Evidence-first database performance intelligence for PostgreSQL and Snowflake.
Detect regressions. Investigate plans. Produce bounded recommendations. Verify the outcome.
Database performance tools often stop at a chart, and AI tools often start guessing before they have enough evidence. DataCurrent joins the full investigation loop while keeping the distinction between observation, heuristic, hypothesis, and verified result explicit.
| Detect | Investigate | Recommend | Verify |
|---|---|---|---|
| Collect real workload statistics and identify regressions | Correlate history, structured plans, schema, indexes, and lineage | Generate cited, confidence-scored proposals with tradeoffs | Compare candidates in isolation with median, p95, and result agreement |
The product includes a polished landing experience, a working product workspace, production APIs, scheduled collectors, and a reproducible commerce workload. PostgreSQL and Snowflake are first-class providers—not visual placeholders.
Visit the public product site →
- Workload overview — current health, regressions, latency trends, and expensive fingerprints.
- Query Explorer — redacted normalized SQL, historical metrics, provider context, and evidence state.
- Query Lab — read-only plan analysis with Monaco and structured plan findings.
- AI Investigator — OpenAI Agents SDK, narrow evidence tools, typed output, citations, confidence, and limitations.
- Data Lineage — catalog-backed and parsed relationships with evidence classification.
- Benchmarks — isolated, repeated comparison with result digests and visible methodology.
- Connections — encrypted PostgreSQL and Snowflake configuration with capability checks.
- Operations — Celery scheduling, Redis locks, retention, request IDs, health checks, and Docker packaging.
The control plane never sends database credentials or arbitrary row data to the model. Database connections are encrypted at rest, collection uses read-only roles, AI receives minimized authorized evidence through narrow tools, and empirical verification requires an explicitly marked benchmark connection.
The public product surface is served at datacurrent.jonzhang.dev through Porkbun-managed authoritative DNS and the Cloudflare-backed Sites edge. The containerized application runtime is a separate deployment surface, so the public experience remains available without placing the database control plane in the request path.
Read the architecture, security model, and operations guide.
Requirements: Docker Desktop with Compose v2. That is enough for the full stack.
git clone https://github.com/jonxsong/DataCurrent.git
cd DataCurrent
cp .env.example .env.local
docker compose up --buildThen open:
- Product:
http://localhost:3000 - API health:
http://localhost:8000/api/v1/healthz - OpenAPI:
http://localhost:8000/docs - Seeded demo PostgreSQL:
localhost:55432
The included demo database enables pg_stat_statements and creates deterministic customers, products, orders, payments, inventory, and event data. Five intentionally interesting scenarios live in benchmarks/scenarios.
For AI investigations, set OPENAI_API_KEY in .env.local. Never commit that file. The rest of DataCurrent continues to work without an API key.
DataCurrent uses the official Snowflake Python connector and supports:
- encrypted connection configuration and a bounded connectivity test;
SNOWFLAKE.ACCOUNT_USAGE.QUERY_HISTORYworkload collection;INFORMATION_SCHEMAcatalog discovery;- read-only
EXPLAIN USING JSONin Query Lab; - provider-aware lineage and isolated benchmark execution.
The recommended role grants only warehouse usage, database/schema usage, imported privileges on the SNOWFLAKE database, and the minimum object privileges needed for catalog inspection. See the exact setup SQL and validation steps in the Snowflake guide.
# JavaScript dependencies
corepack enable
pnpm install --frozen-lockfile
# Python environment (Python 3.12)
uv sync --frozen --all-packages --group dev
# Complete local quality gate
uv run ruff format --check apps/api workers/collector
uv run ruff check apps/api workers/collector
uv run pytest apps/api/tests -q
pnpm --filter @datacurrent/web lint
pnpm --filter @datacurrent/web typecheck
pnpm --filter @datacurrent/web test
pnpm --filter @datacurrent/web build
pnpm --filter @datacurrent/web test:e2eCI runs the same Python, TypeScript, unit, browser, and image-build checks. Dependabot maintains npm, uv, Docker, and GitHub Actions dependencies.
The public product site is live on the custom domain through the no-cost Sites edge. The production profile for the full application runs the real Next.js, FastAPI, Celery, PostgreSQL, and Redis services behind automatic HTTPS; its Google Compute Engine rollout is deployment-ready but intentionally deferred until cloud compute is desired. See the deployment guide for provisioning, secrets, DNS, backups, verification, and rollback.
apps/api/ FastAPI control plane, connectors, analysis, AI, migrations
apps/web/ Next.js landing page and application workspace
workers/collector/ Celery collection, retention, and regression jobs
infrastructure/docker/ Reproducible PostgreSQL development environment
infrastructure/deployment/ Hardened HTTPS production deployment
benchmarks/scenarios/ Deterministic workload investigations
docs/ Architecture, security, Snowflake, and operations
site-preview/ Public .chatgpt.site source package
- Evidence before explanation. Every factual finding cites its source.
- Read-only by construction. Privileges, parsing, timeouts, host policy, and isolated targets enforce the boundary.
- AI proposes; people decide. The model cannot execute schema changes.
- Measurement is labeled. Cumulative statistics are not presented as traces, and a recommendation is not a performance claim.
- Verification changes the status. Measured results, result agreement, environment, and method remain attached to the proposal.
