<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Soumak Das on Medium]]></title>
        <description><![CDATA[Stories by Soumak Das on Medium]]></description>
        <link>https://medium.com/@soumakdas94?source=rss-6f4539057cdd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*QgKB6xTKSdRHPhxbDzu4Sw.jpeg</url>
            <title>Stories by Soumak Das on Medium</title>
            <link>https://medium.com/@soumakdas94?source=rss-6f4539057cdd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 01:52:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@soumakdas94/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Snowflake Architect & Data Engineer — Cortex LLM]]></title>
            <link>https://soumakdas94.medium.com/snowflake-architect-data-engineer-cortex-llm-0e484882f619?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e484882f619</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[snowflake-cortex]]></category>
            <category><![CDATA[cortex]]></category>
            <category><![CDATA[snowflake]]></category>
            <category><![CDATA[interview-questions]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Thu, 30 Apr 2026 10:21:45 GMT</pubDate>
            <atom:updated>2026-04-30T10:21:45.198Z</atom:updated>
            <content:encoded><![CDATA[<h3>Cortex AI • Governance • External APIs • 2025–2026 Features</h3><h3>🧠 SECTION 1: Cortex LLM Functions — Deep Dive</h3><p><strong>Q1. What is the full Cortex AI Function suite and how do they differ architecturally?</strong></p><p><em>Answer:</em> Snowflake Cortex gives instant access to industry-leading LLMs from Anthropic, Mistral, Reka, Meta, and Google, including Snowflake Arctic — all fully hosted and managed by Snowflake, requiring no setup, keeping data within Snowflake, with the performance, scalability, and governance you expect.</p><p>The complete function taxonomy:</p><p>CategoryFunctionPurpose<strong>Generative</strong>AI_COMPLETEFree-form LLM completion, single prompt or multi-turn conversation<strong>Extraction</strong>AI_EXTRACTExtract structured fields from text, docs, images<strong>Classification</strong>AI_CLASSIFYClassify text/images into user-defined categories<strong>Translation</strong>AI_TRANSLATETranslate between languages<strong>Transcription</strong>AI_TRANSCRIBEAudio/video → text with timestamps and speaker info<strong>Embeddings</strong>AI_EMBED / EMBED_TEXT_768Generate vectors for semantic search<strong>Similarity</strong>AI_SIMILARITYCosine similarity between two inputs without explicit vectors<strong>PII Redaction</strong>AI_REDACTDetect and remove PII from unstructured text<strong>Document Parse</strong>AI_PARSE_DOCUMENTExtract from PDFs, images with schema-aware table extraction<strong>Answer Extraction</strong>EXTRACT_ANSWERQA over unstructured text</p><p>AI_CLASSIFY, AI_TRANSCRIBE, and multimodal AI pipelines running entirely inside Snowflake — enabling text, image, audio, and video intelligence without external services or data movement — reached GA in November 2025.</p><p><strong>Q2. How does </strong><strong>AI_COMPLETE work, and what models can you route to? How do you architect model selection?</strong></p><p><em>Answer:</em> AI_COMPLETE accepts a model name and a prompt string. Available models include Claude Opus 4.5, Claude Sonnet 4.5 (Anthropic), GPT-5.2 (OpenAI), DeepSeek-R1, Meta Llama, and Mistral Large 2 — all accessible via SQL, Python, or REST API within Snowflake&#39;s governed environment.</p><pre>-- Basic usage<br>SELECT AI_COMPLETE(&#39;claude-sonnet-4-5&#39;,<br>  &#39;Summarize the following contract: &#39; || contract_text) AS summary<br>FROM contracts;</pre><pre>-- With system prompt (structured output)<br>SELECT AI_COMPLETE(<br>  &#39;mistral-large2&#39;,<br>  [{&#39;role&#39;: &#39;system&#39;, &#39;content&#39;: &#39;Return JSON only&#39;},<br>   {&#39;role&#39;: &#39;user&#39;, &#39;content&#39;: &#39;Extract: &#39; || text}]<br>) AS structured_result<br>FROM documents;</pre><p><strong>Model selection strategy for architects:</strong></p><ul><li>Use auto for Cortex Agents — Snowflake picks the best available model and upgrades automatically.</li><li>Use Claude Sonnet/Opus for complex reasoning and document understanding.</li><li>Use DeepSeek-R1 for math, code, and multi-step reasoning tasks.</li><li>Use Mistral for speed-sensitive, high-volume batch inference.</li><li>Snowflake recommends using a warehouse no larger than MEDIUM when calling Cortex AI Functions — larger warehouses don’t increase performance but increase costs.</li></ul><p><strong>Q3. What is Cortex Provisioned Throughput and when would an architect use it?</strong></p><p><em>Answer:</em> Cortex Provisioned Throughput reached GA in May 2025. It reserves dedicated token throughput capacity for LLM inference, eliminating the variable latency and rate limits of shared serverless inference. Use it when you have predictable, high-volume production AI workloads (e.g., nightly batch NLP pipelines over millions of rows, customer-facing RAG apps with SLA requirements). Trade-off: higher fixed cost vs. on-demand, but guaranteed throughput and consistent latency.</p><p><strong>Q4. How do you build a production batch NLP pipeline using Cortex AI Functions as a data engineer?</strong></p><p><em>Answer:</em> Pattern:</p><pre>-- Step 1: Land raw documents in a Snowflake table<br>-- Step 2: Apply AI functions in a single SQL pass<br>CREATE OR REPLACE TABLE enriched_support_tickets AS<br>SELECT<br>  ticket_id,<br>  raw_text,<br>  AI_CLASSIFY(&#39;ticket category&#39;, raw_text,<br>    [&#39;Billing&#39;, &#39;Technical&#39;, &#39;Account&#39;, &#39;Shipping&#39;]) AS category,<br>  AI_EXTRACT(raw_text, &#39;Extract: customer_name, product, issue_description&#39;) AS entities,<br>  AI_COMPLETE(&#39;mistral-large2&#39;,<br>    &#39;Rate urgency 1-5 and explain: &#39; || raw_text) AS urgency_score,<br>  AI_REDACT(raw_text) AS redacted_text<br>FROM raw_support_tickets;</pre><p>Key engineer considerations:</p><ul><li>Wrap in a <strong>Dynamic Table</strong> for automatic incremental refresh as new tickets arrive.</li><li>Use a <strong>MEDIUM warehouse</strong> — Cortex AI functions are GPU-bound serverlessly; warehouse size doesn’t accelerate them.</li><li>Monitor spend via CORTEX_FUNCTIONS_USAGE_HISTORY view — billed per token.</li><li>Cortex AI Functions in incremental Dynamic Table refresh is supported since September 2025 — no need for manual task orchestration.</li></ul><p><strong>Q5. How do you govern access to Cortex AI Functions? What’s the </strong><strong>AI_FUNCTIONS_USER role?</strong></p><p><em>Answer:</em> Access to AI_COMPLETE is governed by the CORTEX_USER or AI_FUNCTIONS_USER database role and the USE AI FUNCTIONS account-level privilege. Model access controls govern which models a user can use — secondary roles can obscure permissions, so disable them temporarily when verifying.</p><p>The AI_FUNCTIONS_USER database role became GA in April 2026, providing a clean, grantable role for scoping AI function access to specific users or service accounts without granting broader privileges.</p><pre>-- Grant AI function access to a role<br>GRANT DATABASE ROLE SNOWFLAKE.AI_FUNCTIONS_USER TO ROLE DATA_ANALYST;</pre><pre>-- Budget cap for AI<br>-- Use the new Budgets for AI Features (GA April 2026)<br>ALTER ACCOUNT SET AI_FEATURES_BUDGET = 1000; -- credits</pre><h3>🤖 SECTION 2: Cortex Agents, Analyst &amp; Search</h3><p><strong>Q6. Explain the full Cortex Agent architecture. How does planning, tool use, and reflection work?</strong></p><p><em>Answer:</em> Cortex Agents plan tasks, use tools to execute them, and generate responses. Agents use Cortex Analyst for structured data and Cortex Search for unstructured data, along with LLMs. Planning: agents parse requests to orchestrate a plan, explore options, split tasks into subtasks, and route across tools. Tool use: agents retrieve data efficiently using Cortex Search for unstructured and Cortex Analyst for structured data. Reflection: after each tool use, agents evaluate results to determine next steps — asking for clarification, iterating, or generating a final response.</p><pre>User Query<br>    ↓ (LLM plans)<br>Is it a structured data question?<br>    YES → Cortex Analyst (generates SQL from Semantic Model → executes)<br>    NO  → Cortex Search (hybrid vector+keyword search on documents)<br>    BOTH → Multi-step: Search for context → Analyst for numbers → Compose answer<br>    ↓<br>Reflection: is answer complete? → If not, iterate<br>    ↓<br>Final Response (with citations, charts if applicable)</pre><p>Custom tools (stored procedures, UDFs) can be added for business-specific logic.</p><p><strong>Q7. What is Cortex Analyst and what is a Semantic Model? How do you design one?</strong></p><p><em>Answer:</em> Cortex Analyst is a fully managed, sophisticated agentic AI system that handles the complexities of generating highly accurate text-to-SQL responses, accelerating delivery of high-precision, self-serve conversational analytics to business teams while avoiding complex RAG solution patterns, model experimentation, and GPU capacity planning.</p><p>A <strong>Semantic Model</strong> is a YAML file that describes:</p><ul><li>Table and column business names (aliases human-friendly language)</li><li>Verified queries (pre-approved SQL examples for common questions)</li><li>Relationships between tables (join paths)</li><li>Metrics definitions (e.g., “revenue” = SUM(order_value) WHERE status=&#39;completed&#39;)</li><li>Dimension hierarchies</li></ul><pre># semantic_model.yaml snippet<br>tables:<br>  - name: orders<br>    description: &quot;All customer orders&quot;<br>    base_table: SALES_DB.PUBLIC.ORDERS<br>    columns:<br>      - name: revenue<br>        description: &quot;Total order value&quot;<br>        expr: order_value<br>        data_type: NUMBER<br>    verified_queries:<br>      - question: &quot;What was total revenue last month?