When developers migrate from PostgreSQL to DuckDB, the first thing they usually look for is “materialized views.” PostgreSQL’s MATERIALIZED VIEW syntax is clean, but in practice it comes with several pain points—schema changes require DROP and rebuild, refresh operations lock tables and hurt concurrency, and it doesn’t support columnar storage optimization.
DuckDB takes a different design philosophy: it doesn’t need traditional materialized views. Because DuckDB natively supports columnar storage (Parquet), the combination of CREATE TABLE AS SELECT (CTAS) + Parquet direct reads achieves a more flexible and efficient materialization scheme than traditional MVs.
Today, I’ll walk you through building a production-grade, high-performance query backend—from zero to one—using pure SQL + Python.
Why DuckDB Doesn’t Need Traditional Materialized Views
To understand this, let’s look at three core weaknesses of traditional materialized views:
1. Schema changes are expensive
In PostgreSQL, once an MV is created, changing a field requires DROP and rebuild—queries are unavailable during this process. In data product scenarios, customer requirements change constantly, and MV’s rigid structure slows down iteration.
2. Table locks during refresh When refreshing an MV, the database locks the underlying table. In high-concurrency scenarios, this becomes a bottleneck. Imagine 50 users checking your BI dashboard simultaneously while a refresh operation blocks the entire system.
3. No columnar advantages Traditional RDBMS stores data row-wise. Even with MVs, queries require full row scans. Parquet is a columnar compressed format—you only read the columns you need, achieving 5-10x performance differences.
DuckDB’s approach is simpler and more direct: store aggregation results directly as Parquet files. Reading Parquet files is 5-10x faster than reading row-based tables. Combined with CTAS pre-aggregation, query latency drops from seconds to milliseconds.
Complete Walkthrough: From Raw Data to Product in 4 Steps
Step 1: Export Data to Parquet (Zero ETL)
This is the most critical step—converting raw data to Parquet format. All subsequent queries are based on this file.
import duckdb
import time
conn = duckdb.connect(':memory:')
# Simulate e-commerce order data (1M rows)
conn.execute("""
CREATE TABLE orders AS
SELECT
gen AS order_id,
CASE (gen % 5)
WHEN 0 THEN 'Electronics'
WHEN 1 THEN 'Clothing'
WHEN 2 THEN 'Books'
WHEN 3 THEN 'Home'
ELSE 'Sports'
END AS product_category,
CAST(random() * 500 + 10 AS DOUBLE) AS amount,
TIMESTAMP '2026-01-01' + (gen % 270) * INTERVAL '1 day' +
(gen % 24) * INTERVAL '1 hour' +
(gen % 60) * INTERVAL '1 minute' AS created_at
FROM generate_series(1, 1000000) AS t(gen)
""")
# Write to parquet — this is the key!
parquet_path = '/tmp/orders.parquet'
conn.execute(f"COPY orders TO '{parquet_path}' (FORMAT PARQUET)")
conn.close()
print(f"Parquet file created: {parquet_path}")
Benefit analysis: Parquet is a columnar compressed format. 1M order rows compress from 80MB to 12MB. More importantly, subsequent queries let DuckDB read only the needed columns and skip irrelevant bytes—something row storage cannot do.
Step 2: CTAS Pre-Aggregation (MV Replacement)
With the Parquet file ready, create a pre-aggregation table. This step completely replaces traditional materialized view functionality.
conn = duckdb.connect(':memory:')
# One-time creation of materialized table
conn.execute("""
CREATE TABLE IF NOT EXISTS mv_daily_stats AS
SELECT
DATE_TRUNC('day', created_at) AS dt,
product_category,
COUNT(*) AS order_count,
ROUND(SUM(amount), 2) AS total_revenue,
ROUND(AVG(amount), 2) AS avg_order_value,
COUNT(DISTINCT order_id) AS unique_customers
FROM read_parquet('/tmp/orders.parquet')
WHERE created_at >= CURRENT_DATE - INTERVAL '90' DAY
GROUP BY 1, 2
""")
# Add index to accelerate filtered queries
conn.execute("CREATE INDEX IF NOT EXISTS idx_mv_dt ON mv_daily_stats(dt)")
conn.close()
Key design points:
WHERE created_at >= CURRENT_DATE - INTERVAL '90' DAYcontrols data range, preventing unbounded table growthCREATE INDEXaccelerates high-frequency filtered queries- Table name prefixed with
mv_for easy identification and management
Step 3: Performance Benchmark (Real Results)
Run the same query across three approaches:
import time
conn = duckdb.connect(':memory:')
# Warm up: ensure Parquet file is cached in memory
conn.execute("SELECT COUNT(*) FROM read_parquet('/tmp/orders.parquet')")
# Test 1: Parquet direct read aggregation
start = time.time()
for _ in range(10):
conn.execute("""
