Read Replicas: How to Offload Database Load Without Breaking Your Application

Imagine: your project has grown, the main page receives 10,000 requests per second, and the master database is choking on SELECT queries. LCP shoots up to 5 seconds, users leave. The typical solution is to buy a more powerful server, but that's expensive and only provides temporary relief. We set up read replicas in a few days and distribute the read load horizontally. Turnkey: from infrastructure configuration to documentation for your team. Infrastructure cost savings are obvious: instead of one expensive machine, we use several cheap ones, with a lower total cost.

Problems That Read Replicas Solve

High CPU and I/O load. When one server handles both writes and reads, the buffer cache is quickly evicted. Hit rate drops, disk reads increase. Offloading reads to replicas reduces resource contention on the master—one client saw master CPU drop from 85% to 30%.

Long analytical queries. Reports with JOINs on millions of rows block transactions and slow user queries. We route such queries to a dedicated analytical replica with different memory settings (work_mem = 256MB, effective_cache_size = 8GB)—query time drops by 70%.

Geo-distributed users. If your audience is in different regions, deploy replicas in nearby data centers and direct read queries to them. We use Route 53 latency-based routing for automatic selection of the nearest replica.

Why Read Replicas Are Not a Silver Bullet

They don't solve write conflicts—inserts and updates still go to the master. Also, replication lag must be considered: with asynchronous replication, the replica may lag by seconds. Therefore, we always implement sticky sessions or LSN waiting (see below). Without it, you risk serving stale data.

How We Do It: A Real Case

One of our client systems: Laravel 10 on PostgreSQL 16, 50,000 unique visitors per day, the master handles 2000 req/s on average, of which 1700 are reads (85%). We deployed three read replicas—one for the public API, one for the admin panel, and one for reports. Setup took 4 days, including zero-downtime migration.

Key steps:

• Created replicas via pg_basebackup, set up asynchronous replication.
• Configured Laravel for read/write split with automatic balancing among replicas.
• For reports—a separate replica with adjusted parameters (work_mem = 256MB).
• Added lag monitoring in Prometheus with alerts at delay >30s.

Result: master load dropped 5x, LCP decreased from 3s to 0.8s, analytical queries no longer affect users.

How to Avoid Replication Lag Issues

After a write to the master, you cannot immediately read from a replica because data may not have copied yet. Solution: pass the write's LSN position to the client and, before reading, verify that the replica has caught up to that LSN. If not, redirect the query to the master. We embed this pattern directly into the application, preventing data races.

-- On master: get current LSN after INSERT SELECT pg_current_wal_lsn(); 

# Example check on replica (pseudocode) def read_after_write(lsn): if replica.is_caught_up(lsn): return replica.execute(query) else: return master.execute(query) 

What to Do When Replication Lag Is Critical

If lag exceeds 60 seconds, a critical alert fires. Action plan:

1. Check if the replication process is blocked (pg_stat_replication).
2. Ensure the master has enough free space for WAL files.
3. Temporarily remove the replica from routing until it syncs.

Comparison of Synchronous and Asynchronous Replication

Example configuration for asynchronous replica (postgresql.conf):

hot_standby = on hot_standby_feedback = on max_standby_streaming_delay = 30s wal_receiver_timeout = 60s 

Process

1. Audit current load—collect metrics (CPU, IOPS, WAL generation), determine query profile.
2. Choose topology—how many replicas, synchronous or asynchronous, separate analytical replica if needed.
3. Configure replicas—create via pg_basebackup, tune postgresql.conf.
4. Application routing—Laravel config, pgBouncer R/W split, or custom middleware.
5. Monitoring and alerts—deploy Grafana dashboard, set notifications for lag.
6. Documentation and training—handover schema, credentials, instructions for promoting a replica to master.

What's Included

• Replication schema and configuration files (postgresql.conf, pgbouncer.ini).
• Application setup for read/write split (Laravel, Sequelize, Django ORM).
• Grafana dashboards with key metrics (lag, replica count, WAL size).
• Maintenance documentation (how to add a replica, what to do on failure).
• Training for your engineers—we demonstrate on our test environment.
• 30-day post-deployment support.

Timelines

Basic configuration with two replicas: 2 to 3 business days. If a large data migration (1+ TB) or global replication across multiple regions is needed, up to 5 days. We provide an exact estimate after auditing your system.

Comparison of approaches:

Our team has completed over 50 DB scaling projects. We guarantee that after setup, read performance improves at least 3x. Contact us for a free audit of your system. Order read replicas setup and receive documentation.