Platform Updates & Releases
Delivering the Most Enterprise-Ready Postgres, Built for Snowflake
Instances are created via CREATE POSTGRES INSTANCE SQL commands or Snowsight and run on Snowflake-managed VMs across AWS and Azure in all major regions. The open-source pg_lake extension lets Postgres query and write Apache Iceberg tables directly — zero-copy access to your Snowflake data lake from any Postgres-compatible ORM or driver. Full ACID compliance, row-level locking, and Postgres wire protocol compatibility are supported, meaning existing applications connect without code changes.
Project SnowWork: The Easiest Way for Business Users to Get Work Done
SnowWork ships with persona-specific AI "profiles" for finance, sales, marketing, and operations — each pre-loaded with curated agent skills that understand function-specific KPIs, workflows, and terminology. It orchestrates across Snowflake data, Cortex AI models, and enterprise systems to produce finished outputs: board-ready slide decks, churn-risk spreadsheets, supply chain reports. Unlike generic AI agents, SnowWork is built on a single unified metric layer with governed definitions, cross-cloud interoperability, and full auditability. Currently available to a limited research preview cohort.
Snowflake Announces Intent to Acquire Observe for AI-Powered Observability
The acquisition brings Observe's AI-powered log correlation, distributed tracing, and operational intelligence into Snowflake's environment. Observe processes and correlates telemetry data (logs, traces, metrics) at enterprise scale and enables root-cause analysis and anomaly detection. Integration with Snowflake's data platform means observability signals can be correlated directly with business data — a powerful capability for platform engineering teams running Snowflake-native applications and for data engineering teams building on Snowpipe Streaming.
Iceberg Write Support for Databricks Unity Catalog on Azure — Now GA
The GA release enables full DML operations (MERGE, INSERT, UPDATE, DELETE) from Snowflake against Iceberg tables governed by Databricks Unity Catalog on Azure Data Lake Storage. This was previously AWS-only; Azure parity is now complete. The integration uses Iceberg's open REST Catalog API for catalog registration, and write operations produce standard Iceberg v2-compliant files visible to any compatible engine. This positions Snowflake as a viable write-side compute engine in Databricks-governed lakehouse architectures.
Snowpipe Streaming Now Supports Error Logging (GA)
Error logging is enabled via the ON_ERROR parameter in the Snowpipe Streaming channel configuration. Failed rows are written to a designated error log table with the original record payload, error type, timestamp, and channel metadata. Teams can query this table directly in Snowflake to build alerting, dead-letter queue workflows, or automated retry logic using Snowflake Tasks. This eliminates the need to instrument external observability tooling just to surface streaming ingestion errors — everything stays in the warehouse.
Medical & Health Data Classifiers Now Available in Data Classification (GA)
The new classifiers detect healthcare-specific sensitive data types including ICD codes, NPI numbers, DEA numbers, medical record identifiers, and health insurance fields. Classification runs via SYSTEM$CLASSIFY_TABLE() or through scheduled classification policies, automatically tagging columns with semantic labels. Those tags can then drive dynamic masking policies and row-level access policies — enabling end-to-end automated PHI governance natively in Snowflake with no external tools required.
Cortex AI & ML
Announcing OpenAI GPT-5.2 on Snowflake Cortex AI — $200M Partnership
GPT-5.2 is accessible via SNOWFLAKE.CORTEX.COMPLETE('openai-gpt-5-2', prompt) in SQL, through the Cortex REST API, and is coming to Snowflake Intelligence. The model supports a 200K context window, multimodal inputs (text, images, audio via Cortex), and structured JSON output via the response_format parameter. It outperforms GPT-4 significantly on complex reasoning, long-context understanding, and agentic tool-calling — benchmarked at near-perfect accuracy on multi-step task coordination. Pricing follows the Cortex credit consumption model.
