Architecture

IronLayer is built as a dual-engine system where determinism and AI advisory are cleanly separated.

Dual-Engine Design

Layer A (Core Engine) is purely deterministic. Given the same git diff and current state, it produces byte-for-byte identical plan JSON. This is enforced by:

• Content-based plan IDs (SHA-256 of the canonical plan JSON)
• Sorted keys in all JSON output
• No timestamps in plan content
• Versioned canonicalization (v1) for hash stability

Layer B (AI Engine) provides advisory-only intelligence. AI analysis is attached as metadata annotations — it never modifies the execution plan. If the AI engine is unavailable, the plan executes normally without advisory data.

Multi-Tenant Isolation

Every table in the PostgreSQL database includes a tenant_id column. Row-level security (RLS) policies are enforced at the database level, not just the application layer. Tenant context is set per-session.

Data Flow

1. Git diff — IronLayer detects changed SQL files between two commits
2. Model discovery — SQL files are parsed with SQLGlot (Databricks dialect) to extract table references
3. DAG construction — NetworkX builds a dependency graph from the discovered models
4. Plan generation — The planner determines execution order, incremental strategy, and watermark ranges
5. AI augmentation (optional) — Cost prediction, risk scoring, and SQL optimization suggestions are attached
6. Approval gate — Plans require explicit approval (or auto-approve in dev mode)
7. Execution — SQL is executed on the target Databricks workspace via the Databricks SQL API
8. Telemetry — Runtime metrics are captured (PII-scrubbed) and stored for cost prediction training

Security Model

• JWT authentication with short-lived tokens (1 hour) and refresh tokens (24 hours)
• RBAC with admin, engineer, and viewer roles
• Credential encryption using Fernet (AES-128-CBC + HMAC-SHA256, PBKDF2 key derivation)
• SQL safety guard blocks dangerous statements (DROP, TRUNCATE, DELETE, ALTER)
• Hash-chained audit log for tamper-evident compliance
• Rate limiting per-user and per-endpoint
• Token revocation via JTI blacklist

Tech Stack