--- title: "Database Design Best Practices" source: https://refact.co/insights/digital-product/database-design-best-practices author: "Hossein Karami" date: "2026-03-25" --- # Database Design Best Practices Your product should not feel slow. If dashboards take forever to load, new features feel risky, or data looks off, the problem is often the database. Not your UI. Not your framework. The database is the foundation, and small mistakes early can become expensive problems later. Many founders treat the database like plumbing. It works, so nobody touches it. Then traffic grows, reporting breaks, and every change feels risky. We have seen this across many products, and the fix is usually simple in theory: clean up the foundation before the cracks spread. You do not need to become a database engineer. You need a few clear rules that keep data clean, queries fast, and future changes manageable. When you start changing schemas in production, a safe [data migration service](https://refact.co/services/data-migration) process matters too. This guide walks through 10 database design best practices we rely on when building products meant to last. ## 1\. Normalization and Logical Schema Design As your app grows, data can become messy fast. Normalization helps prevent that by reducing duplicate data and keeping relationships clear. Instead of one giant table that tries to hold everything, split data into focused tables. Each table should have one job. Tables should connect through IDs. Example: a publishing platform might start with one `articles` table that also stores author names and category names. A better setup splits this into `articles`, `authors`, and `categories`. The `articles` table stores `author_id` and `category_id`, not repeated text. > **Key insight:** Start with Third Normal Form, or 3NF. It is a strong default for most SaaS and content products, and it prevents common data mistakes. One more rule, assume your schema will change. Plan for safe updates early so changes stay repeatable and low risk. ## 2\. Scalable Primary Key Strategy Your primary keys affect performance, security, and scaling. Changing them later is painful, so decide early. If your product will be distributed, multi-tenant, or merged across systems, UUIDs can prevent ID collisions. If your system is simple and centralized, sequential `bigint` IDs are smaller and often faster. Example: a SaaS app with many tenants may use UUIDs for `users` and `projects`. An ecommerce store may prefer sequential IDs for `orders` because they are compact and easy to analyze. > **Key insight:** If you choose UUIDs, consider UUID v7. It sorts better than random UUIDs, which can help index performance. Write this decision down in your engineering notes. It should not live only in someone’s head. ## 3\. Appropriate Indexing Strategy If your app slows down as data grows, indexing is often the first place to look. Indexes help the database find rows without scanning full tables. Think of an index like the index in a textbook. Without it, you flip every page. With it, you jump to the right spot. A publishing tool might index `publication_date`, `status`, and `author_id`. An ecommerce store might index `product_sku` and `category_id`. A membership site might use a combined index on `user_id` and `subscription_status`. > **Key insight:** Index columns used in `WHERE`, `JOIN`, and `ORDER BY`. Use tools like `EXPLAIN ANALYZE` in PostgreSQL to confirm the index is actually being used. Indexes also have a cost. They use disk space and can slow writes. Review them over time and remove the ones you do not need. If your product runs on Postgres and query speed is a growing problem, a [PostgreSQL development team](https://refact.co/technologies/postgresql) can help you tune both schema and query patterns. ## 4\. Referential Integrity and Foreign Key Constraints Orphan records are a quiet problem. You might have a payment with no order, or a subscription that points to a user who no longer exists. Foreign keys prevent this by forcing valid relationships between tables. The database becomes the guardrail, not just your application code. Example: in a membership product, `subscriptions.user_id` should reference `users.id`. In a billing system, `payments.subscription_id` should reference `subscriptions.id`. > **Key insight:** Decide deletion rules early with `ON DELETE` and `ON UPDATE`. Use `CASCADE` only when you are very sure. Use `RESTRICT` when deletion would break history or reporting. Test these rules with real sample data. A bad cascade rule can remove far more than you expected. ## 5\. Efficient Data Type Selection Data types are not just preferences. They affect storage, index size, query speed, and correctness. Example: do not store money in `FLOAT`. You will get rounding errors. Use `DECIMAL(10,2)`. For counters that can grow fast, use `BIGINT` instead of `INTEGER`. > **Key insight:** Use `TIMESTAMP WITH TIME ZONE` for user-facing dates. It prevents timezone bugs that lead to confusion and support tickets. Be careful with flexible fields too. Use `TEXT` for long content. Use structured columns for values you filter on often. If you store metadata as JSON, be clear about how you will query and index it. ## 6\. Soft Delete and Audit Trail Patterns Hard deletes feel simple until someone deletes the wrong thing. Then you need an undo button that does not exist. Soft deletes keep records but hide them from normal queries. A common pattern is `is_deleted` plus `deleted_at`. Your app then filters to show only active records. Example: an ecommerce store may keep deleted orders for audits. A membership product can track churn without losing history. A training platform can prove when content changed. > **Key insight:** Add an audit log. Track what changed, who changed it, and when. This helps with debugging and can support compliance. Set retention rules too. You might keep soft-deleted records for 30 days for recovery, then remove them later if that matches your policy. ## 7\. Denormalization for Performance-Critical Scenarios Normalization keeps data clean, but it can make some read paths slower. For read-heavy systems, denormalization can be worth it. Denormalization means storing some repeated data on purpose so you can avoid joins. Example: store `category_name` on `products` to avoid joining `categories` on every product list. Or store a cached `subscription_count` on `plans` so you do not run `COUNT(*)` all day. > **Key insight:** Only denormalize after you find a real bottleneck. Measure first, then change the structure. Also document how the duplicated data stays accurate. Triggers, background jobs, or application writes can all work, but memory alone is not a system. ## 8\. Partitioning and Sharding Strategy When tables become massive, a single table gets harder to query and harder to maintain. Partitioning breaks one logical table into smaller physical parts. Example: partition event logs by month. Archive old partitions without touching new data. For multi-tenant systems, you might partition by `tenant_id` to keep each customer’s data grouped. > **Key insight:** Do not partition too early. It adds complexity. Consider it when tables are truly huge or query patterns are very consistent. Retrofitting partitioning into a live system is risky. If you think you will need it, plan ahead before your table reaches a painful size. ## 9\. Multi-Tenancy and Data Isolation If you run a SaaS product, multi-tenancy mistakes can cause data leaks. That is not just a bug. It is a trust problem and sometimes a compliance problem too. A common approach is shared tables with a `tenant_id` column. Every query must filter by it, and your indexes should include it. Example: in a membership platform with many organizations, `users`, `subscriptions`, and `payments` should all include `organization_id`. Some industries may need a database per tenant, but that adds ongoing operational work. > **Key insight:** Put `tenant_id` on every tenant-owned table and include it in indexes. Then add a second layer like PostgreSQL Row-Level Security when it fits. Never trust the client to tell you the tenant. Tenancy should come from the authenticated session and be enforced on every request. If your product needs secure reporting, user roles, and data views across accounts, [custom portals and dashboards](https://refact.co/services/portals) often need this enforced at every layer. ## 10\. Connection Pooling and Resource Management Opening a new database connection for every request is slow and expensive. Under traffic spikes, it can also crash your database. Connection pooling keeps a set of ready connections your app can reuse. Tools like PgBouncer, or built-in pools in most backend stacks, help a lot. Example: a Node.js app might use `pg-pool` with a capped max size. A high-traffic store might put PgBouncer in front of Postgres to handle many short-lived requests safely. > **Key insight:** Monitor pool usage. If you see pool exhaustion, either raise the cap or find the slow queries holding connections too long. Connection issues are often a symptom, not the root cause. Slow queries, bad indexes, and oversized workloads usually show up here first. ## Top 10 Database Design Best Practices Comparison | Pattern | Implementation Complexity | Resource / Performance Impact | Expected Outcomes | Ideal Use Cases | Key Tips | | --- | --- | --- | --- | --- | --- | | Normalization and Logical Schema Design | Moderate to high upfront planning | Low storage waste, joins can cost on reads | Consistent data, fewer anomalies | SaaS, publishing, evolving content | Start at 3NF, document relationships, denormalize only after measuring | | Scalable Primary Key Strategy | Low to medium, must decide early | UUIDs cost more space, sequential IDs are compact | Fewer collisions, better scaling options | Multi-tenant SaaS for UUIDs, centralized systems for sequential IDs | Consider UUID v7, document the choice and stick to it | | Appropriate Indexing Strategy | Medium, needs ongoing review | Big read gains, some write and storage cost | Faster queries and happier users | Search, filtering, dashboards | Index WHERE, JOIN, and ORDER BY columns, confirm with EXPLAIN | | Referential Integrity and Foreign Key Constraints | Low to medium | Small write overhead, prevents corruption | Cleaner relationships, safer deletes | Orders and payments, subscriptions and users | Set delete and update rules, test cascades carefully | | Efficient Data Type Selection | Medium | Strong impact on storage and speed | Accurate data and smaller indexes | Money, timestamps, counters, IDs | Use DECIMAL for money, timezone-aware timestamps, BIGINT for growth | | Soft Delete and Audit Trail Patterns | Medium | More storage, slightly more query logic | Recovery and history | Compliance, churn analysis, finance | Use deleted\_at, add audit logs, set retention rules | | Denormalization for Performance-Critical Scenarios | Medium to high | Faster reads, more write complexity | Faster pages in hot paths | Read-heavy dashboards, feeds | Measure first, document sync rules | | Partitioning and Sharding Strategy | High | Major gains at huge scale | Handles billions of rows better | Event logs, large SaaS, regional data | Pick a partition key that matches query patterns | | Multi-Tenancy and Data Isolation | High | Varies by approach | Prevents cross-tenant leaks | SaaS products | Always include tenant\_id, index it, consider Row-Level Security | | Connection Pooling and Resource Management | Low to medium | Lower connection overhead | More stable under spikes | High-traffic apps | Set timeouts, watch pool usage, find slow queries | ## So, what is the next step for your database? This is a lot to take in. You do not need to fix everything this week. Pick the one problem causing the most pain right now, fix it, and measure the result. ### Make your first move - **Pages are slow?** Start with indexing and query analysis. One good index can make a noticeable difference. - **Data looks inconsistent?** Add foreign keys and tighter constraints so bad relationships cannot enter your system. - **Building SaaS and worried about customer isolation?** Review your tenancy rules and make sure every table and index supports tenant filtering. A database that kind of works will eventually slow product work down. A database you can trust becomes a real advantage. If you want a second set of eyes on your foundation, [talk to our team](https://refact.co/contact). We will help you find the highest-impact fixes and plan changes without breaking production.