#What's Current (as of September 2026)

Read this before any interview. Quoting a renamed product is a small factual error that reads as a large signal: this person stopped learning. For a role whose value is being current across engagements, that is expensive.

Everything below is sourced. Verify status on the day — preview/GA moves fast.

#1. Names that changed (fix these first)

If you still say…	Say this instead
Delta Live Tables (DLT)	Lakeflow Declarative Pipelines (Spark Declarative Pipelines)
Workflows	Lakeflow Jobs
Ingestion connectors / Arcion	Lakeflow Connect
AI Gateway (for model traffic)	Unity AI Gateway — now broader: agents, tools, MCP, spend
Online tables	Lakebase (serverless Postgres on lake storage)
"Databricks SQL warehouse" for everything	Still right — but Lakehouse//RT now exists for real-time
Model Registry (workspace)	Models in Unity Catalog

#2. Data + AI Summit 2026 — what actually shipped

Summit ran 15–18 June 2026 at Moscone, 30,000+ attendees, 20+ launches. The organising theme: enterprise AI competition has moved from "who builds the best agent" to "who governs the data, context, cost and security underneath it." That sentence is worth borrowing.

#Real-time: Lakehouse//RT and the Reyden engine

A new real-time SQL warehouse type powered by Reyden, a new engine ("Reynold's Dream Engine") using an asynchronous execution model. Reported: ~10 ms on smaller datasets, sub-100 ms on larger, sub-100 ms at ~12,000 queries/second, up to 16× better than dedicated real-time serving layers, with latency staying flat as concurrency rises. It queries Delta and Iceberg tables in Unity Catalog directly — no separate serving store, no copy. Beta, read-only workloads at launch. (Databricks blog, newsroom, docs)

Why it matters architecturally: it attacks the "separate serving layer" tax — the Pinot/Druid/ ClickHouse/StarTree pattern where you duplicate data into a real-time store, fragment governance, and maintain parallel pipelines. Strong material for a design scenario involving operational dashboards or customer-facing analytics.

#OLTP: Lakebase and LTAP

Lakebase — serverless Postgres decoupled from storage, on lake storage, governed by Unity Catalog: sub-500 ms instance creation, scale-to-zero, sub-500 ms branching (copy-on-write), and new cross-cloud/cross-region disaster recovery. Lakebase Search (beta) adds hybrid vector + full-text retrieval in Postgres.
LTAP — Lake Transactional/Analytical Processing (announced June 2026): operational data written directly into open columnar formats (Delta/Iceberg) registered in Unity Catalog, so transactional data is immediately queryable for analytics with no CDC pipeline and no second copy. (Databricks newsroom, docs)

Why it matters: if LTAP delivers, a whole category of architecture — the CDC pipeline from the operational database into the lake — becomes unnecessary for new builds. Treat it as directional, not as something to promise a customer today, and say exactly that in the room.

#Open formats: Iceberg v3 GA

Iceberg v3 is GA, with managed Iceberg tables in the runtime. v3 brings deletion vectors, row lineage and the VARIANT type, compatible across Delta, Iceberg, Parquet and Spark — allowing Delta and Iceberg tables to share physical Parquet files without rewriting data. (Summit recap) The practical consequence: "Delta vs Iceberg" is now a catalog question, not a format question.

#AI governance: Unity AI Gateway

Runtime governance for models, agents, MCP services, skills and tools — the layer beneath the agent stack:

Cost: unified spend visibility, attribution by user/team/tool/use case, hard spend caps that halt requests, and smart routing to cheaper models by task complexity.
Contextual service policies (beta): allow / deny / require approval on actions — pushing to a repo, writing to sensitive storage — enforced by context (user, agent, model, content).
Guardrails: PII, prompt injection, jailbreak, unsafe content.
MCP registry: managed MCP services (Drive, Jira, Confluence, Slack, GitHub, SharePoint) plus custom registrations, with access management and audit.
Tracing: unified agent tracing through MLflow; traces analysable in Genie and Lakewatch. (Databricks blog, product page)

The crisp distinction: Unity Catalog governs assets; Unity AI Gateway governs interactions.