Cortex Code CLI Expands to Support Any Data, Anywhere
Cortex Code CLI now connects to external data sources, allowing the agent to understand schemas and relationships across multiple systems when generating SQL, Python, or Snowpark code. The tool uses Snowflake's semantic layer and metadata store to provide contextually accurate suggestions. It is embedded in Snowsight, available via CLI, and integrates with developer IDEs. Teams report 40-60% reduction in time writing boilerplate data transformation code; over 4,400 customers active as of early 2026.
Serverless LLM Fine-Tuning Using Snowflake Cortex AI
Fine-tuning is initiated via the SNOWFLAKE.CORTEX.FINETUNE() SQL function or Python API. Supported base models include Mistral 7B and Llama 3 variants. Training data must be formatted as instruction-response pairs in a Snowflake table. The function returns a model identifier usable as a drop-in replacement for any CORTEX.COMPLETE() call. Training jobs run asynchronously and are monitored via STATUS checks. Credits are consumed only during training compute time — no always-on GPU cluster required.
Gen AI in Action: Real-World Cortex AI Customer Outcomes
Siemens Energy used CORTEX.COMPLETE() and EXTRACT_ANSWER() combined with Streamlit in Snowflake to build an internal knowledge assistant over 500K+ pages of engineering documentation — deployed entirely within Snowflake's security perimeter with no external model APIs. Alberta Health Services used Cortex for clinical note automation, reducing physician documentation burden. Terakeet reported 98% faster market opportunity identification using Cortex over structured SEO data. All deployments run with zero data egress from Snowflake.
Architecture & Engineering
Announcing Apache Iceberg V3 Support on Snowflake (Public Preview)
Iceberg V3 introduces: (1) the Variant data type for schema-less semi-structured data (JSON/XML); (2) geospatial types (GEOMETRY, GEOGRAPHY) at the file format level; (3) deletion vectors that mark deleted rows in a separate bitfield file rather than rewriting data files — dramatically improving write performance for UPDATE/DELETE-heavy workloads; (4) nanosecond-precision timestamps; (5) row-level lineage tracking via persistent row IDs. Support covers both Snowflake-managed Iceberg tables and external catalog integrations (Glue, Unity Catalog, Polaris). Use with caution in production during preview.
Extending Snowflake Data Sharing to Open Table Formats
Data providers can now create Snowflake Marketplace listings or private shares exposing Iceberg and Delta Lake tables directly. Consumers receive read access to native format files in cloud storage — no data copy, no ETL, no conversion. Sharing works across regions and clouds. The integration uses Snowflake's Polaris open catalog for metadata registration, and consumers can query shared tables through Snowflake, Spark, Trino, or any Iceberg-compatible engine. Over 820 providers and 3,400+ datasets are now on Snowflake Marketplace.
Next-Gen Data Engineering: 6 Snowflake Features Transforming How You Build
The six features: (1) Dynamic Tables — define a SQL query, Snowflake handles incremental refresh and scheduling; (2) Tasks with DAG support — multi-step workflows without Airflow or Dagster; (3) Native dbt execution on Snowflake compute without container infrastructure; (4) Streams + Tasks for CDC — row-level change detection with triggered downstream processing; (5) Snowpipe for continuous ingestion; (6) Notebooks for exploratory data engineering. Together these replace significant external tooling for most mid-size data engineering teams running 10-50 pipelines.
Introducing the Snowflake Well-Architected Framework
The five pillars: (1) Operational Excellence — monitoring, alerting, change management; (2) Security & Governance — RBAC, data classification, dynamic masking, row policies; (3) Reliability — failover, data recovery, task retry logic; (4) Performance Efficiency — warehouse sizing, clustering, query optimization, caching; (5) Cost Optimization — resource monitors, auto-suspend tuning, materialized views. Unlike cloud-provider frameworks focused on infra, Snowflake WAF addresses platform-native controls and SQL-level best practices. Full training curriculum available at learn.snowflake.com.
Tutorials & How-Tos
The Long Game, Part 3: The Architect's Anchor
Goodrich builds on the Snowflake Well-Architected Framework to propose the concept of an "Architect's Anchor" — automated checks, naming conventions, and shared code patterns enforced at the platform layer. Practical coverage includes: governance-as-code using Snowflake Tasks and Data Metric Functions for automated WAF compliance checks; naming convention enforcement via object tagging policies; and a shared SQL library pattern for common platform operations. Includes working SQL examples for automated WAF pillar monitoring.
