Databricks Cost Intelligence
· Architecture
OBSERVE → MODEL → CONSTRAIN → SIMULATE → EXPLAIN → APPROVE
READ-ONLY FIRST
CONCEPT v0.1
ARCHITECTURE
COST FLOW
AI REASONING
GOVERNANCE
Architecture view active
Databricks Cost Intelligence architecture
A read-only architecture that observes a running Databricks estate, builds workload digital twins, simulates constrained alternatives, uses governed AI for explanation, and requires human approval before validation or action.
OBSERVED DATABRICKS ESTATE
OPTIMIZATION INTELLIGENCE PLATFORM
DECISION & ACTION
EVIDENCE-BASED · CONSTRAINED · EXPLAINABLE
READ-ONLY COLLECTION · NO WORKLOAD REDESIGN
Current Compute Fleet
All-purpose · Jobs · Pipelines
CONFIG
01
Runtime Telemetry
CPU · memory · shuffle · spill
OBSERVED
02
Cost & Job History
DBU · active hours · SLA history
COST
03
Live Capability Catalog
Node types · DBR · policies
CURRENT
04
Cloud & Contract Rates
Infrastructure · discounts · region
RATES
05
Telemetry & Catalog Sync
Read-only · scheduled · versioned
COLLECT
System Table Aggregator
Selective 30–90 day evidence windows
Spark Event Parser
Stages · tasks · spill · skew · backlog
Capability Snapshot
Nodes · runtimes · policies · lifecycle
Pricing Adapters
DBU + cloud + customer effective rate
Release & Docs Sync
DBR notes · EOL · compatibility RAG
Schema Normalization
Cloud-neutral node & config model
NO RAW CUSTOMER DATA SENT TO AI
Workload Digital Twin
Observed behavior · not a questionnaire
PROFILE
Optimization Lakehouse
Telemetry snapshots · config history
Catalog versions · outcomes · audit
Workload Signature
Interactive · batch · streaming
CPU · memory · I/O · shuffle · driver
Operational Envelope
Concurrency · failure · retries · SLA
Idle windows · peak vs. average
Current Cost Baseline
DBU + infrastructure + active hours
Per cluster · job · owner · workspace
VERSIONED EVIDENCE WINDOW
Candidate Engine
Hard rules before AI
Workspace / region availability
DBR · Photon · ARM compatibility
Policies · access mode · libraries
SLA · spot · GPU · local disk
Full config candidates, not SKUs
VALID CANDIDATES ONLY
Cost & Risk Simulator
Cost · performance · risk
$10K
ILLUSTRATIVE MONTHLY SAVE
Runtime & worker profile
Failure / OOM / SLA probability
Savings range + confidence
Configurable AI Reasoning
Structured evidence in · explanation and migration plan out
AI GATEWAY
FAST
Extract · classify
Release-note sync
HAIKU CLASS
STANDARD
Analyze · explain
Config diff · Q&A
SONNET CLASS
DEEP
Complex migration
High-value · low confidence
OPUS CLASS
Fresh DBR RAG · champion/challenger · token budgets · prompt versioning
LLM never invents candidates or calculates the authoritative price
Recommendation
Evidence-backed savings decision
CURRENT
All-Purpose
DBR 14.3 · fixed 8
PROPOSED
Jobs Compute
LTS · autoscale 2–6
$9–11K
ILLUSTRATIVE SAVINGS RANGE
Performance impact
Confidence & evidence strength
Risk & migration warnings
Databricks App
Review · approve · validate · export
HUMAN
Approval required before change
Canary / shadow validation
Export JSON · policy · Terraform
OPTIONAL APPLY · REVALIDATE · ROLLBACK
Safety Envelope
MVP DEFAULT
Read-only workspace access
No automatic cluster mutation
Savings shown as a range
High-risk changes require canary
START: ALL-PURPOSE + JOBS · ONE CLOUD
Trust, Governance & Continuous Learning
Controls every recommendation from evidence window to realized outcome
CONTROL PLANE
OAuth · RBAC · Least Privilege
Catalog & Price Snapshot Lineage
Model · Prompt · Rule Versions
AI Usage · Tokens · Budgets
Approval & Immutable Audit Trail
Evaluation Dataset & Regression
Champion / Challenger Rollout
Validation & Rollback Evidence
Realized Savings & Feedback
LEGEND & OPERATING PRINCIPLE
Evidence / decision flow
Control / catalog flow
Outcome / learning feedback
THE LLM EXPLAINS — IT DOES NOT AUTHORIZE
YASAR KOCYIGIT / DAQ
ARCHITECTURE SIGNATURE · CONCEPT v0.1
VALIDATION OUTCOME · REALIZED SAVINGS · CALIBRATION