#Agents: Agent Bricks and Omnigent

Agent Bricks expanded: managed memory, sandboxed execution, document intelligence, more model support; 100,000+ agents reportedly built.
Omnigent — an open-source meta-harness sitting above harnesses (Claude Code, Codex, custom), with three pillars: composition, control, collaboration. A managed variant on Databricks inherits Unity AI Gateway governance and isolated execution. (Databricks docs)

#Semantic layer: Metrics, Domains, Business Glossary

Unity Catalog Metrics (public preview) — KPIs like revenue and churn as governed, reusable objects queryable from SQL, BI, APIs and agents; multi-fact relationships, parameterisation, materialisation.
Domains (public preview) — business-aligned grouping of assets, which also scopes agent context retrieval.
Business Glossary (preview) — authoritative business definitions linked to assets. (Atlan recap)

Why it matters: this is the answer to "every dashboard computes revenue differently", and it is the substrate agents need to answer business questions consistently.

#Genie

Genie One (GA) — an agentic interface across structured and unstructured data and 50+ enterprise applications, producing documents and artefacts, on web and mobile. Genie Ontology — a continuously-built enterprise knowledge graph grounding agent answers (OntoRank ranking; millisecond retrieval). Genie Agents (GA) — turn a conversation into a reusable agent deployable to Slack/Teams. Plus Genie Code, Genie ZeroOps (autonomous pipeline/model incident triage), and Genie App Builder.

#Data engineering

Lakeflow Connect past 100+ connectors; ZeroBus (push API to Delta, sub-5s latency, high throughput, Kafka-compatible API in beta); Spark Declarative Pipelines with real-time mode (~5 ms end-to-end); Lakeflow Designer (no-code → open Spark declarative pipelines); Lakeflow Jobs with 50+ integrations (1.7 billion job runs/month); Lakebridge migration tooling.

OpenSharing — vendor-neutral sharing protocol under the Linux Foundation, extending Delta Sharing to AI assets, Iceberg REST clients and on-prem sources (28,000+ recipients, ~33% of shares cross-platform). Lakewatch — agentic SIEM on open lakehouse format, claimed ~80% lower TCO than legacy SIEMs; Panther acquisition intent. CustomerLake — agentic CDP. AI Runtime — serverless GPU compute (public preview). Databricks Free Edition expanded to include Genie Code, serverless GPUs, Lakebase, Agent Bricks and Lakeflow Designer.

#3. How to use this in the room

Do: reference one or two of these in service of an architecture point. "For that operational dashboard requirement I'd look at Lakehouse//RT rather than standing up a separate serving store — it's beta and read-only today, so I'd validate it against your latency profile before committing."

Don't: recite the launch list. Feature recitation is the most-flagged senior failure mode, and a summit-recap monologue is the purest form of it.

The meta-point worth making once:

"The pattern across the last year is Databricks collapsing adjacent systems into the governed lake — the OLTP database, the real-time serving store, the SIEM, the semantic layer, the agent control plane. Whether each individual one wins is an open question, but the strategy is coherent: if the data and the governance are already here, every additional copy elsewhere is cost, risk and drift."

That shows you think about the vendor's strategy, not just its features — which is exactly what a chief architect is paid for.

#4. Caveats to carry

Beta is not production. Lakehouse//RT is beta and read-only; LTAP is announced/coming; several Genie capabilities are preview. Recommending a preview product for a customer's critical path without flagging it is a credibility loss that can lose a room.
Vendor benchmarks are vendor benchmarks. Quote the 16× and the 12,000 QPS as claimed, and say you would validate on the customer's workload. That caveat raises your score; it is truth-seeking in action.
Verify before you assert. Check current documentation the morning of the interview for anything you plan to state as fact.