&quot;<br>        sql: &quot;SELECT SUM(order_value) FROM orders WHERE ...&quot;</pre><p>Verified queries in semantic views became supported in April 2026, allowing architects to pre-certify trusted SQL patterns for LLM grounding.</p><p><strong>Architect tip:</strong> Poor semantic model design = poor Analyst answers. If your data model has issues like tables missing primary keys, inconsistent naming, or complex unnormalized schemas, the quality of generated SQL will suffer — the agent can’t fix poor data architecture on its own.</p><p><strong>Q8. What is Cortex Search and how does hybrid search work internally?</strong></p><p><em>Answer:</em> Cortex Search provides fully managed text embedding, hybrid search (semantic + keyword), and retrieval for finding information by asking questions within a given set of documents.</p><p>Internally:</p><ul><li><strong>Embedding phase</strong> — Documents are chunked, embedded using Snowflake’s embedding models, and stored in a managed vector index.</li><li><strong>Query phase</strong> — User query is embedded, vector similarity search runs (semantic), AND BM25 keyword search runs in parallel.</li><li><strong>Reranking</strong> — Both result sets are merged and reranked by a cross-encoder model for final relevance.</li></ul><pre>-- Create a Cortex Search service<br>CREATE CORTEX SEARCH SERVICE my_doc_search<br>  ON content_column<br>  ATTRIBUTES doc_type, author<br>  WAREHOUSE = my_wh<br>  TARGET_LAG = &#39;1 hour&#39;<br>  AS SELECT doc_id, content_column, doc_type, author<br>     FROM documents;</pre><p><strong>Architect consideration:</strong> Cortex Search charges based on index size and persistence time — not per query. For large document corpuses, estimate index storage costs before provisioning. Cortex Search’s embedding models are optimized for English; non-English text gets lower-quality semantic search results.</p><p><strong>Q9. How do you architect a multi-source Cortex Agent with both structured and unstructured data?</strong></p><p><em>Answer:</em> Real-world scenario: A financial analyst asks — <em>“What were our top 5 customers by revenue last quarter, and what are their contract renewal terms?”</em></p><p>This requires:</p><ol><li><strong>Structured</strong> → Cortex Analyst → SQL on sales_fact table → top 5 customers + revenue</li><li><strong>Unstructured</strong> → Cortex Search → query contracts document index → retrieve renewal terms for those customers</li><li><strong>Agent orchestration</strong> → combine both results → compose final answer with citations</li></ol><pre># Cortex Agent API call<br>import requests</pre><pre>response = requests.post(<br>  f&quot;https://{account}.snowflakecomputing.com/api/v2/cortex/agent:run&quot;,<br>  headers={&quot;Authorization&quot;: f&quot;Bearer {token}&quot;},<br>  json={<br>    &quot;model&quot;: &quot;auto&quot;,<br>    &quot;messages&quot;: [{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: user_query}],<br>    &quot;tools&quot;: [<br>      {&quot;tool_type&quot;: &quot;cortex_analyst_text_to_sql&quot;,<br>       &quot;cortex_analyst_text_to_sql&quot;: {&quot;semantic_model_file&quot;: &quot;@stage/model.yaml&quot;}},<br>      {&quot;tool_type&quot;: &quot;cortex_search&quot;,<br>       &quot;cortex_search&quot;: {&quot;service&quot;: &quot;MY_DB.MY_SCHEMA.my_doc_search&quot;,<br>                         &quot;max_results&quot;: 5}}<br>    ]<br>  }<br>)</pre><h3>💻 SECTION 3: Cortex Code</h3><p><strong>Q10. What is Cortex Code and what can it do beyond basic SQL autocomplete?</strong></p><p><em>Answer:</em> Cortex Code is an AI coding agent purpose-built for the enterprise data stack, launched in November 2025 and significantly expanded in 2026, now used by more than 50% of Snowflake customers to accelerate development workflows. It supports writing, debugging, and refactoring SQL, Python, and Scala with context-aware suggestions tailored to your specific tables and views; managing catalogs, setting permissions, creating users, and optimizing costs using conversational commands; and building and deploying Snowflake Intelligence agents directly from natural language.</p><p>Cortex Code now supports external data systems including AWS Glue, Databricks, and Postgres, extending Snowflake’s data-native intelligence to enable development without migrating data into Snowflake.</p><p>Key capabilities:</p><ul><li>Context-aware: it reads your schema, table stats, and clustering info before suggesting query optimizations.</li><li>Can generate full Snowpark Python pipelines from natural language descriptions.</li><li>Can explain Query Profile bottlenecks in plain English.</li><li>Can suggest and apply clustering key changes based on query history.</li></ul><p><strong>Q11. How does Cortex Code differ from a general-purpose coding assistant like GitHub Copilot for a Snowflake data engineer?</strong></p><p><em>Answer:</em> Cortex Code has <strong>Snowflake context</strong> that external tools don’t:</p><ul><li>It knows your actual table schemas, column statistics, and clustering state.</li><li>It understands your RBAC model and can suggest properly scoped grants.</li><li>It can read QUERY_HISTORY and METERING_HISTORY to generate cost optimization recommendations grounded in your actual usage — not generic advice.</li><li>All code suggestions run inside Snowflake’s security perimeter — no data leaves.</li><li>It can directly execute suggestions within Snowsight without copy-pasting.</li></ul><h3>🛡️ SECTION 4: Governance — Snowflake Horizon Deep Dive</h3><p><strong>Q12. What is Snowflake Horizon Catalog and how does it differ from the older “Horizon” concept?</strong></p><p><em>Answer:</em> Horizon Catalog is the universal catalog for your entire data estate. It provides context and governance for AI, enables any architecture across clouds and regions, works with any engine and data format, and has zero risk of vendor lock-in. It applies sensitive data classification, retention, and data access policies the same way everywhere — giving every engine a shared view of metadata, lineage, and rules.</p><p>The evolution:</p><ul><li><strong>Old Horizon</strong> — A label for Snowflake’s collection of governance features (tagging, masking, row policies).</li><li><strong>Horizon Catalog (2025+)</strong> — Expanded to be a universal catalog: governs not just Snowflake-native tables but also Iceberg tables, external tables, Marketplace data products, and data accessed by external engines (Spark, AWS Glue, Databricks) via Apache Polaris REST APIs.</li></ul><p>Horizon Catalog governs data inside Snowflake and in external storage through external tables and Iceberg tables. It carries governance into the Marketplace by sharing data products through internal exchanges while preserving tags and access rules.</p><p><strong>Q13. How do Data Metric Functions (DMFs) work and how do you architect a data quality monitoring system?</strong></p><p><em>Answer:</em> DMFs are system or custom functions that compute quality metrics on table columns on a schedule. They are Snowflake’s native approach to data observability.</p><pre>-- Create a custom DMF<br>CREATE DATA METRIC FUNCTION null_percent(<br>  arg_t TABLE(col1 NUMBER))<br>RETURNS NUMBER AS<br>&#39;SELECT RATIO_TO_REPORT(COUNT(*)) WITHIN GROUP (ORDER BY col1 IS NULL) * 100<br> FROM arg_t&#39;;</pre><pre>-- Associate with a table<br>ALTER TABLE orders<br>  ADD DATA METRIC FUNCTION null_percent ON (revenue);</pre><pre>-- Schedule<br>ALTER TABLE orders<br>  SET DATA_METRIC_SCHEDULE = &#39;15 MINUTES&#39;;</pre><pre>-- Query results<br>SELECT * FROM SNOWFLAKE.LOCAL.DATA_QUALITY_MONITORING_RESULTS;</pre><p><strong>Architect pattern:</strong></p><ul><li>Attach DMFs to Silver/Gold layer tables — not Bronze (raw data is expected to be dirty).</li><li>Use system DMFs for common checks: SNOWFLAKE.CORE.NULL_COUNT, DUPLICATE_COUNT, FRESHNESS.</li><li>The updated privilege model allows non-owners to associate a data metric function with an object — enabling data stewards to attach quality rules without needing table ownership.</li><li>Trigger alerts via CREATE ALERT when DMF scores breach thresholds.</li></ul><p><strong>Q14. Design a comprehensive PII governance architecture in Snowflake using Horizon tools.</strong></p><p><em>Answer:</em></p><p><strong>Layer 1 — Discovery:</strong></p><ul><li>Enable <strong>Automatic Sensitive Data Classification</strong> — ML models scan column names and sample values to tag PRIVACY_CATEGORY (Name, Email, SSN, etc.) automatically.</li><li>Medical and health data classifiers in sensitive data classification reached GA in April 2026 — extend automatic discovery to HIPAA-relevant fields.</li><li>Excluding specific objects from sensitive data classification is GA since November 2025 — exclude known non-sensitive tables from scan overhead.</li></ul><p><strong>Layer 2 — Protection:</strong></p><pre>-- Dynamic masking for SSN<br>CREATE MASKING POLICY mask_ssn AS (val STRING) RETURNS STRING -&gt;<br>  CASE WHEN CURRENT_ROLE() IN (&#39;PII_ACCESS_ROLE&#39;) THEN val<br>       ELSE &#39;***-**-&#39; || RIGHT(val, 4)<br>  END;</pre><pre>ALTER TABLE customers MODIFY COLUMN ssn SET MASKING POLICY mask_ssn;</pre><pre>-- Row-level policy for regional compliance<br>CREATE ROW ACCESS POLICY region_policy AS (region STRING) RETURNS BOOLEAN -&gt;<br>  region = CURRENT_SESSION_PARAMETER(&#39;USER_REGION&#39;)<br>  OR CURRENT_ROLE() = &#39;DATA_STEWARD&#39;;</pre><p><strong>Layer 3 — Audit:</strong></p><ul><li>Query ACCESS_HISTORY to see which users accessed which tagged columns and when.