Google Cloud ClickHouse
How InstantML-hosted storage uses InstantML-owned ClickHouse on Google Cloud.
InstantML-hosted workspaces store run metadata and metrics in an InstantML-owned ClickHouse deployment on Google Cloud, while account records live in a separate control-plane Postgres database.
Premium workspaces that want product data in their own infrastructure can use customer-owned ClickHouse instead.
Hosted topology
| Source | Service | Storage |
|---|---|---|
| Browser / dashboard | Control and data services | Control-plane Postgres (account records) + InstantML-owned ClickHouse (tenant data) |
| Python SDK / uploader | Data service | InstantML-owned ClickHouse on Google Cloud |
| Artifact upload/download paths | Data service | Private object-storage artifact buckets |
The control service stores account, session, billing, API-key, and tenant-route records in Postgres. The data service stores projects, runs, metrics, console logs, attributes, artifacts, reports, imports, and usage snapshots for the authenticated workspace in its routed ClickHouse database.
Tenant isolation
Tenant databases are routed by workspace. Each tenant database owns its product tables and metric tables, and account records stay in the separate control-plane database.
Hosted artifact bytes are not stored in ClickHouse. They live in private per-workspace object-storage buckets; ClickHouse stores only the artifact metadata: byte count, hash, MIME type, and an opaque storage reference. Downloads stream through the InstantML API rather than exposing bucket URLs.
Usage accounting follows the same boundary. Metric-point counts are always bound to the resolved workspace. Exact warehouse-storage bytes are counted only when a ClickHouse database is scoped to a single workspace; shared cells and customer-owned deployments use workspace-local artifact and metadata accounting instead.
Production and staging environments run as separate services with separate control-plane databases, so staging traffic cannot write production records.
What this means for you
- The SDK and dashboard APIs are the same regardless of which storage path your workspace routes to.
- Run-list and chart endpoints stay bounded; raw metric history is fetched through dedicated series endpoints.
- Artifact uploads and downloads always go through InstantML API routes.
- Customer-owned GCP ClickHouse remains available for Premium teams.
Current benchmark signal
The hosted GCP ClickHouse path has been benchmarked through the production data service against the normal-runs-50k showcase project: 50,000 runs and 522,000,000 metric points (measured 2026-05-23).
| Query | p95 |
|---|---|
| Project newest-100 run list | 236 ms |
| Metric-best sort | 307 ms |
| Project overview | 418 ms |
| 1,000-point chart read from a 20,000-step series | 224 ms |
See Benchmarks for the full methodology, fairness notes, and comparison data.