Replit Storage: Next-Generation Infrastructure

As AI model development and the installation of large-scale dependencies have increased, Replit has significantly expanded its storage limit from the previous 1 GiB to 256 GiB or more.

The previous infrastructure ran on btrfs file systems and LVM thin pools. Each time a new Repl was created, it went through a complex process of creating a virtual device, setting quotas via overlayfs, and then deserializing a snapshot of the file system to prepare it.

This approach caused serious problems, including:

• Slow boot times: As capacity grew, loading the file system took several minutes, degrading the user experience.
• High CPU load: The snapshot creation and serialization process consumed about 50% of system CPU.
• System fragility: As capacity grew, the recovery and boot process became increasingly unstable.

To solve this, Replit is transitioning its infrastructure away from saving the entire disk as a snapshot and toward a method that loads the necessary data on demand, exactly when it's needed.