</li><li>Use POLICY_REFERENCES() to audit all tables with masking/row policies applied.</li></ul><p><strong>Layer 4 — AI-aware:</strong></p><ul><li>Apply AI_REDACT in ETL pipelines before data reaches shared analytics layers.</li><li>Cortex AI functions <strong>inherit Snowflake’s RBAC</strong> — a user cannot use AI_COMPLETE to circumvent masking policies on columns they can&#39;t directly read.</li></ul><p><strong>Q15. What is the Trust Center and how do you use it for continuous security posture management?</strong></p><p><em>Answer:</em> Trust Center supports detection findings, event-driven scanners, and organization-level findings. Trust Center notifications in Snowsight are GA since November 2025.</p><p>The Trust Center provides:</p><ul><li><strong>Security scanners</strong> — Pre-built checks for: MFA enforcement gaps, overly permissive roles, publicly accessible stages, network policy gaps, password policy compliance.</li><li><strong>Event-driven scanners</strong> — Trigger scans when account configuration changes (e.g., a new role is created with ACCOUNTADMIN).</li><li><strong>Organization-level findings</strong> — Aggregate security posture across all accounts in your Snowflake organization.</li><li><strong>Custom scanners</strong> — Write your own compliance checks as SQL queries.</li></ul><p><strong>Architect use case:</strong> Integrate Trust Center findings with your SIEM (via NOTIFICATION INTEGRATION) to alert the security team when critical findings are detected, without manual polling.</p><h3>🌐 SECTION 5: External APIs &amp; Network Access</h3><p><strong>Q16. How does External Network Access work in Snowflake, and how do you architect it securely?</strong></p><p><em>Answer:</em> External Network Access (ENA) allows Snowflake code (Snowpark UDFs, stored procedures, SPCS services) to make outbound HTTP/HTTPS calls to external APIs — controlled via <strong>Network Rules</strong> and <strong>External Access Integrations</strong>.</p><pre>-- Step 1: Define allowed external endpoints<br>CREATE NETWORK RULE my_api_rule<br>  TYPE = HOST_PORT<br>  MODE = EGRESS<br>  VALUE_LIST = (&#39;api.myvendor.com:443&#39;);</pre><pre>-- Step 2: Store credentials securely<br>CREATE SECRET my_api_key<br>  TYPE = GENERIC_STRING<br>  SECRET_STRING = &#39;{&quot;api_key&quot;: &quot;sk-...&quot;}&#39;;</pre><pre>-- Step 3: Create integration<br>CREATE EXTERNAL ACCESS INTEGRATION my_api_integration<br>  ALLOWED_NETWORK_RULES = (my_api_rule)<br>  ALLOWED_AUTHENTICATION_SECRETS = (my_api_key)<br>  ENABLED = TRUE;</pre><pre>-- Step 4: Use in a Snowpark UDF<br>CREATE FUNCTION call_external_api(payload STRING)<br>RETURNS STRING<br>LANGUAGE PYTHON<br>EXTERNAL_ACCESS_INTEGRATIONS = (my_api_integration)<br>SECRETS = (&#39;api_key&#39; = my_api_key)<br>AS $$<br>import requests, _snowflake<br>key = _snowflake.get_generic_secret_string(&#39;api_key&#39;)<br>r = requests.post(&quot;https://api.myvendor.com/enrich&quot;,<br>                  headers={&quot;Authorization&quot;: key},<br>                  json={&quot;data&quot;: payload})<br>return r.json()[&#39;result&#39;]<br>$$;</pre><p><strong>Security governance checklist:</strong></p><ul><li>Network Rules scope to minimum required hosts/ports — no 0.0.0.0/0.</li><li>Secrets stored in Snowflake Secrets (never hardcoded in function body).</li><li>Audit external calls via QUERY_HISTORY — ENA calls are logged.</li><li>Outbound network traffic to stages and volumes on Google Cloud Storage supports private connectivity (GA) — use Private Service Connect for GCP to avoid public internet exposure.</li></ul><p><strong>Q17. How do you call Snowflake’s own REST APIs from external applications? What are Programmatic Access Tokens (PATs)?</strong></p><p><em>Answer:</em> Snowflake exposes REST APIs for SQL execution, Cortex Agents, Snowpipe, and more. Authentication options:</p><ul><li><strong>OAuth 2.0</strong> — For user-delegated access (BI tools, web apps).</li><li><strong>Key-Pair JWT</strong> — For service accounts and machine-to-machine.</li><li><strong>Programmatic Access Tokens (PATs)</strong> — The single-role restriction for service users was removed for PATs, allowing PATs to carry multiple roles for flexible service account access.</li></ul><pre># Calling Cortex Agent REST API externally<br>import requests, snowflake.connector</pre><pre># Generate PAT or use key-pair JWT<br>token = get_snowflake_jwt(account, user, private_key)</pre><pre>response = requests.post(<br>  f&quot;https://{account}.snowflakecomputing.com/api/v2/cortex/agent:run&quot;,<br>  headers={<br>    &quot;Authorization&quot;: f&quot;Bearer {token}&quot;,<br>    &quot;Content-Type&quot;: &quot;application/json&quot;,<br>    &quot;X-Snowflake-Authorization-Token-Type&quot;: &quot;KEYPAIR_JWT&quot;<br>  },<br>  json={<br>    &quot;model&quot;: &quot;claude-sonnet-4-5&quot;,<br>    &quot;messages&quot;: [{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: query}],<br>    &quot;tools&quot;: [{&quot;tool_type&quot;: &quot;cortex_analyst_text_to_sql&quot;, ...}]<br>  },<br>  stream=True  # Supports Server-Sent Events (SSE) streaming<br>)</pre><p><strong>Architect pattern:</strong> Use key-pair JWT for backend services. Use OAuth with PKCE for frontend/user-delegated flows. Never use username/password for API access in production.</p><h3>🔷 SECTION 6: Snowflake Openflow &amp; Ingestion Architecture</h3><p><strong>Q18. What is Snowflake Openflow and how does it change the ingestion architecture?</strong></p><p><em>Answer:</em> Openflow automates data ingestion and integration from varied sources — including unstructured data — to unify enterprise data within Snowflake’s data lakehouse. The managed, cloud-native implementation of Apache NiFi provides visual drag-and-drop data flow creation.</p><p>Snowflake Openflow enables enterprises to connect all of their data from disparate sources and catalogs, making it accessible for AI agents to drive value — all with consistent, built-in security and governance.</p><p><strong>What it replaces/supplements in the architecture:</strong></p><ul><li>Eliminates need for external Kafka Connect + Debezium for many CDC use cases.</li><li>Replaces standalone Apache NiFi clusters for event-driven file ingestion.</li><li>Supports processing events inside your VPC/VNet — data never leaves your network boundary until it lands in Snowflake.</li><li>Direct integration with Cortex for enriching unstructured data during ingestion.</li></ul><p><strong>Data engineer decision matrix:</strong></p><p>Use CaseRecommended ApproachFile-based batch ingestionCOPY INTO / SnowpipeReal-time row-level CDC from DBSnowpipe Streaming + Debezium or OpenflowAPI / webhook dataOpenflow (built-in HTTP source connector)Kafka-based streamingSnowflake Kafka Connector or OpenflowUnstructured file enrichmentOpenflow → Cortex AI Functions inline</p><h3>🏗️ SECTION 7: New Platform Features — 2025/2026</h3><p><strong>Q19. What is Snowflake Postgres and when would you recommend it?</strong></p><p><em>Answer:</em> Snowflake Postgres was announced as public preview in November 2025, enabling customers to leverage real-time transactional data faster and more efficiently.</p><p>It provides a Postgres-compatible transactional database layer inside Snowflake, bridging the OLTP/OLAP gap. Recommended when:</p><ul><li>Applications need low-latency, row-level transactional writes (Snowflake native tables have ~1s latency minimum).</li><li>You want to eliminate a separate Postgres RDS instance and consolidate into one governed platform.</li><li>You need Postgres wire protocol compatibility for existing application drivers.</li></ul><p><strong>Architect caution:</strong> This is not a replacement for Aurora/RDS for high-TPS OLTP systems. It targets hybrid workloads — transactional data that is immediately queryable for analytics.</p><p><strong>Q20. What are Interactive Tables and Warehouses?</strong></p><p><em>Answer:</em> Interactive Tables and Warehouses deliver immediate, interactive insights with generally available availability, redefining how enterprises can build and power AI agents and apps by helping organizations turn data into near real-time experiences.</p><p>Interactive Warehouses are optimized for sub-second query latency on small, frequently-accessed datasets — unlike Standard Warehouses which are optimized for throughput on large scans. Designed for embedded analytics, live dashboards, and AI agent data retrieval where latency SLAs are milliseconds, not seconds.</p><p><strong>Q21. What is Snowflake Intelligence and what new capabilities did the April 2026 update add?</strong></p><p><em>Answer:</em> Snowflake Intelligence is the converged agentic experience for business users built on Cortex Agents + Cortex Analyst + Cortex Search.</p><p>As of April 2026, Snowflake Intelligence now serves as a personal work agent for business users that adapts over time by learning individual preferences and workflows to deliver more relevant results and automate tasks — all while enabling deep, trusted insights grounded in governed enterprise data.</p><p>New April 2026 capabilities:</p><ul><li>Deep Research (public preview soon) helps users answer their most complex questions with fully cited, multi-step reports, using an agentic architecture to reason across structured data, unstructured content, and external context.</li><li>Artifacts (GA soon) allow users to save and share analyses, visualizations, and workflows, turning one-off outputs into reusable knowledge so teams can build on each other’s work and scale insights across the organization.</li><li>Personalization — learns user-specific query patterns and preferred data sources over time.