SQL Server 2025 RC0: AI-Ready, Secure-by-Default On-Prem Database

Microsoft’s first Release Candidate (RC0) for SQL Server 2025 is here, and it’s more than a stability checkpoint—it’s a statement of direction that blends built-in AI, developer‑friendly T‑SQL, and secure‑by‑default networking into a single, on‑premises database platform that looks and feels cloud‑ready. Microsoft says preview adoption for SQL Server 2025 is running at roughly twice the pace of SQL Server 2022, a sign that features like native vector search, JSON innovations, and regex support are resonating with teams building next‑gen applications. RC0 also formalizes a new opt‑in “preview features” switch that lets you test cutting‑edge capabilities without destabilizing production workloads, a welcome change for shops that demand guardrails while experimenting with AI and analytics. (microsoft.com, techcommunity.microsoft.com)

---

Background​

SQL Server 2025 entered public preview in mid‑May, with incremental community technology previews (CTPs) surfacing through major events and user groups. Today’s RC0, dated August 22, 2025, signals the sprint to general availability: focus shifts from net‑new features to ship‑blocking bugs, ergonomics, and operational polish. Microsoft is using RC0 to reinforce its core themes—AI where your data lives, security that defaults to strong encryption, and a balanced model for trying features early without compromising the database’s supportability. (techcommunity.microsoft.com, microsoft.com)

• RC0 introduces the PREVIEW_FEATURES database‑scoped configuration, a mechanism to toggle specific innovations on a per‑database basis—distinct from instance‑level trace flags that traditionally controlled internal behaviors. (microsoft.com)
• The company highlights customer interest in “built‑in AI,” native JSON, and T‑SQL regex, underscoring how 2025’s headline items straddle both modern application patterns and bread‑and‑butter developer productivity. (microsoft.com)

---

What’s actually new in RC0​

Microsoft’s RC0 announcement and companion engineering posts outline a raft of improvements. The following are the most consequential for Windows and cross‑platform SQL Server environments.

A safer path to innovation: PREVIEW_FEATURES​

RC0 debuts a database‑scoped configuration—PREVIEW_FEATURES—that you can enable with a simple ALTER DATABASE SCOPED CONFIGURATION statement. It gates a rotating set of previewed capabilities such as vector indexing and AI text chunking. Crucially, enabling the switch doesn’t put your database into a “non‑GA” state, though you should treat preview features as non‑production and plan cleanup before disabling the switch if you’ve created persisted preview objects. (learn.microsoft.com, devblogs.microsoft.com)

• Example: ALTER DATABASE SCOPED CONFIGURATION SET PREVIEW_FEATURES = ON; enables eligible preview experiences in the target database. (devblogs.microsoft.com)
• The model is intentionally the opposite of broad trace flags: it’s scoped to a database, developer‑facing, and designed to isolate experimental functionality. (microsoft.com)

Built‑in AI: vectors, models, and text chunking​

SQL Server 2025 embeds AI primitives right into T‑SQL: a native vector data type, vector distance/search functions, and an approximate vector index (DiskANN) for fast similarity search. New AI functions like AI_GENERATE_EMBEDDINGS and AI_GENERATE_CHUNKS allow you to generate embeddings and chunk text in‑database. RC0 further expands model options by allowing local ONNX models and clarifies how to invoke external model endpoints via REST. (techcommunity.microsoft.com, devblogs.microsoft.com)

• You can register external models (Azure OpenAI, OpenAI, Ollama, or local ONNX) with CREATE EXTERNAL MODEL and call them from T‑SQL. Many shops will enable REST calls in a controlled database and then define models for embeddings or text tasks. (devblogs.microsoft.com)
• DiskANN vector indexing and VECTOR_SEARCH bring approximate nearest‑neighbor retrieval into SQL Server, cutting query latencies for RAG and recommendation workloads at scale. (techcommunity.microsoft.com)
• Microsoft positions these capabilities as “ground‑to‑cloud,” consistent across on‑premises SQL Server and Azure SQL, which eases hybrid deployment strategies. (techcommunity.microsoft.com)

Developer‑friendly SQL: native JSON and regex​

Two long‑requested areas see major movement:

