Data¶
- Sampled analytics. Cloudflare's adaptive GraphQL analytics are sampled and extrapolated. On small sites fine-grained breakdowns can be unstable, and a low-volume path may return 0 rows — zero does not mean absent. Very large sites will find attribution too coarse; FlareHound targets small-to-medium sites.
- Different units from the dashboard. Adaptive counts include Worker subrequests and cache operations. On a live Free zone FlareHound measured ≈ 1.29× the dashboard's request count and ≈ 1.84× its edge 5xx (cache-operation artifacts). Thresholds are in FlareHound's units; see Calibration.
- Ingestion lag (about 5–10 minutes) sets the detection floor. This is a watchdog, not a real-time shield — edge rules remain the shield.
- No referer dimension. Traffic cannot be split by referring page, so "forged referer bypass" cannot be detected from analytics.
- Worker CPU is available only as quantiles (FlareHound uses P50), not as a total; any cost reading is an estimate.
- Retention. Cloudflare keeps about 31 days.
-backfillcan recover at most the last 30 days; anything older cannot be rebuilt. The archive holds hourly aggregates, never individual requests. - Database files level off, older ones never shrink. Pruned rows free pages that new rows reuse. Only
archive.db/console.dbfiles created by v1.18 or later give space back to the disk after a large prune (for example after loweringarchive.hourly_retention_days); files from earlier versions are not rebuilt and keep their size. See disk space.
Plans¶
What analytics expose depends on the plan. FlareHound probes each zone and degrades instead of failing. Verified on a live Free zone:
| Capability | Free | Effect when missing |
|---|---|---|
clientAsn |
not available | No ASN Top lists, no new-ASN detection, no ASN suggestions; the daily report says so |
clientASNDescription |
not available | Network type from the built-in table only |
| Firewall events dataset | not available | Firewall events, block attempts, audit hit statistics and --observe show "n/a" — never 0 |
verifiedBotCategory |
available | Used for crawler verification |
| Host and country dimensions | available | — |
Rule quotas also differ by plan (custom rules Free 5 / Pro 20 / Business 100 / Enterprise 1,000; rate-limit rules 1 / 2 / 5 / 100), and on Free the rate-limit period and mitigation timeout must be 10 seconds.
Detection¶
- Hard thresholds. Slow growth (+5% a day) and attacks kept just under the line stay invisible; a legitimate surge (a launch, a viral post) looks the same as an attack in the numbers.
- Baseline problems. If yesterday itself was abnormal, the comparison is distorted. The 4-week median baseline (once the archive has enough samples) reduces this, but a long-lasting anomaly can still become the new normal.
- No semantics. Rules cannot know whether a new ASN is a crawler fleet or your own new service. Register known sources under
known_faces. - Per-country and per-range shares are not measured when blocking; FlareHound says so in the plan.
- IP location is a library's reading, not a fact. It is the registered/routed country of an address — a residential proxy shows its exit, a data-centre address its data centre — and libraries disagree at the edges (DB-IP lite places Google's IPv6 resolver
2001:4860:4860::8888in CA, GeoLite2 in US). There is deliberately no per-country traffic view built from it: countries inferred from the top IPs would be a heavily biased sample. - No AI in the detection path: findings are deterministic and reproducible from the same data.
Scope¶
- Cloudflare only. Origin internals (processes, databases) are out of scope.
- One process per deployment; runs on your own machine. A Cloudflare Workers deployment is not provided.