Fly.io Volumes: Persistent Storage for Stateful AI Applications

# Create a 10 GB volume in the London region
fly volumes create ai_data --size 10 --region lhr
 
# List volumes for your app
fly volumes list
 
# Extend a volume (non-destructive, instant)
fly volumes extend vol_abc123 --size 20

[mounts]
  source = 'ai_data'
  destination = '/data'

A Fly Volume is tied to a specific region and a specific machine. If you scale to multiple machines, each machine needs its own volume — data is not automatically replicated between volumes.

FROM ollama/ollama:latest
 
ENV OLLAMA_MODELS=/data/models
 
# Startup script: pull model if not already present
COPY start.sh /start.sh
RUN chmod +x /start.sh
CMD ["/start.sh"]

#!/bin/sh
# start.sh
ollama serve &
sleep 5
 
# Pull model only if not already cached on the volume
if [ ! -d "/data/models/llama3.2" ]; then
  echo 'Pulling llama3.2...'
  ollama pull llama3.2
fi
 
wait

# fly.toml
app = 'my-ollama'
primary_region = 'lhr'
 
[mounts]
  source = 'ollama_models'
  destination = '/data'
 
[[vm]]
  memory = '8gb'
  cpu_kind = 'performance'
  cpus = 4

import os, sqlite3
 
# Store SQLite database on the persistent volume
DB_PATH = os.environ.get('DB_PATH', '/data/app.db')
 
def get_db():
    conn = sqlite3.connect(DB_PATH)
    conn.row_factory = sqlite3.Row
    return conn
 
# Database survives deploys because /data is a Fly Volume

SQLite on a Fly Volume is a surprisingly capable stack for AI applications serving under ~10,000 requests/day. It eliminates the cost and complexity of a managed Postgres instance for smaller deployments.

# Create a snapshot of a volume
fly volumes snapshots create vol_abc123
 
# List snapshots
fly volumes snapshots list vol_abc123
 
# Restore from snapshot (creates a new volume)
fly volumes create restored_data \
  --snapshot-id vs_abc123 \
  --size 10 --region lhr