</li></ul><h3>🧪 SECTION 8: Architect Scenario Questions (2025/2026 Focused)</h3><p><strong>Q22. Design an AI-powered contract intelligence system entirely in Snowflake.</strong></p><p><em>Answer:</em></p><pre>Stage (PDF contracts)<br>  → Snowpipe Streaming auto-ingest<br>  → AI_PARSE_DOCUMENT (extract clauses, parties, dates, values)<br>  → AI_EXTRACT (extract: renewal_date, penalty_clause, governing_law)<br>  → AI_CLASSIFY (contract_type: NDA, MSA, SOW, Lease)<br>  → AI_REDACT (remove PII for analyst-facing views)<br>  → Gold table: contracts_enriched (structured + semi-structured)<br>  → Cortex Search service on full contract text<br>  → Cortex Analyst semantic model on contracts_enriched<br>  → Cortex Agent (answers: &quot;Which contracts expire in Q3 with auto-renew?&quot;)<br>  → Snowflake Intelligence UI for business users</pre><p>Governance: Tag CONTRACT_VALUE, RENEWAL_DATE with sensitivity tags. Apply row access policy by deal team. Audit via ACCESS_HISTORY.</p><p><strong>Q23. A Cortex Analyst implementation is generating incorrect SQL for ~20% of business queries. How do you debug and fix it as an architect?</strong></p><p><em>Answer:</em></p><ol><li><strong>Check the Semantic Model</strong> — the most common root cause. Are column descriptions business-friendly? Are join paths defined? Are ambiguous metric definitions (e.g., “revenue” could mean gross or net) clarified?</li><li><strong>Add Verified Queries</strong> — Pre-certify correct SQL for the most common 20–30 questions as examples. Support for verified queries in semantic views reached GA in April 2026. The LLM uses these as few-shot examples.</li><li><strong>Use Cortex Analyst Routing Mode</strong> — Cortex Analyst Routing Mode entered preview in November 2025 — routes ambiguous questions to clarification flows instead of hallucinating SQL.</li><li><strong>Monitor</strong> via CORTEX_ANALYST_USAGE_HISTORY — identify patterns in failed queries.</li><li><strong>Normalize the data model</strong> — if tables are over-denormalized or have non-intuitive schemas, Analyst struggles. Introduce intermediary semantic views.</li><li><strong>Test iteratively</strong> in the Cortex LLM Playground — the Cortex LLM Playground lets you perform side-by-side comparisons of text completions between any two models or different prompt configurations, then copy the generated SQL directly to a worksheet.</li></ol><p><strong>Q24. How do you extend Snowflake Horizon governance to external engines (Spark, AWS Glue, Databricks) querying Iceberg tables?</strong></p><p><em>Answer:</em> External engines connect to the Iceberg REST endpoint at https://&lt;account_identifier&gt;.snowflakecomputing.com/polaris/api/catalog using External OAuth, key-pair/JWT, or Programmatic Access Token. Snowflake Horizon Catalog vends credentials to the metadata — no need to manage storage credentials in the external engine. Access checks use Snowflake users and roles; you can scope different tokens by role (e.g., DATA_ENGINEER vs DATA_ANALYST).</p><p>This means: Horizon Catalog enforces row access and data masking policies on Apache Iceberg tables that you query from Apache Spark through Snowflake Horizon Catalog — the same policies that apply inside Snowflake apply to Spark queries. True unified governance across engines.</p><pre># Spark config for Horizon Catalog<br>spark.conf.set(&quot;spark.sql.catalog.snowflake&quot;,<br>  &quot;org.apache.iceberg.spark.SparkCatalog&quot;)<br>spark.conf.set(&quot;spark.sql.catalog.snowflake.catalog-impl&quot;,<br>  &quot;org.apache.iceberg.rest.RESTCatalog&quot;)<br>spark.conf.set(&quot;spark.sql.catalog.snowflake.uri&quot;,<br>  &quot;https://myaccount.snowflakecomputing.com/polaris/api/catalog&quot;)<br>spark.conf.set(&quot;spark.sql.catalog.snowflake.token&quot;, pat_token)</pre><blockquote><strong><em>Summary cheat sheet for interviews:</em></strong></blockquote><p>Cortex = AI suite: LLM Functions + Analyst + Search + Agents + Code + Intelligence</p><p>Horizon = Governance suite: Catalog + Trust Center + DMFs + Classification + Masking + Lineage</p><p>Openflow = Apache NiFi-based managed ingestion (replaces external ETL connectors)</p><p>Iceberg + Horizon Catalog = multi-engine lakehouse with unified governance via Polaris REST API</p><p>External Network Access = controlled outbound API calls from Snowpark with secrets and network rules</p><p>PAT / OAuth / Key-Pair JWT = three authentication patterns for REST API integrations</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e484882f619" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 2026 Snowflake Interview Guide]]></title>
            <link>https://soumakdas94.medium.com/the-2026-snowflake-interview-guide-a90209440fa0?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/a90209440fa0</guid>
            <category><![CDATA[architecture]]></category>
            <category><![CDATA[snowflake]]></category>
            <category><![CDATA[interview]]></category>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <category><![CDATA[interview-questions]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Wed, 29 Apr 2026 08:19:59 GMT</pubDate>
            <atom:updated>2026-04-29T08:19:59.523Z</atom:updated>
            <content:encoded><![CDATA[<p>The Snowflake interview landscape has fundamentally shifted. If you’re stepping into an interview for a Senior Data Engineer or Architect role today, simply reciting documentation won’t get you the job. Hiring managers in complex enterprise environments aren’t looking for someone who can just define a micro-partition or explain basic time travel; they are looking for platform leaders who can design cost-efficient, AI-ready data pipelines and enforce bulletproof governance across the organization. After years of architecting enterprise-grade Lakehouse solutions and navigating the realities of the modern data stack, I’ve seen exactly where candidates stumble. This isn’t your standard list of trivia — it’s the definitive guide to the architectural scenarios, cost optimization strategies, and Cortex AI integrations that actually determine who gets the offer.</p><p>Here is a deep, architect-level Snowflake interview guide — structured around the domains that matter most at that seniority level:</p><h3>❄️ Snowflake Architect-Level Interview Questions</h3><h3>🏛️ SECTION 1: Account &amp; Org Design Strategy</h3><p><strong>Q1. When would you design a multi-account Snowflake strategy vs. a single-account strategy? What are the trade-offs?</strong></p><p><em>Answer:</em> A key architect decision is identifying use cases appropriate for single vs. multiple account strategies, considering isolation/segmentation needs, key constraints, and which features can be leveraged across accounts.</p><p>ConsiderationSingle AccountMulti-AccountCost managementSimpler, centralizedPer-account budgets, more granularData isolationRole-based, logicalPhysical — hard boundaryCross-team governanceEasierRequires Org-level policiesDR / Business continuityOne failover planIndependent failover per accountReplication overheadNoneCross-account replication needed</p><p><strong>Best practice:</strong> Use multi-account for regulatory isolation (e.g., PII in a separate account), environment separation (dev/staging/prod), or regional data residency compliance. Use single-account when governance complexity and cost outweigh isolation benefits.</p><p><strong>Q2. How does Snowflake’s object hierarchy affect architectural decisions?</strong></p><p><em>Answer:</em> The hierarchy flows from Organization → Account → Database → Schema → Tables, Views, Stages, File Formats, Functions, Procedures, Streams, and Tasks. Design decisions cascade down — account-level parameters affect all objects below; database roles scope within a database, while account roles scope across databases.</p><p>Architect impact:</p><ul><li>Grant structure design (functional roles vs. access roles)</li><li>Schema-level isolation for multi-tenancy</li><li>Database-level Iceberg/external volume defaults</li><li>Task and Stream ownership boundaries</li></ul><p><strong>Q3. How would you design a multi-tenancy architecture in Snowflake?</strong></p><p><em>Answer:</em> Three patterns, each with trade-offs:</p><p><strong>Silo Model</strong> — One database (or account) per tenant. Maximum isolation, but high overhead and scaling costs.</p><p><strong>Pool Model</strong> — Shared tables with a tenant_id column. Cost-efficient, but requires Row Access Policies on every sensitive table to enforce data isolation and prevent cross-tenant leakage.</p><p><strong>Bridge Model</strong> — Shared schemas with tenant-specific views and masking policies. Balances isolation and efficiency.</p><p>For regulated industries (healthcare, fintech), the Silo or Bridge model is preferred. At scale, use <strong>Row Access Policies</strong> + <strong>Column Masking Policies</strong> + <strong>Secondary Roles</strong> to enforce tenant boundaries within the Pool model.</p><h3>⚙️ SECTION 2: Compute Architecture &amp; Warehouse Design</h3><p><strong>Q4. How do you design a warehouse strategy for a mixed workload environment (ETL + BI + ad-hoc analytics + ML)?</strong></p><p><em>Answer:</em> Workload isolation through dedicated warehouses per use case is the foundational principle:</p><ul><li><strong>ETL Warehouse</strong> — Large/XL, no auto-suspend (or long suspend window), single-cluster. Predictable, scheduled workloads.</li><li><strong>BI/Dashboard Warehouse</strong> — Medium, multi-cluster enabled, aggressive auto-suspend. Bursty concurrency.</li><li><strong>Ad-hoc / Data Science Warehouse</strong> — XL or 2XL with Query Acceleration Service enabled for large unpredictable scans.</li><li><strong>Snowpark / ML Warehouse</strong> — Snowpark-Optimized type for memory-intensive Python workloads.