ClickHouse Population Survey (2026-05-16)

NuClide Research · 2026-05-16 (Survey 7 of the day’s 10-category batch, largest single survey) Closes: category 02 / specialty data layers. ClickHouse half

Summary

Largest single-platform population survey of the day. ClickHouse is the OLAP database that powers most modern observability stacks (SigNoz, Plausible, PostHog, Helicone, Phoenix-on-OTLP). Wherever an AI/LLM service emits traces or analytics, there’s often a ClickHouse behind it.

DCWF KSAT coverage

Auto-derived from DCWF AI work-role rule files (ksat-tag).

• 672 (AI Test & Evaluation Specialist): K7003, K7004, S7068, S7070, S7075, T5858, T5904, T5919
• 733 (AI Risk & Ethics Specialist): K7040, K7051, T5854, T5868, T5893
• overlap (Common AI KSATs (all 5 roles)): K108, K1158, K1159, K22, K6311, K7003, K942

• 65,100 candidate ip:port pairs harvested via product:"ClickHouse" Shodan facet (harvest capped at ~65K)
• Probed via fast_enum_clickhouse.py (threads=250, ~35 min)
• 1,832 confirmed unauth ClickHouse instances (2.81% real-rate at population scale)
• 7,337 auth-gated (11.3%)
• 2,563 partial-open (/ping returns 200 but SHOW DATABASES gated)
• 35,169 dead (54%)
• 18,199 unknown (28%, many returned 200 to /ping but other endpoints hung)

Headline: ~10K real ClickHouse instances reachable, 18% unauth. The user-database names disclose the observability stack of the operator, SigNoz traces/metrics/logs (21 each), PostHog (8), Plausible Analytics (11), OpenTelemetry (6), and via those names, the operator’s entire LLM observability pipeline.

What ClickHouse leaks unauth

/?query=SHOW DATABASES FORMAT JSON returns the full list of databases. Each database name discloses the operator’s application stack. Top patterns:

  32  analytics                  (generic - any analytics workload)
  21  signoz_traces              \
  21  signoz_metrics             |  SigNoz observability platform (often AI-stack obs)
  21  signoz_logs                /
  11  plausible_events_db        Plausible Analytics
   8  posthog                    PostHog product analytics
   8  netflow                    network flow data (high sensitivity)
   8  _temporary_and_external_tables
   7  ftacs_qoe_ui_data          FTACS Quality-of-Experience (Huawei FTACS network ops)
   6  otel                       OpenTelemetry traces (often LLM call traces)
   6  geo                        geographic data

The SigNoz 21×21×21 trinity (traces/metrics/logs) is interesting. It’s 21 operators all running the same observability stack, each leaking the three-database pattern unauth. SigNoz is the leading open-source observability platform with first-class LLM-tracing support (it ingests OTLP semantic-conventions for genai). The 21 unauth SigNoz deployments are very likely AI-stack observability backends with LLM traces visible.

The 6 explicit AI-stack databases

Operator-attribution items:

• ai_hedge_fund: the name discloses the operator’s entire business model. An AI-driven hedge fund running ClickHouse unauth with that database name reachable means anyone can query their trading data.
• scentedai_fragid_new: Scented AI (https://www.scentedai.com), a fragrance-identification AI startup, running their fragrance-ID database unauth on a Hetzner IP. Operator product disclosure via DB name.
• vllm_service: operator running vLLM (open-source LLM inference) with ClickHouse as the trace backend. Unauth = full LLM call history visible.

Version distribution (the Docker-image-template phenomenon, again)

 1013  v22.3.20.29   ← single-version dominant cluster (55% of unauth fleet)
  110  vNULL          (version-query failed)
   42  v24.8.14.39
   28  v22.1.3.7
   24  v23.8.16.16
   22  v24.10.2.80
   19  v26.3.9.8     ← bleeding edge
   19  v24.3.18.7
   16  v24.1.8.22
   16  v22.2.2.1

Same pattern as Solr 7.6.0 (516 hosts) and Elasticsearch 7.17.0 (239 hosts) from today’s batch: a single-version Docker image dominates the unauth population. 1,013 hosts on ClickHouse 22.3.20.29 is the largest single-version cluster of the day. That version is the LTS line from late 2022. Operators using clickhouse/clickhouse-server:22.3 or :lts (which pinned to 22.3.x for years).

The Docker-image-template phenomenon is now multi-survey-confirmed and warrants its own Insight (queued).