• Native JSON data type and JSON index: The engine now stores JSON in a binary format and exposes a JSON index that accelerates JSON_VALUE/JSON_PATH_EXISTS/JSON_CONTAINS, reducing the friction of mixing relational and semi‑structured data. It’s GA in Azure SQL and available in the SQL Server 2025 preview, with an explicit JSON index syntax. (devblogs.microsoft.com, techcommunity.microsoft.com)
• T‑SQL regex suite: Functions like REGEXP_LIKE, REGEXP_REPLACE, and REGEXP_SUBSTR make complex pattern matching first‑class. In RC0, REGEXP_LIKE becomes SARGable and regex functions support LOB types such as varchar(max)/nvarchar(max). Note that REGEXP_LIKE requires compatibility level 170, while others don’t. (techcommunity.microsoft.com, stackoverflow.com)

These changes unlock cleaner data validation, powerful parsing, and richer search semantics without round‑tripping to application code or CLR.

Secure‑by‑default connections and modern crypto​

SQL Server 2025 pushes encryption decisions left. RC0 introduces and enables TLS 1.3 by default on Linux, and the broader 2025 cycle bakes in TDS 8.0 so encryption starts at pre‑login. SQL Agent and Linked Servers now default to mandatory encryption, with support rolling across high availability and replication features. Expect fewer silent downgrades and more explicit failures when clients are outdated or misconfigured. (techcommunity.microsoft.com)

• TLS 1.3 on by default in RC0 (Linux) and TDS 8.0 enforcement tighten the handshake from TCP onward. DBAs should audit client driver versions ahead of upgrades. (techcommunity.microsoft.com)

Linux momentum and Ubuntu 24.04 (preview)​

Adoption on Linux continues to rise, and RC0 adds preview support for Ubuntu 24.04 LTS. Microsoft emphasizes this is for evaluation, not production, giving teams a contemporary dev/test baseline—particularly in WSL2 or containerized scenarios—while production guidance remains on supported 2022 images. (techcommunity.microsoft.com)

Azure integration: managed identity and Arc​

Two threads converge: PolyBase connectivity to Azure Storage with managed identity, and broader Entra integration via Azure Arc. The upshot is less secret sprawl and simpler governance for outbound Azure calls (e.g., staging data to Data Lake) and selected inbound scenarios. For many on‑prem shops, Arc is the key that unlocks managed identity, making “passwordless” patterns viable without moving the database to the cloud. (techcommunity.microsoft.com)

Fabric mirroring with real resource governance​

Mirroring to Microsoft Fabric now benefits from Resource Governor alignment, letting you carve out workload groups and, if desired, their own pool to prevent ingestion and synchronization flows from starving OLTP. It’s a pragmatic bridge between operational stores and near‑real‑time analytics without standing up entirely separate data movement stacks. (microsoft.com)

---

Under‑the‑hood engine improvements​

Beyond the marquee features, RC0 and the 2025 train bring quieter but important operational wins:

• CE Feedback for expressions now persists across restarts/failovers, stabilizing plans in workloads that rely on complex calculated predicates. (techcommunity.microsoft.com)
• Time‑bound Extended Events sessions automatically stop after the specified horizon, a small but handy quality‑of‑life improvement for diagnostics hygiene. (techcommunity.microsoft.com)
• Tempdb space resource governance lets you cap tempdb consumption per workload group, dramatically reducing “bully query” blast radius. Paired with ADR in tempdb, it’s a meaningful step toward resilient tempdb operations. (techcommunity.microsoft.com)
• Type tweaks and function upgrades—like JAROWINKLER return type adjustments and REGEXP_ enhancements—smooth rough edges that could become papercuts at scale. (techcommunity.microsoft.com)

---

Why this Release Candidate matters for Windows shops​

If you run SQL Server on Windows, RC0 clarifies how 2025 serves both modernization and day‑to‑day reliability:

• Built‑in AI without bolt‑on stores: By keeping embeddings, indexes, and inference orchestration inside the database boundary, you can secure, audit, back up, and govern AI artifacts with the same tooling you use today. That’s a direct productivity win for DBAs who would otherwise be asked to own a separate vector database. (techcommunity.microsoft.com)
• Safer experimentation: The PREVIEW_FEATURES switch removes the “all or nothing” feel of early adoption, granting per‑database control and an exit strategy. (learn.microsoft.com)
• End‑to‑end encryption by default: TDS 8.0/TLS 1.3 reduce the chance that an old client quietly drops you into weak or no encryption. Your security team will appreciate the new defaults; your operations team gets cleaner failure modes during cutovers. (techcommunity.microsoft.com)
• Windows‑first AI flows: RC0’s Windows walkthroughs show how to enable external REST endpoints, register local ONNX or Ollama models, and instrument calls with Extended Events—without standing up a separate inference service tier. (devblogs.microsoft.com)

---

Hands‑on: a quick evaluation plan for RC0​

The following sequence is designed for a Windows‑hosted dev or staging environment. It emphasizes isolation, auditability, and reversible changes.