SELECT
DATE_TRUNC('day', created_at) AS dt,
product_category,
COUNT(*) AS order_count,
ROUND(SUM(amount), 2) AS total_revenue
FROM read_parquet('/tmp/orders.parquet')
WHERE created_at >= CURRENT_DATE - INTERVAL '90' DAY
GROUP BY 1, 2
""").fetchall()
parquet_time = (time.time() - start) / 10
# Test 2: CTAS materialized table query
conn.execute("""
CREATE TABLE IF NOT EXISTS mv_daily_stats AS
SELECT
DATE_TRUNC('day', created_at) AS dt,
product_category,
COUNT(*) AS order_count,
ROUND(SUM(amount), 2) AS total_revenue
FROM read_parquet('/tmp/orders.parquet')
WHERE created_at >= CURRENT_DATE - INTERVAL '90' DAY
GROUP BY 1, 2
""")
start = time.time()
for _ in range(10):
conn.execute("""
SELECT * FROM mv_daily_stats
WHERE dt >= CURRENT_DATE - INTERVAL '30' DAY
ORDER BY total_revenue DESC
""").fetchall()
ctas_time = (time.time() - start) / 10
# Test 3: CTAS + index query
conn.execute("CREATE INDEX idx_mv_dt ON mv_daily_stats(dt)")
start = time.time()
for _ in range(10):
conn.execute("""
SELECT * FROM mv_daily_stats
WHERE dt >= CURRENT_DATE - INTERVAL '30' DAY
ORDER BY total_revenue DESC
""").fetchall()
idx_time = (time.time() - start) / 10
print(f"Parquet direct read: {parquet_time:.4f}s")
print(f"CTAS materialized: {ctas_time:.4f}s ({parquet_time/ctas_time:.1f}x faster)")
print(f"CTAS + index: {idx_time:.4f}s ({parquet_time/idx_time:.1f}x faster)")
conn.close()
Results (DuckDB 1.5.5, 1M rows):
Parquet direct read: 0.0302s
CTAS materialized: 0.0033s (9.2x faster)
CTAS + index: 0.0027s (11.1x faster)
Interpretation: CTAS materialized table is 9x faster than Parquet direct read. With an index added, it’s even faster. At data volumes above 1M rows, the gap widens further.
Step 4: Encapsulate as a Data Product Backend
Wrap the logic into a reusable Python class—this is the critical step toward productization.
import duckdb
import json
from datetime import datetime
class DataProductBackend:
"""Data product backend built with DuckDB + CTAS, ~50 lines of code"""
def __init__(self, parquet_path):
self.conn = duckdb.connect(':memory:')
self.parquet_path = parquet_path
self._build()
def _build(self):
"""Build materialized table"""
self.conn.execute(f"""
CREATE TABLE IF NOT EXISTS mv_stats AS
SELECT
DATE_TRUNC('day', created_at) AS dt,
product_category,
COUNT(*) AS order_count,
ROUND(SUM(amount), 2) AS total_revenue,
ROUND(AVG(amount), 2) AS avg_order_value,
COUNT(DISTINCT order_id) AS unique_customers
FROM read_parquet('{self.parquet_path}')
WHERE created_at >= CURRENT_DATE - INTERVAL '90' DAY
GROUP BY 1, 2
""")
self.conn.execute("CREATE INDEX IF NOT EXISTS idx_dt ON mv_stats(dt)")
def refresh(self):
"""Refresh materialized table (rebuild)"""
self.conn.execute("DROP TABLE IF EXISTS mv_stats")
self._build()
print(f"[{datetime.now()}] Materialized table refreshed")
def top_categories(self, days=30, limit=10):
"""Query top N categories"""
return self.conn.execute(f"""
SELECT * FROM mv_stats
WHERE dt >= CURRENT_DATE - INTERVAL '{days}' DAY
ORDER BY total_revenue DESC LIMIT {limit}
""").fetchall()
def daily_report(self, days=7):
"""Generate daily report"""
return self.conn.execute(f"""
SELECT dt,
SUM(order_count) AS orders,
SUM(total_revenue) AS revenue,
AVG(avg_order_value) AS avg_value
FROM mv_stats
WHERE dt >= CURRENT_DATE - INTERVAL '{days}' DAY
GROUP BY 1
ORDER BY 1 DESC
LIMIT {days}
""").fetchall()
def close(self):
self.conn.close()
Usage example:
backend = DataProductBackend('/tmp/orders.parquet')
# Query top 5 categories
print("Top 5 Categories (Last 30 Days):")
for row in backend.top_categories(days=30, limit=5):
print(f" {row[0]} | {row[1]:12s} | Orders:{row[2]:5d} | Revenue:¥{row[3]:>12,.2f}")
# Generate daily report
print("\nLast 7 Days Report:")
for row in backend.daily_report(days=7):
print(f" {row[0]} | Orders:{row[1]:5d} | Revenue:¥{row[2]:>12,.2f}")
backend.close()
Three Refresh Strategies for Production
In production, source data keeps updating. How do you refresh the materialized table? Here are three strategies:
Strategy A: Scheduled Full Rebuild (Simplest, daily updates)
def daily_refresh(parquet_path):
"""Rebuild materialized table every day at midnight"""
conn = duckdb.connect(':memory:')
conn.execute(f"""
CREATE TABLE mv AS
SELECT
DATE_TRUNC('day', created_at) AS dt,
product_category,
COUNT(*) AS order_count,
SUM(amount) AS total_revenue
FROM read_parquet('{parquet_path}')
WHERE created_at >= CURRENT_DATE - INTERVAL '90' DAY
GROUP BY 1, 2
""")
return conn
Best for: Data < 5M rows, daily updates, relaxed real-time requirements. Rebuild typically takes < 1 second.
