Snowflake Developers

Part 1 separated the ontology, knowledge graph, and semantic layer. Part 2 builds the full stack on Snowflake. Snowflake Data Superhero Satish Tirumalasetti’s implementation uses: • KG_NODES and KG_EDGES as the physical layer • VARIANT for properties across node and edge types • Ontology metadata for classes, relationships, and rules • A stored procedure that generates typed views Adding “Cinematographer” becomes an INSERT, not an ALTER TABLE. A Cortex Agent then routes metric-driven questions to a semantic view and relationship questions to a recursive traversal UDF. Continue the series with the complete implementation and working code: https://bit.ly/3Vd9A80

Teams can now build Streamlit in Snowflake apps together entirely in Snowsight. With support for shared workspaces now generally available, you can move an app from a personal workspace into a shared workspace and collaborate without using Git. Anyone with access to the workspace can run the development app. Only the workspace owner role can deploy it, so changes don’t reach the live app until they’re ready. Prefer source control? Git-backed workspaces let you edit in Snowsight or locally and sync through Git. 🎥 Watch it in action below, then explore the full guide: https://bit.ly/4h8w1DU

Adding Kimi K3 to Snowflake Cortex AI gives you an open-weight model to evaluate without changing where you build. Use the kimi-k3 model ID with AI_COMPLETE in SQL, or call it through Snowflake’s Chat Completions endpoint using the OpenAI SDK. Available at launch through Cortex AI Functions and Cortex Inference: https://bit.ly/4xO3Zm1

A sub-second query on an idle system tells you very little about an interactive workload. McKnight Consulting Group tested Snowflake under rising concurrency using: • 500 million rows, roughly 1 TB uncompressed • 13 queries spanning point lookups, aggregations, window functions, and self-joins • Concurrency levels from 1 to 32 workers • An X-Small Interactive Warehouse with a configured fallback warehouse At 32 concurrent workers, Snowflake completed 1.68 queries per second with a 0% error rate. Full methodology and results: https://bit.ly/4rwzyiP

PostgreSQL 19 is delayed. That’s the release process doing its job. Since beta began, 53 changes have been reverted as testing and post-commit review uncovered design flaws, compatibility issues, and incorrect results. Among the notable changes: • SQL Property Graph Queries are now PostgreSQL 20 material • ALTER TABLE MERGE/SPLIT PARTITION has been reverted • GROUP BY ALL was pulled after review found it could produce wrong results • REPACK CONCURRENTLY and postgres_fdw statistics import remain at risk If your roadmap depends on a PostgreSQL 19 feature, verify its status now. PostgreSQL 18 remains the safe upgrade target today. See what was reverted, what remains at risk, and what is still expected to ship: https://bit.ly/4yXchJB

Auto-gen Agents for Shared Data can generate a Semantic View and Cortex Agent from an existing share or listing, with no modeling required. Snowflake uses the share metadata and listing description to detect metrics, dimensions and relationships, then creates an agent with context-aware instructions. Review and refine the generated objects, attach them to the share, and updates flow to consumers without republishing. You can explore the data in natural language using your own warehouse compute and token usage. For a better starting point, use descriptive column names, add table and column comments, and write a detailed listing description. Check out the quickstart with sample data: https://bit.ly/46BzGUP

There are several ways to evaluate Claude Opus 5.5 – now in public preview on Snowflake Cortex AI: • Select it with /model in Snowflake CoCo CLI • Pass claude-opus-5-5 to AI_COMPLETE in SQL • Call it through Cortex Inference using the Anthropic SDK and your Snowflake account endpoint Details: https://bit.ly/4AtHqFR

Your agent may have completed the request without an error, but it retrieved the wrong document and gave you a confidently wrong answer. Traditional observability can tell you the request completed. It can’t tell you whether the response was accurate, relevant, or useful. Agent Observability in Observe, Inc. makes quality and cost production signals alongside latency and errors: • Online LLM-as-judge evaluations against production traffic, with results connected to the underlying interactions • Agent Explorer for stepping through prompts, completions, retrieved context, tool calls and retries to find where a workflow diverged from the intended outcome • Latency, error and cost metrics derived from agent traces • Context Graph for connecting agent traces to supporting services, APIs and databases for end-to-end visibility A lower-cost model might not be the better choice if it fails more evaluations. You need quality and cost in the same place to make that call. Dive in: https://bit.ly/4yOBaH7

Building AI spend controls by hand means a metering query, a scheduled task, a procedure that revokes a role, and another to undo it on the first of the month. Per-user quotas are now GA. That plumbing is gone. A quota is a first-class Snowflake object. Set a monthly and daily per-user credit limit, turn on enforcement, and Snowflake runs the loop: • The object is the API. One call for current config, one for who's in scope, one for who's blocked, one for full enforcement history • Blocked users get a "quota exhausted" error at the call site, not a generic auth failure • Scope is declarative through user tags and resolves at evaluation time, so retagging moves a user between quotas • Blocks expire on their own at the UTC cycle boundary. Raise the limit and they clear If you need custom behavior, register your own stored procedure at a threshold and the quota hands you the breaching user IDs. Get started: https://bit.ly/4yimBf9

A standard warehouse exposes over 15 properties to get right; Snowflake Adaptive Compute collapses that into two: ⬩ MAX_QUERY_PERFORMANCE_LEVEL (MQPL) sets how far up the cost-performance curve a single query may climb. ⬩ QUERY_THROUGHPUT_MULTIPLIER (QTM) controls how much concurrent work runs before queries queue. Together, they define an admission-control token budget. For example, with MQPL set to Large and QTM set to 4, that budget could run the equivalent of four Large queries, eight Medium queries, 16 Small queries or any mix that fits. Queries that don’t fit wait until capacity is returned. Warehouse sizes, multi-cluster settings, Query Acceleration Service configuration, and suspend and resume policies all stop being yours to manage. Underneath, every job across your account’s Adaptive Warehouses routes to a shared compute pool dedicated to that account. The warehouses remain distinct endpoints for governance and reporting, while compute capacity can be used across workloads instead of remaining isolated by warehouse. Take a closer look at per-query autosizing, shared compute and the admission control underneath 👇 https://bit.ly/4y9vRCo