Snowflake Cost Optimization: 12 Proven Techniques to Cut Your Bill by 40%
Key techniques: (1) Auto-suspend <=60 seconds on non-interactive warehouses — 15-25% cost reduction within 24 hours; (2) Warehouse right-sizing by testing smaller tiers; (3) Warehouse consolidation by performance SLA rather than by team/domain; (4) Query optimization before infrastructure tuning — partition pruning analysis via QUERY_PROFILE; (5) Clustering key strategy — 70-90% scan reduction on range-filtered large tables; (6) Resource monitors with hard credit limits; (7) Materialized views for frequently reused sub-queries; (8) Result cache exploitation. VWH compute accounts for 60-80% of total Snowflake spend.
Data Engineering Pipelines with Snowpark Python
Covers building a multi-stage pipeline using Snowpark DataFrames for incremental data processing. Key patterns: Snowpark stored procedures for transformations (deployed via session.sproc.register()), Python Tasks for orchestration, schema evolution handling with DataFrame.merge(), and type-safe schema definitions using StructType. The pipeline runs entirely on Snowflake Virtual Warehouses — no external compute or container infrastructure. Compatible with Python 3.9-3.13, supporting pandas, scikit-learn, and XGBoost via Snowflake's Anaconda channel.
Snowflake Well-Architected Framework: Cost Optimization & FinOps Guide
Covers the five core FinOps levers: (1) Resource Monitors — setting credit quotas per warehouse with NOTIFY and SUSPEND triggers; (2) Auto-suspend and auto-resume configuration for idle compute elimination; (3) Query profiling using ACCOUNT_USAGE.QUERY_HISTORY to surface the top cost-per-query offenders; (4) Clustering strategy selection using SYSTEM$CLUSTERING_INFORMATION(); (5) Materialized View cost/benefit analysis for frequently reused aggregations. All patterns include ready-to-run SQL examples from the official Snowflake framework.
Use Cases & Customer Stories
Secrets of Gen AI Success: Real-World Customer Stories
Featured production patterns: Sigma Computing embedded Cortex AI into BI dashboards generating natural-language query explanations and auto-suggested follow-on analyses. Terakeet used COMPLETE() over structured SEO data to identify market opportunities 98% faster. IntelyCare applied sentiment analysis (CORTEX.SENTIMENT()) on workforce feedback to reduce nurse burnout indicators. The common technical prerequisite: a clean semantic layer (Snowflake Metrics Layer), column-tagged PII classification, and inference workloads isolated on a dedicated warehouse.
Secrets of Migration Success: Customer Stories from Penske, PayPal & Guitar Center
Penske migrated a complex manufacturing data estate, eliminating on-premise database licenses and consolidating ETL pipelines onto Snowflake Tasks. PayPal standardized access governance across a fragmented data environment using RBAC and dynamic masking policies, achieving SOC 2 compliance readiness within the Snowflake perimeter. Guitar Center's migration unlocked Cortex AI for personalized customer recommendations. Common pattern: 6-12 month migration timelines using Snowflake Migration Service for schema translation, with significant reduction in data engineering maintenance overhead post-migration.
Ecosystem & Industry
Snowflake vs Databricks in 2026: An Honest Comparison
2026 differentiators: Snowflake leads on SQL analytics, governed data sharing (Marketplace + Polaris), multi-cloud governance (Horizon Catalog), and business-user accessibility. Databricks leads on ML/AI workloads (MLflow native, Delta Live Tables), unstructured data, and Unity Catalog federated governance for decentralized architectures. Databricks hit $5.4B ARR at 65%+ growth in February 2026; Snowflake has 9,100+ accounts on AI products with Snowflake Intelligence reaching 2,500 accounts within three months of launch. Apache Iceberg is the neutral ground where both compete on interoperability. Enterprise evaluators increasingly run both in parallel.