</li></ul><p>Key architect consideration: Never let ETL and BI share a warehouse — ETL long-running jobs starve BI concurrency queues.</p><p><strong>Q5. When would you use multi-cluster warehouses, and what are the scaling policy trade-offs (Economy vs. Standard)?</strong></p><p><em>Answer:</em></p><ul><li><strong>Multi-cluster warehouses</strong> add clusters automatically when concurrency queues build up. Ideal for BI tools like Tableau or Power BI with many simultaneous users.</li><li><strong>Standard policy</strong> — Adds clusters as soon as a query is queued. Prioritizes performance, higher cost.</li><li><strong>Economy policy</strong> — Waits for 6+ minutes of queued workload before scaling out. Prioritizes cost, accepts latency.</li></ul><p>Architect guidance: Use Standard for customer-facing dashboards (SLA-bound), Economy for internal analyst tools where slight delay is acceptable.</p><p><strong>Q6. Explain the difference between scaling UP vs. scaling OUT in Snowflake and when you’d use each.</strong></p><p><em>Answer:</em></p><ul><li><strong>Scale UP</strong> (increase warehouse size XS → XL) — More memory and CPU per node. Best for complex queries, large joins, heavy transformations, and memory spills.</li><li><strong>Scale OUT</strong> (multi-cluster) — More warehouses running in parallel. Best for high concurrency, many simultaneous users hitting the same warehouse.</li></ul><p>Common architect mistake: Scaling up to fix a concurrency problem (it won’t help), or scaling out to fix a slow complex query (it won’t help either).</p><h3>🗃️ SECTION 3: Data Modeling &amp; Schema Design</h3><p><strong>Q7. When would you choose Star Schema vs. Data Vault vs. Wide Denormalized tables in Snowflake?</strong></p><p><em>Answer:</em></p><p>ModelBest ForSnowflake Fit<strong>Star Schema</strong>BI/reporting, stable dimensionsExcellent — columnar storage handles joins well<strong>Data Vault</strong>Auditable, source-agnostic ingestion at scaleGood — Hub/Link/Satellite maps well to Snowflake tables<strong>Wide/Flat Tables</strong>ML feature stores, real-time servingExcellent — minimizes join cost; benefits from clustering<strong>OBT (One Big Table)</strong>High-frequency, low-latency dashboardsGood when clustering is tuned correctly</p><p>Snowflake’s MPP engine handles star schema joins efficiently — unlike traditional RDBMS, you don’t need aggressive denormalization. However, for ML pipelines, wide feature tables with clustering keys on entity IDs are often optimal.</p><p><strong>Q8. How do you design clustering keys, and when should you NOT use them?</strong></p><p><em>Answer:</em> Clustering keys reorder micro-partitions to co-locate related data, improving partition pruning for filter-heavy queries.</p><p><strong>When to use:</strong></p><ul><li>Large tables (&gt;1TB) that are frequently filtered on a non-insertion-order column</li><li>Date/time columns used in range filters</li><li>High-cardinality columns used in WHERE clauses alongside low-cardinality ones</li></ul><p><strong>Design rules:</strong></p><ul><li>Put the lower-cardinality column first (e.g., (region, date) not (date, region))</li><li>Limit to 3–4 columns max</li><li>Monitor SYSTEM$CLUSTERING_INFORMATION to measure DML depth and overlap</li></ul><p><strong>When NOT to use:</strong></p><ul><li>Small tables (&lt; 1 TB) — overhead exceeds benefit</li><li>Tables with frequent bulk inserts/deletes (constant re-clustering cost)</li><li>Columns used in equality filters only with high cardinality (use search optimization instead)</li></ul><p><strong>Q9. How do you handle large-scale schema evolution in Snowflake without breaking downstream consumers?</strong></p><p><em>Answer:</em></p><ul><li>Use <strong>Schema Evolution</strong> on Snowpipe/COPY INTO with MATCH_BY_COLUMN_NAME to auto-add new columns.</li><li>Version views on top of base tables — consumers always query v_orders_v2, not raw tables.</li><li>Use <strong>Streams</strong> on base tables so downstream tasks receive only incremental deltas; schema changes don’t break the stream offset.</li><li>For breaking changes (column renames, type changes), maintain backward-compatible views during a deprecation window.</li><li>Tag columns with Data Classification + lineage tracking via Snowflake Horizon to identify all downstream dependencies before making changes.</li></ul><h3>🔄 SECTION 4: Data Pipeline Architecture</h3><p><strong>Q10. Design a near-real-time CDC pipeline from an OLTP source (e.g., Postgres) into Snowflake.</strong></p><p><em>Answer:</em> <strong>Architecture:</strong></p><pre>Postgres → Debezium (CDC) → Kafka → Snowpipe Streaming SDK → Snowflake Landing Table<br>→ Stream → Dynamic Table (or Task + Merge) → Serving Layer</pre><p>Key design decisions:</p><ul><li>Use <strong>Snowpipe Streaming</strong> (not file-based Snowpipe) for row-level, sub-second ingestion</li><li>Land raw CDC events in an append-only staging table (INSERT only, never UPDATE)</li><li>Apply MERGE logic in a Task/Dynamic Table to materialize the “current state” view</li><li>Use <strong>Streams with APPEND_ONLY=TRUE</strong> on staging for efficiency (skips delete-tracking overhead)</li><li>For out-of-order events, use event timestamp in the MERGE key, not ingestion timestamp</li></ul><p><strong>Q11. Streams + Tasks vs. Dynamic Tables — when do you choose one over the other?</strong></p><p><em>Answer:</em></p><p>CriterionStreams + TasksDynamic Tables<strong>Complexity</strong>Higher (manual DAG wiring)Lower (declarative)<strong>Latency control</strong>Precise (cron or event-triggered)Target Lag (approximate)<strong>Error handling</strong>Full control (retry logic in Stored Procs)Limited<strong>Multi-step pipelines</strong>Natural (task DAGs)Chains of dynamic tables<strong>Backfill support</strong>ManualBuilt-in full refresh<strong>Cost</strong>Warehouse or serverlessServerless (preferred)</p><p><strong>Architect guidance:</strong> Prefer Dynamic Tables for greenfield pipelines with standard transformation patterns. Use Streams + Tasks when you need fine-grained control, complex retry logic, or notification-driven triggers.</p><p><strong>Q12. How do you architect a Data Lakehouse on Snowflake using Apache Iceberg?</strong></p><p><em>Answer:</em> Snowflake supports creating and managing external volumes for Apache Iceberg tables via Snowsight, including granting USAGE privileges, verifying storage authentication, and setting external volumes as defaults at account, database, or schema level.</p><p>Key architectural patterns:</p><ul><li><strong>Snowflake-managed Iceberg tables</strong> — Snowflake writes the Iceberg metadata; use when Snowflake is the primary engine. Allows full DML.</li><li><strong>Externally-managed Iceberg tables</strong> — Catalog (e.g., AWS Glue, Databricks Unity Catalog) owns the metadata; Snowflake reads and writes via REST catalog protocol.</li><li><strong>Horizon Catalog (Apache Polaris)</strong> — Write to Snowflake-managed Apache Iceberg tables via external query engines like Apache Spark using a single Horizon Catalog endpoint, leveraging existing users, roles, policies, and authentication.</li></ul><p>Design principle: Use Iceberg when you need multi-engine access (Spark for ML, Snowflake for SQL analytics, Databricks for notebooks) on the same dataset without data duplication.</p><h3>🔐 SECTION 5: Security &amp; Governance Architecture</h3><p><strong>Q13. Design an enterprise RBAC model for Snowflake. What are functional roles vs. access roles?</strong></p><p><em>Answer:</em> The architect must configure an RBAC hierarchy that accounts for privilege inheritance, database roles, system roles and best practices, and the distinction between functional roles and access roles.</p><p><strong>Pattern:</strong></p><pre>SYSADMIN<br>  └── FUNCTIONAL_ROLE_DATA_ENGINEER<br>        └── ACCESS_ROLE_DB_FINANCE_READ<br>        └── ACCESS_ROLE_DB_FINANCE_WRITE<br>  └── FUNCTIONAL_ROLE_ANALYST<br>        └── ACCESS_ROLE_DB_FINANCE_READ<br>        └── ACCESS_ROLE_SCHEMA_MARKETING_READ</pre><ul><li><strong>Access Roles</strong> — Granted privileges on objects (READ, WRITE, MODIFY). Named by object + permission. Never assigned directly to users.</li><li><strong>Functional Roles</strong> — Assigned to users. Aggregate multiple access roles reflecting job function.</li></ul><p>This two-layer model simplifies auditing, reduces privilege sprawl, and makes onboarding/offboarding clean — just change the functional role.</p><p><strong>Q14. How do you implement column-level and row-level security for a regulated dataset (e.g., PII in GDPR scope)?</strong></p><p><em>Answer:</em></p><ul><li><strong>Dynamic Data Masking</strong> — Create a masking policy that returns NULL or a hash for users without the PII_ACCESSrole. Apply to SSN, email, phone columns. The data is never decrypted at query time for unprivileged users.</li><li><strong>Row Access Policy</strong> — Define a mapping table (role → allowed tenant/region). Attach the policy to the table. Snowflake filters rows before returning results — transparent to the query.</li><li><strong>Column Tagging</strong> — Tag PII columns with SNOWFLAKE.CORE.PRIVACY_CATEGORY. Use Horizon&#39;s data classification to auto-discover and surface untagged PII.</li><li><strong>Audit</strong> — Use SNOWFLAKE.ACCOUNT_USAGE.QUERY_HISTORY and ACCESS_HISTORY views to track which users accessed tagged columns and when.</li></ul><p><strong>Q15. What is Tri-Secret Secure and when would an architect recommend it?</strong></p><p><em>Answer:</em> Tri-Secret Secure involves encrypting data using a composite key that combines a Snowflake-managed key with a customer-managed key via a cloud KMS (AWS KMS, Azure Key Vault, or GCP Cloud KMS). This means Snowflake alone cannot decrypt data — both Snowflake and the customer must participate.