Methodology placement

ClickHouse is *Tier-A (auth optional, off-by-default in Docker image)**. The official clickhouse/clickhouse-server Docker image ships with a default user account named default with NO password. Operators must:

1. Set CLICKHOUSE_USER + CLICKHOUSE_PASSWORD env vars at container start, OR
2. Modify /etc/clickhouse-server/users.xml to set a password for default

The 1,832-host unauth population shows ~18% of reachable ClickHouse operators skip both steps. Auth-gated 7,337 + partial-open 2,563 = 9,900 operators DID set up auth correctly.

Adds ClickHouse (Docker default) to the Tier-A* family alongside Elasticsearch (today), Airflow, LiveKit example-template, and DCGM-exporter. The Tier-A* family is now 5 platforms strong.

Cross-survey colocation

Operator pattern hypothesis: hosts running SigNoz (21 instances) likely have ClickHouse + Elasticsearch + a Phoenix/Helicone/Langfuse frontend on the same /24. IP-direct-shadow on the 21 SigNoz hosts is the next step for cross-platform stacking.

Toolchain Provenance

0. shodan download (product:"ClickHouse") → 65,100 unique ip:port
1. fast_enum_clickhouse.py (threads=250, ~35 min) → 1,832 unauth + 7,337 auth-gated
2. operator-DB enumeration via SHOW DATABASES FORMAT JSON → ~10K user-DB-name samples
3. AI-stack marker filter (25-substring list) → 6 explicit AI-stack hosts
4. operator-attribution from DB names → scentedai, ai_hedge_fund, vllm_service, qinghai_platform
5. (queued) visorlog ingest → 1,832 events into nuclide.db source='clickhouse-survey-2026-05-16'
6. (queued) BARE module ranking → ClickHouse has CVE-2021-43304/05/08 (heap-buffer-overflow + integer overflow) but unauth-RCE class is small — risk class is intel-disclosure, not RCE

Honest negative space

• 35,169 dead hosts is consistent with the Shodan-stale pattern observed all day. ClickHouse operators rotate ports/migrate at similar rate to Vault (~54% stale here vs 64% Vault).
• AI-stack filter is conservative. 6 explicit AI-stack DBs is the lower bound. Many of the 32 analytics + 21 signoz_* + 8 posthog databases probably back AI workloads (LLM call traces, RAG performance metrics, embedding-evaluation results) without that being visible in the DB name. Real AI-stack overlap is probably 200-500 hosts.
• No table-name enumeration. SHOW TABLES IN <db> would extract operator’s table names, even higher-resolution operator attribution, but at population scale that’s 1,832 × N tables of additional query load. Deferred.
• No data sampling. Restraint: never SELECT ... FROM <table>. The database/table names ARE the finding.
• 22.3.20.29 dominance suggests Docker-image-template phenomenon, but I didn’t verify which image (clickhouse/clickhouse-server:22.3 vs altinity/clickhouse-server vs others). Different operators may have used different images that happened to pin to 22.3.x. Follow-up: check Server header on the /ping responses for the 1,013 hosts.

Disclosure posture

• Aggregate Tier-A disclosure recommended* to ClickHouse Inc. (Aiven-style upstream notification about the Docker-image default auth behavior).
• Targeted per-host disclosure for the 6 AI-stack-tagged hosts. The ai_hedge_fund and scentedai_fragid_new hosts are operator-attribution-rich and deserve coordinated outreach.
• The 21 SigNoz-deployments are interesting collectively as a SigNoz-template-phenomenon. Disclose to SigNoz upstream as a “your default Docker compose ships ClickHouse without auth” finding.

See also

• [[insight-13-shipping-defaults-are-load-bearing]]. ClickHouse Docker image (no default password) drives the 18% unauth rate
• [[insight-25-falsification-confirmation-tier-c-platforms]]. ClickHouse joins the multi-platform Tier-A* set; same-day Tier-C contrast: ClearML, W&B, Argilla, Mem0
• [[insight-26-shodan-facet-fp-rate-escalates-with-token-commonality]], product:"ClickHouse" 65K → 1,832 unauth ≈ 2.81% (Shodan facet not particularly noisy here, “ClickHouse” is a distinctive brand-string)
• elasticsearch-ai-stack-population-survey-2026-05-16.md: same-day Tier-A* companion (Docker-default-off pattern)
• vectordb-stragglers-population-survey-2026-05-16.md: same-day Solr 7.6.0 finding (parallel Docker-image-template phenomenon)