snowflake-development – skill Dữ liệu & Khoa học cho agent AI

Snowflake Development

You are a Snowflake development expert. Apply these rules when writing SQL, building data pipelines, using Cortex AI, or working with Snowpark Python on Snowflake.

When to Use

• When the user asks for help with Snowflake SQL, data pipelines, Cortex AI, or Snowpark Python.
• When you need Snowflake-specific guidance for dbt, performance tuning, or security hardening.

SQL Best Practices

Naming and Style

• Use snake_case for all identifiers. Avoid double-quoted identifiers — they create case-sensitive names requiring constant quoting.
• Use CTEs (WITH clauses) over nested subqueries.
• Use CREATE OR REPLACE for idempotent DDL.
• Use explicit column lists — never SELECT * in production (Snowflake's columnar storage scans only referenced columns).

Stored Procedures — Colon Prefix Rule

In SQL stored procedures (BEGIN...END blocks), variables and parameters must use the colon : prefix inside SQL statements. Without it, Snowflake raises "invalid identifier" errors.

BAD:

CREATE PROCEDURE my_proc(p_id INT) RETURNS STRING LANGUAGE SQL AS
BEGIN
    LET result STRING;
    SELECT name INTO result FROM users WHERE id = p_id;
    RETURN result;
END;

GOOD:

CREATE PROCEDURE my_proc(p_id INT) RETURNS STRING LANGUAGE SQL AS
BEGIN
    LET result STRING;
    SELECT name INTO :result FROM users WHERE id = :p_id;
    RETURN result;
END;

Semi-Structured Data

• VARIANT, OBJECT, ARRAY for JSON/Avro/Parquet/ORC.
• Access nested fields: src:customer.name::STRING. Always cast: src:price::NUMBER(10,2).
• VARIANT null vs SQL NULL: JSON null is a distinct VARIANT null value, not the string "null" or SQL NULL. Use IS_NULL_VALUE to distinguish it. The JSON load option is STRIP_NULL_VALUES = TRUE; removing null-valued elements changes data semantics and must be intentional.
• Flatten arrays: SELECT f.value:name::STRING FROM my_table, LATERAL FLATTEN(input => src:items) f;

MERGE for Upserts

MERGE INTO target t USING source s ON t.id = s.id
WHEN MATCHED THEN UPDATE SET t.name = s.name, t.updated_at = CURRENT_TIMESTAMP()
WHEN NOT MATCHED THEN INSERT (id, name, updated_at) VALUES (s.id, s.name, CURRENT_TIMESTAMP());

Data Pipelines

Choosing Your Approach

| Approach | When to Use |

|----------|-------------|

| Dynamic Tables | Declarative transformations. Default choice. Define the query, Snowflake handles refresh. |

| Streams + Tasks | Imperative CDC. Use for procedural logic, stored procedure calls. |

| Snowpipe | Continuous file loading from S3/GCS/Azure. |

Dynamic Tables

CREATE OR REPLACE DYNAMIC TABLE cleaned_events
    TARGET_LAG = '5 minutes'
    WAREHOUSE = transform_wh
    AS
    SELECT event_id, event_type, user_id, event_timestamp
    FROM raw_events
    WHERE event_type IS NOT NULL;

Key rules:

• Set TARGET_LAG progressively: tighter at top, looser at bottom.
• Incremental DTs cannot depend on Full refresh DTs.
• SELECT * breaks on schema changes — use explicit column lists.
• Change tracking must stay enabled on base tables.
• Views cannot sit between two Dynamic Tables.

Streams and Tasks

CREATE OR REPLACE STREAM raw_stream ON TABLE raw_events;

CREATE OR REPLACE TASK process_events
    WAREHOUSE = transform_wh
    SCHEDULE = 'USING CRON 0 */1 * * * America/Los_Angeles'
    WHEN SYSTEM$STREAM_HAS_DATA('raw_stream')
    AS INSERT INTO cleaned_events SELECT ... FROM raw_stream;

-- Tasks start SUSPENDED — you MUST resume them
ALTER TASK process_events RESUME;

Cortex AI

Function Reference

| Function | Purpose |

|----------|---------|

| AI_COMPLETE | LLM completion (text, images, documents) |

| AI_CLASSIFY | Classify into categories (up to 500 labels) |

| AI_FILTER | Boolean filter on text/images |

| AI_EXTRACT | Structured extraction from text/images/documents |

| AI_SENTIMENT | Sentiment score (-1 to 1) |

| AI_PARSE_DOCUMENT | OCR or layout extraction |

| AI_REDACT | PII removal |

Deprecated (do NOT use): COMPLETE, CLASSIFY_TEXT, EXTRACT_ANSWER, PARSE_DOCUMENT, SUMMARIZE, TRANSLATE, SENTIMENT, EMBED_TEXT_768.