</p><p>Recommend when: regulatory frameworks (HIPAA, FedRAMP High, financial services) require proof that the cloud vendor cannot unilaterally access customer data. Trade-off: adds KMS latency (~5–10% query overhead) and operational complexity — if the customer-managed key is rotated or revoked, all access stops immediately.</p><h3>💰 SECTION 6: Cost Architecture &amp; FinOps</h3><p><strong>Q16. How do you architect a cost governance model for a large Snowflake deployment?</strong></p><p><em>Answer:</em> Layers of cost control:</p><ol><li><strong>Resource Monitors</strong> — Set credit quotas at account, warehouse, or resource group level. Trigger notifications or auto-suspend when thresholds hit.</li><li><strong>Warehouse right-sizing</strong> — Use QUERY_HISTORY to measure actual credit consumption and queue time. Downsize warehouses that are consistently underutilized.</li><li><strong>Auto-suspend tuning</strong> — Default 10 min is wasteful for sporadic workloads. Set to 60–120 seconds for BI warehouses.</li><li><strong>Query Acceleration Service</strong> — Offloads eligible queries to serverless compute; can reduce warehouse size needed for outlier queries.</li><li><strong>Search Optimization</strong> — Evaluate cost vs. query frequency before enabling; it adds storage overhead.</li><li><strong>Budgets for AI features</strong> — Now GA (April 2026), allowing architects to cap Cortex/AI spend separately from compute budgets.</li><li><strong>Storage costs</strong> — Manage Time Travel retention (default 1 day for dev tables, 90 days only where needed). Aggressive Time Travel on large transactional tables is a hidden cost driver.</li></ol><p><strong>Q17. A team reports unexpectedly high Snowflake credits. Walk through your troubleshooting approach as an architect.</strong></p><p><em>Answer:</em> Step-by-step diagnostic:</p><ol><li>Query ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY — identify which warehouse consumed the most credits and when.</li><li>Drill into QUERY_HISTORY for that warehouse/time window — find long-running or high-credit queries.</li><li>Check QUERY_ACCELERATION_ELIGIBLE flag — are large scans missing QAS?</li><li>Look at AUTO_CLUSTERING_HISTORY — automatic re-clustering can silently consume significant credits on frequently updated large tables.</li><li>Check PIPE_USAGE_HISTORY — Snowpipe serverless costs are separate from warehouse credits.</li><li>Review DATA_TRANSFER_HISTORY — cross-cloud or cross-region egress is often an overlooked cost driver.</li><li>Check MATERIALIZED_VIEW_REFRESH_HISTORY — background MV refreshes run on their own serverless compute.</li></ol><h3>🌐 SECTION 7: Disaster Recovery &amp; Business Continuity</h3><p><strong>Q18. How do you design a DR strategy for a mission-critical Snowflake deployment?</strong></p><p><em>Answer:</em> Architects must determine appropriate data recovery solutions, including Time Travel, table types, costs, availability, query performance impacts, data corruption impacts, zero-copy cloning, and Fail-safe considerations.</p><p>Full DR architecture:</p><ul><li><strong>Cross-region replication</strong> — Use CREATE REPLICATION GROUP to replicate databases, shares, and resource monitors to a secondary region. RPO = replication lag (near real-time with Business Critical).</li><li><strong>Failover</strong> — Promote the secondary to primary with ALTER FAILOVER GROUP … PRIMARY. Update connection strings (use DNS-based connection abstraction).</li><li><strong>RTO target</strong> — Failover takes minutes, not hours. Test the failover procedure quarterly.</li><li><strong>Point-in-time recovery</strong> — Time Travel (up to 90 days) handles data corruption. Fail-safe (7 days) is the last resort, requires Snowflake support.</li><li><strong>Client driver resilience</strong> — Use connection retry logic and the Snowflake privatelink endpoint in DR scenarios for private connectivity.</li></ul><h3>🤖 SECTION 8: AI/ML Architecture on Snowflake (Latest)</h3><p><strong>Q19. How would you architect an end-to-end ML pipeline entirely within Snowflake?</strong></p><p><em>Answer:</em></p><pre>Raw Data (Tables/Iceberg)<br>  → Feature Engineering (Snowpark DataFrames / Dynamic Tables)<br>  → Model Training (Snowpark ML / Snowpark-Optimized Warehouse)<br>  → Model Registry (Snowflake Model Registry)<br>  → Inference (Snowpark UDF or Snowpark Container Services endpoint)<br>  → Predictions back into Snowflake tables<br>  → Monitoring (Data Quality Metrics + Cortex)</pre><p>Key components:</p><ul><li><strong>Snowpark Python</strong> for feature engineering without data movement.</li><li><strong>Snowpark-Optimized Warehouses</strong> for memory-intensive training.</li><li><strong>Model Registry</strong> — versioned model storage with metadata; supports rollback.</li><li><strong>SPCS</strong> — for serving models requiring custom containers (e.g., custom PyTorch inference), with Tri-Secret Secure protection for block volumes now generally available.</li><li>Avoid exporting data to external ML platforms — governance, lineage, and security are all preserved inside Snowflake’s perimeter.</li></ul><p><strong>Q20. How do you architect a Retrieval-Augmented Generation (RAG) pipeline using Snowflake Cortex?</strong></p><p><em>Answer:</em></p><pre>Documents (PDFs, emails, contracts) → staged in Snowflake<br>  → Cortex Document AI (chunking + extraction) <br>  → Embedding via EMBED() function<br>  → Cortex Search (vector + keyword hybrid index)<br>  → User Query → Cortex Analyst / Cortex Agent<br>  → LLM answer with grounded citations (AI_COMPLETE)</pre><p>Key architect decisions:</p><ul><li>Keep all data inside Snowflake perimeter — no external vector DB needed (avoids data exfiltration risk).</li><li>Use <strong>Cortex Search</strong> for hybrid retrieval (semantic + keyword) rather than pure vector search — better precision on structured domains.</li><li>Use <strong>Semantic Views</strong> to provide the LLM with a governed, business-friendly data model rather than raw tables.</li><li>Enforce governance: the LLM only accesses data the querying user’s role can see — Cortex inherits Snowflake’s RBAC automatically.</li></ul><h3>🎯 Architect-Level Scenario Questions</h3><p><strong>Q21.</strong> <em>You’re migrating a 500TB Oracle data warehouse to Snowflake. How do you approach this?</em> → Discuss: SnowConvert AI for code migration, phase-by-phase cutover, validation via row count + aggregate checksums, warehouse sizing for migration load, coexistence period with dual-write, and Iceberg as the landing format for flexibility.</p><p><strong>Q22.</strong> <em>A business-critical query takes 45 minutes and involves 10 joins across 5 tables totaling 2TB. How do you optimize it as an architect?</em> → Discuss: Query Profile analysis (identify the bottleneck node), clustering keys on join/filter columns, materialized views or Dynamic Tables for pre-aggregation, QAS enablement, warehouse scale-up, and whether the query pattern justifies a pre-aggregated OBT.</p><p><strong>Q23.</strong> <em>How do you handle GDPR “right to be forgotten” in Snowflake given immutable micro-partitions?</em> → Discuss: Tokenization at ingestion (store only tokens, PII in external vault), soft deletes + view-based masking for near-term, periodic full table rebuilds (CTAS) to purge Time Travel history, and Fail-safe expiry (data fully gone after 7 days post Time Travel expiry).</p><p><strong>Q24.</strong> <em>Design a data mesh architecture using Snowflake.</em> → Discuss: Domain-owned databases with independent RBAC, Snowflake Data Sharing for cross-domain consumption (no copy, governed access), Marketplace listings for curated data products, centralized governance via Horizon (tags, policies applied at the Org level), and Semantic Views as the contractual interface layer between domains.</p><blockquote><strong><em>Interview Tip:</em></strong><em> At the architect level, interviewers reward candidates who articulate </em>why<em> they chose a design and what they traded off — not just </em>what<em> Snowflake features exist. Always anchor your answers in business outcomes: cost, SLA, compliance, and team operability.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a90209440fa0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ Building a Real-Time AQI Dashboard in Snowflake using Streamlit]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://soumakdas94.medium.com/building-a-real-time-aqi-dashboard-in-snowflake-using-streamlit-4602764db30d?source=rss-6f4539057cdd------2"><img src="https://cdn-images-1.medium.com/max/1986/1*OLjxklgLXky7ffG8VwUPfg.png" width="1986"></a></p><p class="medium-feed-snippet">Overview</p><p class="medium-feed-link"><a href="https://soumakdas94.medium.com/building-a-real-time-aqi-dashboard-in-snowflake-using-streamlit-4602764db30d?source=rss-6f4539057cdd------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://soumakdas94.medium.com/building-a-real-time-aqi-dashboard-in-snowflake-using-streamlit-4602764db30d?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/4602764db30d</guid>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <category><![CDATA[snowflake-cortex]]></category>
            <category><![CDATA[streamlit-app]]></category>
            <category><![CDATA[snowflake]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Wed, 29 Apr 2026 07:56:18 GMT</pubDate>
            <atom:updated>2026-04-29T07:56:18.001Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Snowflake advanced interview question — part 2]]></title>
