Observability v2 — ClickHouse only

One store, one agent, a straight line. Customer logs and metrics both live in the per-region ClickHouse droplet — no VictoriaMetrics, no vmagent, no Vector.

The pipeline has three stops: Collect — one OpenTelemetry Collector grabs app logs, Envoy request logs and cAdvisor meters and tags each row with its tenant; Store — a few ClickHouse tables hold it, with retention set from the customer's plan; Read & query — Starbase queries behind a forced tenant filter and Stardeck renders. The 7 metrics are SQL over the rows (four from the request log, three from the scraped meters); logs are a tail. Identity on request rows rides the HTTPRoute name (read from route metadata and split once — not v1's per-scrape envoy_cluster_name regex).

VPC topology

Diagram — v2 VPC topology. The regional VPC (one per region) holds the customer cluster and its single ClickHouse telemetry droplet, so ingest is intra-VPC at $0. One OTel Collector agent per cluster ships app logs, Envoy access logs and cAdvisor metrics to the regional OTel Collector gateway, which batches into ClickHouse — no VictoriaMetrics, no Vector, no vmauth. The central control plane (its own VPC) peers in to read, only through one authed front-door: a read-only ClickHouse user (native RBAC, not a service). Starbase injects the tenant filter (FR-065); the store stays private. Grafana Alloy ships platform series to Grafana Cloud (external) — unchanged. Blue = telemetry / reads · amber = platform · dashed = external.

What's covered, and where

Every requirement, its v2 home, and its status:

Area	Requirement	v2 home	Status
Metrics — latency, RPS, error rate, throughput (L7)	FR-063	Read & query (from otel_http)	MVP
Metrics — CPU, memory, network	FR-050	Read & query (scraped otel_metrics_*)	MVP
Logs — app runtime, live tail (build logs → §16.6, outside ClickHouse)	FR-049	Read & query (otel_logs)	MVP
Attribution — tenant identity on every row	FR-064	Collect (labels / route metadata)	MVP
Isolation — server-side tenant filter	FR-065	Read & query	MVP
Retention — per plan (7/30/90 d)	FR-051	Store (dictionary + TTL)	MVP · fixed
Transport — private VPC, cross-region reads	FR-071	this page + Read & query	MVP (x-region dormant)
Preview-env metrics — no per-route cardinality blowup	FR-067	rows in otel_http, no active-series limit	MVP
Platform self-monitoring	FR-069	§35.5 (shared section, unchanged)	MVP
Rollup materialized views	—	Store (deferred; raw-first)	post-MVP
Distributed tracing	—	— (OTel-native when needed)	future

What changed from v1

	v1 (superseded)	v2
Stores (per region)	VictoriaMetrics + ClickHouse	ClickHouse only
Metrics agent	vmagent	OTel Collector
Logs agent	Fluent Bit → Vector	OTel Collector
L7 identity	envoy_cluster_name regex, per scrape (EG-version-coupled)	HTTPRoute name from route metadata, split once
Per-plan retention	✗ — metrics global 90 d (FR-051 gap)	✓ — project → days dictionary + TTL (logs and metrics)
Read front-doors	vmauth + ClickHouse user	ClickHouse user
Query languages	PromQL + SQL	SQL

Build order

The throughline is get → store → read: collect the signals, land them in ClickHouse, read them out behind the tenant filter. Each page walks its stop end-to-end with the configs to copy.

1. Collect — the one agent — one collector, three inputs, identity stamped, batched into ClickHouse.
2. Store & retention — the ClickHouse tables, tenant keys, plan-based retention.
3. Read & query — the tenant filter, the 7 metrics as SQL, and log tail.
4. Bootstrap — the ordered bring-up that ties it together.