TO_FILE — Common Error Source

Stage path and filename are SEPARATE arguments:

-- BAD: TO_FILE('@stage/file.pdf')
-- GOOD:
TO_FILE('@db.schema.mystage', 'invoice.pdf')

Use AI_CLASSIFY for Classification (Not AI_COMPLETE)

SELECT AI_CLASSIFY(ticket_text,
    ['billing', 'technical', 'account']):labels[0]::VARCHAR AS category
FROM tickets;

Cortex Agents

CREATE OR REPLACE AGENT my_db.my_schema.sales_agent
FROM SPECIFICATION $spec$
{
    "models": {"orchestration": "auto"},
    "instructions": {
        "orchestration": "You are SalesBot...",
        "response": "Be concise."
    },
    "tools": [{"tool_spec": {"type": "cortex_analyst_text_to_sql", "name": "Sales", "description": "Queries sales..."}}],
    "tool_resources": {"Sales": {"semantic_model_file": "@stage/model.yaml"}}
}
$spec$;

Agent rules:

• Use $spec$ delimiter (not $$).
• models must be an object, not an array.
• tool_resources is a separate top-level object, not nested inside tools.
• Do NOT include empty/null values in edit specs — clears existing values.
• Tool descriptions are the #1 quality factor.
• Never modify production agents directly — clone first.

Snowpark Python

from snowflake.snowpark import Session
import os

session = Session.builder.configs({
    "account": os.environ["SNOWFLAKE_ACCOUNT"],
    "user": os.environ["SNOWFLAKE_USER"],
    "password": os.environ["SNOWFLAKE_PASSWORD"],
    "role": "my_role", "warehouse": "my_wh",
    "database": "my_db", "schema": "my_schema"
}).create()

• Never hardcode credentials.
• DataFrames are lazy — executed on collect()/show().
• Do NOT use collect() on large DataFrames — process server-side.
• Use vectorized UDFs (10-100x faster) for batch/ML workloads instead of scalar UDFs.

dbt on Snowflake

Dynamic table materialization (streaming/near-real-time marts):

{{ config(materialized='dynamic_table', snowflake_warehouse='transforming', target_lag='1 hour') }}

Incremental materialization (large fact tables):

{{ config(materialized='incremental', unique_key='event_id') }}

Snowflake-specific configs (combine with any materialization):

{{ config(transient=true, copy_grants=true, query_tag='team_daily') }}

• Do NOT use {{ this }} without {% if is_incremental() %} guard.
• Use dynamic_table materialization for streaming/near-real-time marts.

Performance

• Cluster keys: Only multi-TB tables, on WHERE/JOIN/GROUP BY columns.
• Search Optimization: ALTER TABLE t ADD SEARCH OPTIMIZATION ON EQUALITY(col);
• Warehouse sizing: Start X-Small, scale up. AUTO_SUSPEND = 60, AUTO_RESUME = TRUE.
• Separate warehouses per workload.
• Estimate AI costs first: SELECT SUM(AI_COUNT_TOKENS('claude-4-sonnet', text)) FROM table;

Security

• Follow least-privilege RBAC. Use database roles for object-level grants.
• Audit ACCOUNTADMIN regularly: SHOW GRANTS OF ROLE ACCOUNTADMIN;
• Use network policies for IP allowlisting.
• Use masking policies for PII columns and row access policies for multi-tenant isolation.

Common Error Patterns

| Error | Cause | Fix |

|-------|-------|-----|

| "Object does not exist" | Wrong context or missing grants | Fully qualify names, check grants |

| "Invalid identifier" in proc | Missing colon prefix | Use :variable_name |

| "Numeric value not recognized" | VARIANT not cast | src:field::NUMBER(10,2) |

| Task not running | Forgot to resume | ALTER TASK ... RESUME |

| DT refresh failing | Schema change or tracking disabled | Use explicit columns, check change tracking |

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.

CREATE OR REPLACE can replace existing objects and their data/configuration; it is not a harmless repeatable migration. Review the exact target, preserve grants/backups as needed and use a planned migration. Function support, deprecation and dynamic-table restrictions depend on the current account/version; verify the cited function in official Snowflake documentation rather than treating broad deprecated-function lists as authority.