            <link>https://soumakdas94.medium.com/snowflake-advanced-interview-question-part-2-fcdc1bbc048b?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/fcdc1bbc048b</guid>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <category><![CDATA[advanced]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[interview]]></category>
            <category><![CDATA[snowflake]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Mon, 26 Aug 2024 08:59:05 GMT</pubDate>
            <atom:updated>2024-08-27T17:45:02.432Z</atom:updated>
            <content:encoded><![CDATA[<h3>Snowflake advanced interview question — part 2</h3><ol><li>How can you extract the fourth character from each value in a specific column that contains city names such as Chennai, Kolkata, and Mumbai?<br> — Answer: You can use the `SUBSTR` function in Snowflake to extract the fourth character. For example:<br>SELECT SUBSTR(col_a, 4, 1) AS fourth_char FROM your_table;</li><li>What is the process to recover accidentally deleted records from a Snowflake table?<br> — Answer: You can recover deleted records using Time Travel in Snowflake. For instance, if you want to restore the table to its state 5 minutes ago:<br> ```sql<br> CREATE TABLE your_table_restored AS<br> SELECT *<br> FROM your_table AT (OFFSET =&gt; -5 * 60);<br> ```</li><li>Explain how to create a table in Snowflake and load data from an external file.<br> — Answer: <br> — Creating the table:<br> ```sql<br> CREATE TABLE your_table (<br> column1 DATA_TYPE,<br> column2 DATA_TYPE,<br> …<br> );<br> ```<br> — Loading data:<br> ```sql<br> COPY INTO your_table<br> FROM <a href="http://twitter.com/your_stage">@your_stage</a><br> FILE_FORMAT = (FORMAT_NAME = ‘your_file_format’);<br> ```</li><li>What are the benefits of using SnowSQL in Snowflake, and when is it most useful?<br> — Answer: SnowSQL is a command-line tool ideal for automating tasks, bulk loading/unloading data, and scripting. It’s beneficial for batch processing and managing Snowflake environments programmatically.<br> 5. What should be considered when loading files into Snowflake using Snowpipe? Is there a file size limit?<br> — Answer: While Snowpipe can handle large files, splitting files into chunks of around 100–250 MB is recommended for better performance. Snowpipe can handle larger files, but processing time may increase with file size.<br> 6. How can you manage double quotes in file formats when loading data into Snowflake?<br> — Answer: You can configure the file format to escape or remove double quotes using the `FIELD_OPTIONALLY_ENCLOSED_BY` parameter in the file format definition:<br> ```sql<br> CREATE FILE FORMAT my_format<br> TYPE = ‘CSV’<br> FIELD_OPTIONALLY_ENCLOSED_BY = ‘“‘;<br> ```<br> 7. Is it possible to perform DML operations on a clone table in Snowflake, and how does it affect storage?<br> — Answer: Yes, you can perform DML operations on a clone table. Only the changes (i.e., new or modified data) consume additional storage, as clones are metadata pointers to the original data.<br> 8. Describe the process of database cloning in Snowflake and whether it includes staging data.<br> — Answer: Database cloning in Snowflake creates a copy of a database almost instantly without duplicating the data. However, cloning does not include data in external stages or other external objects.<br> ```sql<br> CREATE DATABASE new_db CLONE existing_db;<br> ```<br> 9. What datatype should be used to handle JSON and XML data in Snowflake?<br> — Answer: Use the `VARIANT` datatype to handle semi-structured data like JSON and XML in Snowflake.<br> 10. What is the maximum size of the VARIANT datatype in Snowflake?<br> — Answer: The maximum size of the VARIANT datatype in Snowflake is 16 MB.<br> 11. How can you optimize a query that uses a clustering key but is still slow due to large data volumes?<br> — Answer: You may need to adjust the clustering by redefining the key to better match the query pattern, increasing the compute power, or using materialized views. Additionally, you could review the query execution plan and adjust the workload management settings.<br> 12. What happens if a file with the same name is reloaded into an S3 bucket linked to Snowflake?<br> — Answer: If the file is reloaded without being deleted first, Snowflake’s `COPY` command will skip the file if it detects a duplicate based on the `MD5` hash, preventing duplicate data from being loaded.<br> 13. If a file is deleted and then reloaded to S3, will the same data get reloaded into Snowflake?<br> — Answer: Yes, if the file is deleted and then reloaded, Snowflake will treat it as a new file and load the data again, potentially leading to duplicates unless deduplication measures are taken.<br> 14. How do you remove duplicates while retaining one unique record in a Snowflake table?<br> — Answer: You can use a CTE with the `ROW_NUMBER()` function to identify and delete duplicates:<br> ```sql<br> WITH cte AS (<br> SELECT *, ROW_NUMBER() OVER (PARTITION BY unique_column ORDER BY load_time DESC) AS rn<br> FROM your_table<br> )<br> DELETE FROM your_table<br> WHERE rn &gt; 1;<br> ```<br> 15. What is the cost implication of cloning in Snowflake?<br> — Answer: Cloning is cost-effective because it doesn’t duplicate storage for unchanged data. However, changes made to the clone will incur storage costs.<br> 16. What are some common errors encountered when loading data into Snowflake, and how can they be resolved?<br> — Answer: Common errors include file format mismatches, permission issues, and network interruptions. These can be resolved by checking file formats, ensuring appropriate permissions, and retrying the load operation.<br> 17. How do you monitor long-running queries or steps in Snowflake?<br> — Answer: Use the `QUERY_HISTORY` or `WAREHOUSE_LOAD_HISTORY` views to monitor long-running queries and identify bottlenecks.<br> 18. What are the limitations of using SQL within Snowflake’s JavaScript stored procedures?<br> — Answer: Snowflake JavaScript stored procedures may have limitations like no support for certain SQL features (e.g., window functions), lack of transaction control, and limited error handling.<br> 19. How can you efficiently load large volumes of on-premise data into Snowflake?<br> — Answer: Efficiently loading large data volumes can be achieved using Snowpipe for continuous loading, breaking data into smaller chunks, and utilizing the bulk `COPY INTO` command.<br> 20. How does Python integration work in Snowflake, and what are its use cases?<br> — Answer: Python integration in Snowflake, available via Snowpark, allows you to write Python code for data processing directly in Snowflake. It’s ideal for complex data transformations, machine learning, and advanced analytics.<br> 21. What is the process for changing a Snowflake warehouse size from small to medium?<br> — Answer: You can change the warehouse size using the following command:<br> ```SQL<br> ALTER WAREHOUSE your_warehouse_name SET WAREHOUSE_SIZE = ‘MEDIUM’;<br> ```<br> 22. What happens to compute resources in Snowflake when they are no longer in use?<br> — Answer: Snowflake automatically suspends compute resources when they are idle, helping to optimize cost by only charging for active usage.<br> 23. What are the restrictions in a Snowflake Reader account?<br> — Answer: A Snowflake Reader account has limited capabilities, such as restricted DML operations, inability to create warehouses, and limited administrative functions.<br>Sure! Here are the remaining questions with answers:<br> 24. How do you change the data type of a column in a Snowflake table?<br> — Answer: You can change the data type of a column using the `ALTER TABLE` command with the `MODIFY COLUMN` clause. However, Snowflake has limitations, such as not allowing changes to certain data types (e.g., converting `VARCHAR` to `NUMBER` directly). Here’s an example of changing a `VARCHAR` column to a larger size:<br> ```SQL<br> ALTER TABLE your_table<br> MODIFY COLUMN column_name VARCHAR(255);<br> ```<br> 25. What are Snowflake credits, and how are they calculated for each warehouse size?<br> — Answer: Snowflake credits are the currency used to pay for computing usage. Credits are calculated based on the size of the virtual warehouse and the time it is running. For example, a `SMALL` warehouse consumes 1 credit per hour, while a `LARGE` warehouse consumes 4 credits per hour.<br> 26. How do Snowflake credits differ from Data Warehouse credits?<br> — Answer: Snowflake credits are a broader term encompassing compute and storage costs. Data Warehouse credits specifically refer to the compute resources consumed by virtual warehouses. These credits are used when you run queries or perform operations in Snowflake.<br> 27. What are the different types of credits in Snowflake, and what are their purposes?<br> — Answer: <br> — Compute Credits: Used for running queries, loading data, and other compute-intensive operations.<br> — Storage Credits: Used for storing data in Snowflake.<br> — Cloud Services Credits: Used for metadata operations, optimizing query performance, and managing Snowflake’s infrastructure.<br> 28. How can you handle errors encountered when loading files into Snowflake tables?<br> — Answer: Errors can be handled using error handling options in the `COPY INTO` command, such as `ON_ERROR`, which can be set to `CONTINUE`, `SKIP_FILE`, `ABORT_STATEMENT`, etc. Example:<br> ```SQL<br> COPY INTO your_table<br> FROM <a href="http://twitter.com/your_stage">@your_stage</a><br> FILE_FORMAT = (FORMAT_NAME = ‘your_file_format’)<br> ON_ERROR = ‘SKIP_FILE’;<br> ```<br> 29. How do you eliminate entire row duplicates in a flat file before loading it into Snowflake?