• Stand up an RC0 instance in a sandbox. Ensure client tools (SSMS/SqlClient) support TLS 1.3 and TDS 8.0 defaults; plan to test fallbacks for older apps. (techcommunity.microsoft.com)
• Create a dedicated database for AI and preview experimentation. Enable PREVIEW_FEATURES in that database only: ALTER DATABASE SCOPED CONFIGURATION SET PREVIEW_FEATURES = ON; (devblogs.microsoft.com)
• Set database compatibility to 170 if you’ll validate REGEXP_LIKE: ALTER DATABASE YourDB SET COMPATIBILITY_LEVEL = 170; (techcommunity.microsoft.com)
• If testing local embeddings or REST‑based model endpoints, enable REST invocation with sp_configure 'external rest endpoint enabled', 1; RECONFIGURE WITH OVERRIDE; then define an external model (e.g., to a local ONNX or Ollama endpoint). Add an XEvent session to monitor REST and embedding calls. (devblogs.microsoft.com)
• Validate vector operations end‑to‑end: generate embeddings with AI_GENERATE_EMBEDDINGS, build a DiskANN CREATE VECTOR INDEX, and exercise VECTOR_SEARCH alongside standard SQL filters to confirm hybrid search behavior and latency. (techcommunity.microsoft.com)
• Pilot JSON upgrades on a noncritical table: migrate a string column to the native JSON type, create a JSON index, and benchmark typical JSON_VALUE and JSON_PATH_EXISTS predicates. (techcommunity.microsoft.com)
• Trial regex queries against troublesome data quality patterns. Confirm SARGability of REGEXP_LIKE in your schema by reviewing actual execution plans and index usage. (techcommunity.microsoft.com)
• If you use Fabric mirroring, carve out a Resource Governor pool and assign ingestion/sync workload groups to test isolation. Watch CPU, IO, and tempdb under load. (microsoft.com)
• For Azure‑connected scenarios, Arc‑enable the host to test managed identity for outbound storage operations via PolyBase, and review identity governance with your security team. (techcommunity.microsoft.com)

---

Strengths that stand out​

• “AI where your data lives” is real now. SQL Server 2025’s vector and model primitives shrink architecture sprawl and unlock retrieval‑augmented use cases without exporting data to niche databases. Hybrid search—combining vector and relational filters—becomes a first‑class pattern. (techcommunity.microsoft.com)
• Preview without peril. By scoping experimental features at the database level, Microsoft reduces the blast radius of trying new things and aligns on‑prem processes with cloud‑style feature flags. (microsoft.com)
• Security defaults grow up. Enforcing encryption from pre‑login with TDS 8.0/TLS 1.3 and requiring encryption for Agent and Linked Servers tighten posture while illuminating weak links during upgrades. (techcommunity.microsoft.com)
• Developer delight. Native JSON and regex bring long‑awaited ergonomics, and early SARGability work for REGEXP_LIKE shows the engine team is sweating real‑world performance details. (techcommunity.microsoft.com)
• Operational plumbing. Tempdb resource caps and CE Feedback persistence translate directly into fewer “mystery outages” and more predictable plan behavior after maintenance windows and failovers. (techcommunity.microsoft.com)

---

Risks, caveats, and what to watch​

Even with RC0 maturity, some sharp edges remain. A prudent rollout acknowledges them up front.

