Architecture in 60 Seconds
HatiData is built on a two-plane architecture that separates data processing from management. This page gives you the full picture in under a minute.
Two Planes
┌────────────────────────────────────────────────┐
│ Your Infrastructure │
│ │
│ AI Agents ──→ Data Plane (HatiData Proxy) │
│ │ Postgres wire protocol │
│ │ SQL transpilation │
│ │ Query execution │
│ │ Memory + CoT + Triggers │
│ │ Multi-tier caching │
│ ▼ │
│ Object Storage (your bucket) │
└────────────────────────────────────────────────┘
│
PrivateLink / HTTPS
│
┌────────────────────────────────────────────────┐
│ Control Plane (HatiData) │
│ Auth · Policies · Billing · Audit · Dashboard │
└────────────────────────────────────────────────┘
Data plane — Runs where your data lives. Accepts Postgres connections, transpiles Snowflake SQL for execution, and provides agent-native capabilities (memory, chain-of-thought, triggers, branching). In Local mode, it runs on your machine. In Enterprise mode, it runs in your VPC.
Control plane — Manages everything that is not query execution: authentication, policy management, billing, audit, and the dashboard. Communicates with the data plane over PrivateLink (Enterprise) or HTTPS (Cloud).
Multi-Stage Query Pipeline
Every SQL query flows through a multi-stage pipeline. Security and compliance are built in — not bolted on:
| Stage | What Happens |
|---|---|
| Admission | Concurrency control — queue, don't reject |
| Table extraction | Identifies all referenced tables |
| Policy check | ABAC evaluation against agent identity, role, time, tables |
| Cost estimation | Heuristic credit cost based on joins, scans, aggregations |
| Quota check | Per-org credit limits enforced |
| Row-level security | WHERE clauses injected based on agent context |
| Transpilation | Snowflake SQL auto-adapted for execution |
| Snapshot pin | Repeatable reads via point-in-time snapshot pinning |
| Execution | Columnar engine with vectorized processing |
| AI healing | Optional: retry failed queries with AI-suggested corrections |
| Column masking | Sensitive columns masked by role (Admin → full, Agent → redacted) |
| Metering | Credit usage recorded, metrics emitted |
| Audit | PII-redacted, cryptographically hash-chained audit entry written |
Agent-Native Capabilities
Beyond SQL, HatiData provides five capabilities through the same Postgres connection and 24 MCP tools:
| Capability | What It Does | MCP Tools |
|---|---|---|
| Identity | Per-agent authentication, scoping, billing | — (via connection params) |
| Memory | Persistent SQL + vector hybrid search | 5 tools |
| Chain-of-Thought | Immutable hash-chained reasoning traces | 3 tools |
| Triggers | Semantic concept-matching with 4 action types | 4 tools |
| Branching | Copy-on-write schema isolation with merge | 5 tools |
Plus 7 core query tools: run_sql, run_sql_readonly, list_schemas, list_tables, describe_table, run_sql_arrow, get_usage_stats.
Three Deployment Tiers
| Tier | Data Location | Price | Best For |
|---|---|---|---|
| Local | Your machine | Free | Development, prototyping, CI |
| Cloud | HatiData-managed | $29/month | Small teams, startups |
| Enterprise | Your VPC | Custom | Regulated industries, production |
The same SQL, SDKs, and agent code work across all tiers. The only change is the connection string.
Caching (Three Tiers)
| Tier | Backing | Latency | Eviction |
|---|---|---|---|
| Memory | In-memory | Sub-microsecond | TTL + LRU |
| Disk | High-performance local storage | Sub-millisecond | TTL + size cap |
| Object | Cloud object storage | 10–100ms | Snapshot-based |
Transpilation results are cached separately by SQL hash — repeated queries skip the parse-and-rewrite pipeline entirely.
Next Steps
- What is ANDI? — The category definition
- Quickstart — Install and run your first query
- Two-Plane Model — Full architecture deep dive
- Query Pipeline — All pipeline stages in detail