ZeroDB Lakehouse: SQL Analytics Over Your AI Agent Data

## Your AI Agents Generate Data. Now You Can Query It. Every AI agent generates a trail: inference calls, memory operations, tool executions, cost data, performance metrics. Most platforms throw this away. The ones that keep it lock it behind proprietary dashboards you can't query. ZeroDB Lakehouse gives you raw SQL access to all of it. **DuckDB over Parquet. 80+ partitions. One endpoint.** --- ## What Is a Lakehouse? A lakehouse combines the low-cost storage of a data lake (Parquet files on MinIO/S3) with the query power of a data warehouse (DuckDB SQL). No ETL pipelines. No Snowflake bills. No Spark clusters. Your agent data lands in Parquet files automatically. When you want to analyze it, you run SQL.

Traditional	ZeroDB Lakehouse
Deploy Spark + Airflow + Redshift	One API endpoint
Write ETL jobs for each data source	Data lands automatically in Parquet
$500-5,000/month for warehouse	Included in your ZeroDB plan
Wait for nightly batch jobs	Query with 5-minute cache TTL

--- ## Two Endpoints, Two Use Cases ### 1. SQL Query API (Agent Intelligence) Your agents can query their own data. Analyze performance, detect anomalies, forecast costs — with standard SQL. ```bash curl -X POST https://api.ainative.studio/api/v1/internal/lake/query \ -H "X-Internal-Key: $INTERNAL_KEY" \ -H "Content-Type: application/json" \ -d '{ "sql": "SELECT agent_id, COUNT(*) as runs, AVG(latency_ms) as avg_latency FROM agent_run_log WHERE created_at > current_date - 7 GROUP BY agent_id ORDER BY runs DESC LIMIT 10" }' ``` **Response:** ```json { "columns": ["agent_id", "runs", "avg_latency"], "rows": [ ["cos", 351, 1250.3], ["atlas", 104, 890.1], ["aurora", 87, 2100.5] ], "row_count": 3, "duration_ms": 42, "partition": "agent_run_log" } ``` 42 milliseconds. No Spark job. No waiting. ### 2. Analytics API (Platform Overview) Aggregate metrics for your dashboard — no auth required, public data. ```bash curl "https://api.ainative.studio/api/v1/public/platform/lake?days=30" ``` Returns agent runs, success rates, API usage, cost trends, top models, package downloads — everything you need for a metrics dashboard in one call. --- ## 80+ Queryable Partitions Every aspect of your AI platform is queryable. Here's what's in the lake:

Category	Partitions	Example Query
Core	agent_run_log, mcp_request_logs, rlhf_signals, billing_events, cli_telemetry	Agent success rate by day
Platform Intelligence	llm_token_usage, agent_heartbeats, agent_resource_usage, billing_aggregations, agent_traces	Token cost per model per day
Agent Operations	agent_instances, agent_tasks, agent_workload, a2a_interactions, agent_swarm_sessions	P95 task duration by agent type
Business	developer_earnings, developer_payouts, crm/deals, crm/pipeline, commerce/orders	Monthly developer payout totals
Quality & RLHF	ai_feedback, agent_audit_logs, zerodb_rlhf, agent_quality_trends, agent_performance_metrics	Quality score trend over 30 days
Predictions	request_volume_forecast, agent_anomalies, token_cost_forecast	Anomaly detection on agent behavior
External Integrations	slack, email, luma/events, dealroom/sessions, sec_filings, irs_soi	Slack message volume by channel
Sentinel (Security)	threat_intel, ransomware, exploit_intel, certificate_transparency, network_telemetry + 30 more	Threat intel alerts by severity

--- ## Real Queries, Real Answers ### "Which models cost the most this month?" ```sql SELECT model, SUM(tokens) as total_tokens, SUM(cost_credits) as total_cost FROM platform/llm_token_usage WHERE created_at > current_date - 30 GROUP BY model ORDER BY total_cost DESC LIMIT 5 ``` ### "Are any agents degrading?" ```sql SELECT agent_id, AVG(CASE WHEN status = 'success' THEN 1 ELSE 0 END) * 100 as success_pct, COUNT(*) as runs FROM agent_run_log WHERE created_at > current_date - 7 GROUP BY agent_id HAVING AVG(CASE WHEN status = 'success' THEN 1 ELSE 0 END) < 0.95 ORDER BY success_pct ASC ``` ### "What's our developer payout liability?" ```sql SELECT developer_id, SUM(amount_cents) / 100.0 as payout_usd, COUNT(*) as transactions FROM platform/developer_earnings WHERE status = 'pending' GROUP BY developer_id ORDER BY payout_usd DESC ``` ### "How is agent swarm activity trending?" ```sql SELECT DATE_TRUNC('day', created_at) as day, COUNT(DISTINCT session_id) as sessions, COUNT(*) as messages FROM platform/agent_swarm_messages WHERE created_at > current_date - 14 GROUP BY day ORDER BY day ``` --- ## Security: SQL Without the Risk The lake query endpoint validates every query before execution:

Protection	How
No writes	DROP, DELETE, INSERT, UPDATE, ALTER, CREATE all blocked
No file access	read_csv, read_json, COPY, ATTACH blocked
Partition whitelist	Only listed partitions accessible — no arbitrary S3 paths
No system functions	information_schema, pg_catalog, system calls blocked
Query length limit	Max 4,000 characters per query

Your agents get SQL power without the ability to damage anything. --- ## Architecture ``` Your SQL query ↓ API validates → blocks writes, file access, system calls ↓ Rewrites FROM {partition} → read_parquet('s3://ainative-lake/raw/{partition}/date=*/*.parquet') ↓ DuckDB executes in-process (no external DB) ↓ Results in JSON (columns + rows) ``` Two MinIO buckets: - **ainative-lake** — Core platform data with `date=YYYY-MM-DD` partition structure - **sentinel-lakehouse** — Security data with `union_by_name` for schema evolution DuckDB runs embedded in the API process. No separate warehouse service. Sub-100ms for most queries. --- ## How Agents Use the Lakehouse The lake query endpoint is designed for agents — not just humans. An agent can: 1. **Monitor its own performance**: Query `agent_run_log` to check success rates 2. **Detect anomalies**: Query `derived/predictions/agent_anomalies` for flagged behavior 3. **Forecast costs**: Query `derived/predictions/token_cost_forecast` for budget planning 4. **Audit actions**: Query `platform/agent_audit_logs` for compliance 5. **Coordinate with other agents**: Query `platform/a2a_interactions` for communication history The endpoint returns structured JSON that agents can parse and act on — no human interpretation needed. --- ## Getting Started The Lakehouse is available on all ZeroDB plans. **Public analytics (no auth):** ```bash curl "https://api.ainative.studio/api/v1/public/platform/lake?days=30" ``` **SQL queries (requires internal key):** ```bash curl -X POST https://api.ainative.studio/api/v1/internal/lake/query \ -H "X-Internal-Key: $KEY" \ -d '{"sql": "SELECT COUNT(*) FROM agent_run_log"}' ``` - **Documentation**: [docs.ainative.studio/docs/api/data-lake](https://docs.ainative.studio/docs/api/data-lake) - **API Reference**: [docs.ainative.studio/docs/api/overview](https://docs.ainative.studio/docs/api/overview) - **Tool Discovery**: `GET /api/v1/public/tools/catalog` --- *Published June 12, 2026 by the AINative Engineering Team*

ZeroDB Lakehouse: SQL Analytics Over Your AI Agent Data

Check your site's AX Score