Serverless SQL Best Practices¶
Home | Best Practices | Serverless SQL
Best practices for Azure Synapse Serverless SQL pools.
Overview¶
Serverless SQL is ideal for:
- Ad-hoc data exploration
- Power BI DirectQuery
- Data lake querying
- ETL transformations
For detailed guidance, see: Full Serverless SQL Guide
Query Optimization¶
Use Proper File Formats¶
| Format | Query Performance | Compression | Schema Evolution |
|---|---|---|---|
| Parquet | Excellent | High | Good |
| Delta | Excellent | High | Excellent |
| CSV | Poor | Low | None |
| JSON | Poor | Low | Flexible |
-- Optimal: Query Parquet with schema
SELECT *
FROM OPENROWSET(
BULK 'https://storage.dfs.core.windows.net/data/*.parquet',
FORMAT = 'PARQUET'
) WITH (
customer_id INT,
order_date DATE,
amount DECIMAL(18,2)
) AS orders
WHERE order_date >= '2024-01-01';
Partition Pruning¶
-- Enable partition pruning with filepath
SELECT
filepath(1) AS year,
filepath(2) AS month,
COUNT(*) AS record_count
FROM OPENROWSET(
BULK 'https://storage.dfs.core.windows.net/data/year=*/month=*/*.parquet',
FORMAT = 'PARQUET'
) AS sales
WHERE filepath(1) = '2024' AND filepath(2) = '01'
GROUP BY filepath(1), filepath(2);
Create Views¶
-- Create optimized view
CREATE VIEW gold.vw_sales AS
SELECT
customer_id,
order_date,
product_id,
amount
FROM OPENROWSET(
BULK 'https://storage.dfs.core.windows.net/gold/sales/**/*.parquet',
FORMAT = 'PARQUET'
) WITH (
customer_id INT,
order_date DATE,
product_id INT,
amount DECIMAL(18,2)
) AS sales;
Cost Optimization¶
Data Processed Estimation¶
-- Check estimated data processed
SELECT
execution_type,
total_worker_time / 1000000.0 AS cpu_seconds,
total_physical_reads * 8 / 1024.0 AS data_read_mb
FROM sys.dm_exec_query_stats
ORDER BY total_physical_reads DESC;
Best Practices¶
- Use Parquet/Delta - 10-100x less data scanned
- Partition data - Enable partition elimination
- Project only needed columns - Columnar storage benefits
- Filter early - Reduce data processed
Related Documentation¶
- Serverless SQL Architecture
- SQL Performance
Last Updated: January 2025