Snowflake Expands Open Data Strategy with Iceberg V3 and Governance Portability
The governance portability plan enables Snowflake access policies (row-level security, dynamic masking, object tags) to be exported in an open format and applied in other catalog-compatible environments. Combined with Iceberg V3 support and expanded cross-catalog sharing (Glue, Unity Catalog, Polaris), this positions Snowflake as a governance-as-a-service layer for multi-cloud architectures. The strategy directly counters Databricks Unity Catalog's cross-platform governance narrative. No GA date for portability announced — currently a roadmap commitment. Iceberg V3 is in public preview.
SQL Tips of the Week
Find Oversized Warehouses Before Your Next Bill Arrives
Query snowflake.account_usage.query_history grouping by warehouse_name, filtering for warehouses where avg_exec_seconds is under 30 and query count is above 100. Warehouses with high total_credits but low avg_exec_seconds are your best downsizing candidates. Drop one tier (e.g., LARGE to MEDIUM), run for two weeks, and compare. You'll rarely get user complaints and often see 30-50% cost reduction.
Call GPT-5.2 Directly From SQL with CORTEX.COMPLETE()
Use SNOWFLAKE.CORTEX.COMPLETE('openai-gpt-5-2', prompt) to classify customer feedback into predefined categories. Process in batches of 1,000 records at a time to manage Cortex credit consumption. For high-volume classification tasks, consider SNOWFLAKE.CORTEX.CLASSIFY_TEXT() — it's purpose-built for categorization and consumes fewer Cortex credits than COMPLETE(). Reserve COMPLETE() for tasks where full prompt control and complex reasoning matter.
Tag-Based Dynamic Data Masking for Auto-Governed PHI Columns
Create a role-aware masking policy, then attach it to Snowflake's classification tag via ALTER TAG snowflake.core.privacy_category SET MASKING POLICY. Trigger automatic classification with SYSTEM$CLASSIFY_TABLE() using '{"auto_tag": true}'. Once run, any column tagged with PRIVACY_CATEGORY automatically inherits the masking policy — no ALTER TABLE SET MASKING POLICY required. Schedule classification as a recurring Snowflake Task to catch newly added columns automatically as schemas evolve.
Set Up a Resource Monitor with Hard Credit Limits in 60 Seconds
Create a resource monitor with CREDIT_QUOTA, FREQUENCY = MONTHLY, and trigger thresholds: NOTIFY at 75%, NOTIFY at 90%, SUSPEND at 100%. Attach to warehouses with ALTER WAREHOUSE ... SET RESOURCE_MONITOR. Use FREQUENCY = DAILY on dev and sandbox warehouses where runaway spend is most likely — daily limits catch problems before they compound across a weekend. For production warehouses, use MONTHLY with NOTIFY-only triggers so you get early warnings without risking an accidental mid-month pipeline suspension.
Use the CHANGES Clause for Lightweight Incremental Processing
Enable change tracking with ALTER TABLE ... SET CHANGE_TRACKING = TRUE, then query with CHANGES(INFORMATION => DEFAULT) AT(OFFSET => -86400) END(OFFSET => 0) to get all changes in the last 24 hours. CHANGES reads from Snowflake's time-travel history — no Stream objects, no Stream credits, no offset management. The catch: it's bounded by DATA_RETENTION_TIME_IN_DAYS. For permanent incremental pipelines, prefer Streams. For debugging, backfills, and audit use cases, CHANGES is the cleaner tool.
Ditch the Subquery: Use QUALIFY to Filter Window Function Results
Replace nested subqueries using QUALIFY rn = 1 at the end of a SELECT that includes a window function like ROW_NUMBER() OVER (PARTITION BY customer_id ORDER BY event_time DESC) AS rn. QUALIFY is evaluated after window functions but before ORDER BY — meaning you can reference window function aliases directly in the clause. It's a Snowflake/BigQuery feature that doesn't exist in standard SQL, so add a comment when you use it for teammates coming from Postgres or Redshift.