Strategy B: Incremental Merge (Hourly updates)
conn = duckdb.connect(':memory:')
conn.execute("""
CREATE TABLE mv_incremental AS
SELECT dt, product_category,
SUM(order_count) AS order_count,
SUM(total_revenue) AS total_revenue
FROM (
-- Historical data
SELECT DATE_TRUNC('day', created_at) AS dt, product_category,
COUNT(*) AS order_count, SUM(amount) AS total_revenue
FROM orders
WHERE created_at >= CURRENT_DATE - INTERVAL '90' DAY
GROUP BY 1, 2
UNION ALL
-- New data
SELECT DATE_TRUNC('day', created_at), product_category,
COUNT(*), SUM(amount)
FROM new_orders
GROUP BY 1, 2
) t
GROUP BY 1, 2
""")
Best for: Large data (> 5M rows), hourly updates needed. Note: incremental merge assumes data is append-only.
Strategy C: Partitioned Parquet + Multi-File Aggregation (TB-scale)
conn = duckdb.connect(':memory:')
# Read partitioned directory directly, DuckDB auto-detects partitions
conn.execute("""
CREATE TABLE mv_yearly AS
SELECT
DATE_TRUNC('month', created_at) AS month,
product_category,
COUNT(*) AS orders,
ROUND(SUM(amount), 2) AS revenue
FROM read_parquet('/data/orders/*.parquet')
GROUP BY 1, 2
""")
Best for: TB-scale data with time-partitioned storage. DuckDB’s partition pruning automatically skips unnecessary partition files.
Key Design Principles
Follow these principles in production to avoid common pitfalls:
1. Parquet is the core storage format Export all raw data to Parquet first. DuckDB reads it zero-copy, avoiding ORM/ETL layer overhead. Don’t import data into a relational database first— that adds unnecessary steps.
2. CTAS is the simplest materialization
No MV syntax needed—CREATE TABLE AS SELECT is the most朴素 materialized view. Its advantage is flexibility: you can change the SQL and rebuild anytime,不受 Schema 限制 (unrestricted by schema).
3. Add indexes only where needed Only index columns used in filters (e.g., date fields). DuckDB’s VACUUM auto-maintains indexes—no manual intervention needed.
4. Reuse connections
Maintain long-lived connections in Python. Avoid repeated connect() calls—initialization has overhead that adds up.
5. Keep data range bounded
Add WHERE created_at >= ... in CTAS to control data range and prevent unbounded table growth. 90 days is a proven rule of thumb, adjustable per business needs.
Comparison: Traditional MV vs DuckDB CTAS + Parquet
| Dimension | PostgreSQL MV | DuckDB CTAS + Parquet |
|---|---|---|
| Creation syntax | CREATE TABLE | CREATE TABLE AS SELECT |
| Refresh method | REFRESH MATERIALIZED VIEW | DROP + CREATE or incremental merge |
| Schema changes | Must DROP and rebuild | Change SQL and rebuild directly |
| Storage format | Row-wise | Columnar (Parquet) |
| Query performance | Moderate | High (columnar + vectorized) |
| Operational cost | High (locking, maintenance) | Low (no locking, no dependencies) |
Summary
DuckDB’s “materialization” philosophy is: Parquet stores data + CTAS stores results + indexes accelerate queries. This combination delivers 9x speedup on 1M rows, and the gap widens at larger scales.
The key insight: you don’t need any additional storage engine or scheduling tool. Pure SQL + Python gives you a production-grade data product backend. This is DuckDB’s core competitiveness compared to traditional approaches—simple to the extreme, powerful to the practical.
The complete code repository (including incremental refresh scripts, multi-file aggregation examples, and FastAPI integration examples) is published at duckdblab.org with more detailed steps and additional cases.
📖 Want to systematically learn more DuckDB实战 techniques? duckdblab.org has a complete tutorial series from beginner to advanced, covering CTAS materialization, Parquet optimization, and data product —continuously updated.
本文信息
| 项目 | 内容 |
|---|---|
| DuckDB 版本 | v1.5.x（部分功能基于 v2.0 Preview） |
| 最后验证 | 2026-09-12 |
| 测试环境 | Linux / x86_64 / 16GB RAM |
| 官方文档 | DuckDB Documentation |
| GitHub | pengzz9527/duckdb-blog |
如发现错误，欢迎通过 GitHub Issue 或邮件 contact@duckdblab.org 反馈。