• Preview is still preview. The PREVIEW_FEATURES model is safer, not risk‑free. If you create persisted objects (e.g., a preview vector index), you may need to drop or rebuild them before turning preview features off. Treat these features as non‑production until GA guidance changes. (learn.microsoft.com)
• JSON is powerful—and young on‑prem. While the native JSON type is mature in Azure SQL, the SQL Server 2025 preview includes limits and unsupported areas called out by the community (for example, certain replication/CDC scenarios or bulk import behaviors). Validate carefully before large‑scale schema changes. (devblogs.microsoft.com, brentozar.com)
• Regex can be a foot‑gun. SARGability helps, but poorly designed patterns can still thrash CPU and undermine indexing. Ensure compatibility level 170 where required and profile plans—don’t assume LIKE‑style performance. (techcommunity.microsoft.com, stackoverflow.com)
• Client compatibility for TLS 1.3/TDS 8.0. Older drivers and middleware may balk at stricter defaults. Inventory connection strings and provider versions, and plan phased cutovers with capture/replay where possible. (techcommunity.microsoft.com)
• Managed identity nuances on‑prem. Historically, on‑prem SQL Server couldn’t use managed identity for Azure resources without mediation. In 2025, Azure Arc is the enabling layer—great news, but it introduces its own operational considerations in governance and lifecycle. (techcommunity.microsoft.com)
• Fabric mirroring resource planning. Dedicated pools and workload groups add control and complexity. Under‑provisioning will pinch ingestion; over‑provisioning can starve OLTP. Test with production‑like telemetry. (microsoft.com)

---

How RC0 changes day‑one architecture conversations​

SQL Server 2025 RC0 subtly shifts the default answers architects give to common questions:

• “Do we need a separate vector database?” Often, no. If your retrieval and ranking can live within SQL’s vector type, DiskANN index, and filtering, keeping everything in one engine reduces operational overhead and data governance risk. (techcommunity.microsoft.com)
• “How do we safely test AI features?” Use PREVIEW_FEATURES in an isolated database, register a local ONNX or Ollama model, and gate outbound calls with firewall rules and auditing. You can simulate realistic workflows without shipping data to third‑party services. (devblogs.microsoft.com)
• “Can we get to near real‑time analytics without a new pipeline?” Fabric mirroring plus Resource Governor will cover many cases, giving product teams fresher insights while DBAs retain control over resource isolation. (microsoft.com)

---

Practical testing scenarios you can run this week​

• Build a semantic product search: store product descriptions and attributes; generate embeddings via a local ONNX model; add a DiskANN index; expose a parameterized vector + keyword search from your API. Track p95 latency and recall compared to baseline keyword search. (devblogs.microsoft.com, techcommunity.microsoft.com)
• Cleanse inbound feeds with regex: replace brittle application‑side parsing with REGEXP_REPLACE and REGEXP_LIKE in staging tables; confirm SARGable predicates hit supporting indexes. (techcommunity.microsoft.com)
• Migrate a hot JSON workload: convert a nvarchar(max) column to JSON type; apply a JSON index to targeted paths; quantify plan stability and CPU improvements relative to computed‑column workarounds. (techcommunity.microsoft.com)
• Validate secure‑by‑default posture: run connection tests from your app stack with TLS 1.3/TDS 8.0 enforced; document any clients that fail handshakes; plan upgrades or shims. (techcommunity.microsoft.com)
• Pilot Arc‑enabled managed identity: mirror a data mart to Fabric and test outbound storage and key vault calls authenticated via managed identity—no embedded secrets. (techcommunity.microsoft.com)

---

Verdict: RC0 is the right kind of bold​

SQL Server 2025 RC0 threads a tricky needle. It’s assertive about AI and developer experience, yet conservative where it counts—encryption defaults, scoped experimentation, and guardrails around preview features. The inclusion of a vector database’s best tricks—ANN indexes, embeddings, hybrid search—alongside JSON and regex closes gaps that had pushed teams toward extra components and glue code. And by hardening connectivity and offering managed identity via Arc, the platform acknowledges modern security expectations instead of punting them to documentation. (techcommunity.microsoft.com)
There’s still work ahead: production‑grade guidance for JSON on‑prem, careful regex performance education, rigorous client compatibility testing for TLS/TDS changes, and clear upgrade playbooks for Fabric mirroring at scale. But with RC0, the shape of SQL Server 2025 is unmistakable—a Windows‑friendly, AI‑ready, secure‑by‑default database that lets you innovate at your pace while keeping enterprise governance intact. If you’ve been waiting for a reason to kick off serious lab testing, this is the release candidate to install, script, and start benchmarking today. (microsoft.com, techcommunity.microsoft.com)

---

Source: Microsoft Accelerating SQL Server 2025 momentum: Announcing the first release candidate - Microsoft SQL Server Blog