<br> — Answer: You can use an intermediate staging table to load the data, followed by a `DELETE` operation using a CTE or the `ROW_NUMBER()` function to remove duplicates. Example:<br> ```SQL<br> WITH cte AS (<br> SELECT *, ROW_NUMBER() OVER (PARTITION BY * ORDER BY load_time DESC) AS rn<br> FROM staging_table<br> )<br> DELETE FROM staging_table<br> WHERE rn &gt; 1;<br> ```<br> 30. How do you retain one unique record while deleting duplicates from a Snowflake table?<br> — Answer: You can use the `ROW_NUMBER()` function to assign a unique rank to each row and then delete the duplicates:<br> ```SQL<br> WITH cte AS (<br> SELECT *, ROW_NUMBER() OVER (PARTITION BY unique_column ORDER BY load_time DESC) AS rn<br> FROM your_table<br> )<br> DELETE FROM your_table<br> WHERE rn &gt; 1;<br> ```<br> 31. How do you check for long-running steps in SQL within Snowflake?<br> — Answer: You can monitor long-running queries using the `QUERY_HISTORY` or `WAREHOUSE_LOAD_HISTORY` views. These views provide insights into query performance and can help identify bottlenecks.<br> 32. What are the limitations of using SQL within Snowflake’s JavaScript stored procedures?<br> — Answer: Limitations include the inability to execute some SQL commands, limited transaction control, and the lack of support for certain SQL features, such as window functions. Error handling in JavaScript is also different from SQL, requiring careful coding practices.<br> 33. How can you integrate and share data between Unix and Snowflake using SnowSQL on Unix?<br> — Answer: You can use SnowSQL on Unix to execute SQL commands, load data, and export data from Snowflake. Integration typically involves setting up secure connections, configuring stages for data sharing, and using the `PUT` and `GET` commands to move data between Unix and Snowflake.<br> 34. What is the best approach to load a large volume of data from on-premises to Snowflake?<br> — Answer: The best approach involves using Snowpipe for continuous data loading or the `COPY INTO` command for batch loading. Data can be transferred to Snowflake via external stages (e.g., Amazon S3) before loading it into Snowflake. Breaking the data into smaller chunks and parallelizing the load can also improve performance.<br> 35. How do you load data into an external stage in S3 for Snowflake?<br> — Answer: To load data into an external stage in S3, you first create the external stage, then upload the files to S3, and finally load the data into Snowflake using the `COPY INTO` command. Example:<br> ```SQL<br> CREATE OR REPLACE STAGE my_s3_stage<br> URL = ‘s3://mybucket’<br> CREDENTIALS = (AWS_KEY_ID=’your_aws_key’ AWS_SECRET_KEY=’your_secret_key’);<br>COPY INTO your_table<br> FROM <a href="http://twitter.com/my_s3_stage">@my_s3_stage</a><br> FILE_FORMAT = (FORMAT_NAME = ‘your_file_format’);<br> ```<br> 36. How does Python work in Snowflake, and what are its use cases?<br> — Answer: Python in Snowflake, facilitated by Snowpark, allows developers to write Python code directly within Snowflake for data processing and analytics. Use cases include complex transformations, data engineering tasks, and integrating machine learning models into your Snowflake workflows.<br> 37. Does Snowflake support Lambda functions, and if so, how are they used?<br> — Answer: Snowflake doesn’t natively support AWS Lambda functions, but you can invoke Lambda functions externally to trigger Snowflake tasks or process Snowflake data by integrating with AWS services such as SNS or S3.<br> 38. What are the benefits of using an ETL tool with Snowflake?<br> — Answer: Using an ETL tool with Snowflake helps streamline data ingestion, transformation, and loading processes. ETL tools can optimize data pipelines, support complex transformations, and provide robust scheduling and error handling mechanisms, making them ideal for managing large-scale data workflows in Snowflake.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fcdc1bbc048b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Azure Databricks vs Azure Synapse]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">Introduction</p><p class="medium-feed-link"><a href="https://soumakdas94.medium.com/azure-databricks-vs-azure-synapse-a7f9eab98f4e?source=rss-6f4539057cdd------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://soumakdas94.medium.com/azure-databricks-vs-azure-synapse-a7f9eab98f4e?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/a7f9eab98f4e</guid>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[azure]]></category>
            <category><![CDATA[difference]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[synapse]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Fri, 22 Sep 2023 05:58:49 GMT</pubDate>
            <atom:updated>2023-09-22T05:58:49.082Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Unlocking the Power of Data with Snowflake Schema Detection and Schema Evolution]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">At Snowflake, we understand the importance of simplifying the process of working with data. One of Snowflake&#x2019;s highly compelling features&#x2026;</p><p class="medium-feed-link"><a href="https://soumakdas94.medium.com/unlocking-the-power-of-data-with-snowflake-schema-detection-and-schema-evolution-484ea3f863c3?source=rss-6f4539057cdd------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://soumakdas94.medium.com/unlocking-the-power-of-data-with-snowflake-schema-detection-and-schema-evolution-484ea3f863c3?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/484ea3f863c3</guid>
            <category><![CDATA[schema-design]]></category>
            <category><![CDATA[snowflake-computing]]></category>
            <category><![CDATA[snowflake]]></category>
            <category><![CDATA[data-ingestion]]></category>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Mon, 18 Sep 2023 13:26:07 GMT</pubDate>
            <atom:updated>2023-09-18T13:27:04.180Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Snowflake Snowpark: Empowering Data Processing and Analytics]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">Abstract</p><p class="medium-feed-link"><a href="https://soumakdas94.medium.com/snowflake-snowpark-empowering-data-processing-and-analytics-edd7e4f508d3?source=rss-6f4539057cdd------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://soumakdas94.medium.com/snowflake-snowpark-empowering-data-processing-and-analytics-edd7e4f508d3?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/edd7e4f508d3</guid>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <category><![CDATA[snowpark-python]]></category>
            <category><![CDATA[snowpark]]></category>
            <category><![CDATA[snowflake-snowpark]]></category>
            <category><![CDATA[snowflake]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Mon, 18 Sep 2023 10:01:32 GMT</pubDate>
            <atom:updated>2023-09-18T10:01:32.561Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Snowflake SnowPark API — An Introduction]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">The Snowpark library offers a user-friendly toolkit for querying and processing data at scale within the Snowflake platform. With Snowpark&#x2026;</p><p class="medium-feed-link"><a href="https://soumakdas94.medium.com/snowflake-snowpark-api-9f2a52ac6ff0?source=rss-6f4539057cdd------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://soumakdas94.medium.com/snowflake-snowpark-api-9f2a52ac6ff0?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f2a52ac6ff0</guid>
            <category><![CDATA[snowflake-computing]]></category>
            <category><![CDATA[snowpark]]></category>
            <category><![CDATA[snowflake]]></category>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <category><![CDATA[snowpark-python]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Mon, 18 Sep 2023 09:40:56 GMT</pubDate>
            <atom:updated>2023-09-20T07:04:01.153Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Snowflake Sept 2023 Releases]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">SQL Updates Introduction of New Function:</p><p class="medium-feed-link"><a href="https://soumakdas94.medium.com/snowflake-sept-releases-fa29310ec4d?source=rss-6f4539057cdd------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://soumakdas94.medium.com/snowflake-sept-releases-fa29310ec4d?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa29310ec4d</guid>
            <category><![CDATA[new-features]]></category>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <category><![CDATA[snowflake]]></category>
            <category><![CDATA[snowflake-computing]]></category>
            <category><![CDATA[snowflakedb]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Mon, 18 Sep 2023 09:32:35 GMT</pubDate>
            <atom:updated>2023-09-18T09:33:01.154Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Data Governance in Snowflake]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">General availability of the Data Governance interface in Snowflake Snowsight</p><p class="medium-feed-link"><a href="https://soumakdas94.medium.com/data-governance-in-snowflake-a2ee0986dbfc?source=rss-6f4539057cdd------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://soumakdas94.medium.com/data-governance-in-snowflake-a2ee0986dbfc?source=rss-6f4539057cdd------2</link>
            <guid isPermaLink="false">https://medium.com/p/a2ee0986dbfc</guid>
            <category><![CDATA[snowflake]]></category>
            <category><![CDATA[snowflake-computing]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[data-governance-service]]></category>
            <category><![CDATA[snowflake-data-cloud]]></category>
            <dc:creator><![CDATA[Soumak Das]]></dc:creator>
            <pubDate>Thu, 14 Sep 2023 14:20:12 GMT</pubDate>
            <atom:updated>2023-09-14T14:42:47.579Z</atom:updated>
        </item>
    </channel>
</rss>