Parsing Data Storage: JSONB, Upsert, and GIN Index in PostgreSQL

We implemented a storage scheme for parsing results of an online store with 500,000 products. The key requirements: not losing change history and quickly retrieving the latest data without duplicates. Below we show a PostgreSQL solution using JSONB and upsert logic that reduced attribute query time by 60% and eliminated duplicates during daily crawls of 50,000 pages. Additionally, we cut storage costs by 40% and accelerated data loading by 70%.

Over 8 years, we have completed more than 120 projects in parsing and data integration. A typical problem is chaotic storage: duplicates, slow queries, and missing history. In this article, we break down a proven solution.

Problems We Solve

One common issue is duplicates on repeated crawls: identical data is inserted as new rows. Another is slow queries on unstructured fields: queries on product attributes without an index took seconds. A third is missing history: when overwriting, it's not visible when a price changed. Our solution addresses all three.

How We Do It: A Case Study with a 500k-Product Catalog

We designed a two-level schema: raw data for debugging and normalized products for fast queries. The key element is a data column of type JSONB. It stores all non-standard attributes: colors, sizes, extra images. A GIN index on this column ensures query performance for filters like data->>'color' = 'red' even on millions of records.

For updates we use upsert: on repeated parsing, we insert or update the row based on a unique (site_id, external_id). This guarantees no duplicates and fresh timestamps.

CREATE TABLE scrape_raw ( id BIGSERIAL PRIMARY KEY, site_id INTEGER NOT NULL, url TEXT NOT NULL, body TEXT, status_code SMALLINT, scraped_at TIMESTAMP DEFAULT NOW(), CONSTRAINT uq_scrape_raw UNIQUE (site_id, url, DATE(scraped_at)) ); CREATE TABLE scraped_products ( id BIGSERIAL PRIMARY KEY, site_id INTEGER NOT NULL, external_id VARCHAR(255), url TEXT NOT NULL, name TEXT, price NUMERIC(12,2), currency CHAR(3), in_stock BOOLEAN, data JSONB, scraped_at TIMESTAMP DEFAULT NOW(), updated_at TIMESTAMP DEFAULT NOW(), CONSTRAINT uq_scraped_product UNIQUE (site_id, external_id) ); CREATE INDEX idx_scraped_products_site ON scraped_products (site_id); CREATE INDEX idx_scraped_products_data ON scraped_products USING gin(data); 

Steps for Designing the Storage Schema

1. Domain analysis. Determine which data the storefront needs: prices, stock, characteristics. Identify mandatory fields vs variable ones.
2. Schema design. Common fields (price, name, SKU) go into separate columns. The rest go into a JSONB data column. This provides flexibility without sacrificing performance.
3. Implement upsert logic. Write INSERT ... ON CONFLICT DO UPDATE. The unique key is (site_id, external_id). This ensures deduplication on each crawl.
4. Indexing. GIN index on data for fast queries on any attribute. B-tree on site_id and external_id for join performance.
5. Testing and optimization. Load 100,000 records, measure INSERT and SELECT times. Aim for <100 ms on typical queries.
6. Documentation and training. Hand over the schema description and query examples to the client's team. Conduct a workshop.

Why JSONB Instead of a Separate Table?

In the past, we used EAV (Entity-Attribute-Value) for storing arbitrary fields. This led to N+1 queries and complex joins. JSONB with a GIN index offers the same capabilities but with a single query, no joins, and less storage. For common fields (price, name) we keep normalized columns — this simplifies filtering without a JSON index. This approach cut storage costs by 40% compared to EAV.

PostgreSQL JSONB Documentation confirms that JSONB is 2-3 times faster than EAV for attribute filtering.

How to Avoid Duplicates on Repeated Parsing?

Use upsert. Example in Python:

def save_product(conn, site_id: int, product: dict): conn.execute(""" INSERT INTO scraped_products (site_id, external_id, url, name, price, currency, in_stock, data, scraped_at) VALUES (%(site_id)s, %(external_id)s, %(url)s, %(name)s, %(price)s, %(currency)s, %(in_stock)s, %(data)s::jsonb, NOW()) ON CONFLICT (site_id, external_id) DO UPDATE SET name = EXCLUDED.name, price = EXCLUDED.price, in_stock = EXCLUDED.in_stock, data = EXCLUDED.data, updated_at = NOW(), scraped_at = NOW() """, {**product, 'site_id': site_id, 'data': json.dumps(product.get('extra', {}))}) 

This approach guarantees one row per product, and updated_at provides an update history.

Typical Mistakes

What’s Included in the Work

• Schema design tailored to your domain (raw data, products, categories).
• Implementation of upsert logic to avoid duplicates.
• Index configuration (GIN, B-tree) for fast queries.
• Documentation of the structure and operations.
• Team training on working with JSONB.
• Support for 2 weeks after delivery.

Over 8 years, we have accumulated experience solving similar tasks: more than 120 projects, from small stores to marketplaces with millions of products. We guarantee quality and optimization for Core Web Vitals.

Timelines and Contact

Basic schema with upsert and indexes — 1-2 working days. Full solution with documentation and training — up to 5 days. Contact us to get your project evaluated. Get a consultation on schema design for your project. We help you avoid common mistakes and speed up development.