Persistent Storage Versus Snapshotting for Data | Generated by AI

Home 2026.06 English 中文

---

Question: Would using an attached block storage volume make snapshot creation faster compared to snapshotting the root disk?

Answer:

Actually, it’s a nuanced tradeoff. Let me break it down:

Block Storage Volume Snapshots vs Droplet Snapshots

	Droplet Snapshot	Volume Snapshot
What’s captured	Entire root disk	Just the volume
Speed to create	Slow (full disk)	Also slow (same mechanism)
Speed to restore	Fast (CoW lazy)	Also fast (CoW lazy)
Cost	$0.06/GB/month	$0.06/GB/month
Reattach to new droplet	❌ Not directly	✅ Yes, detach/reattach

So snapshot creation speed is roughly the same — both go through the same CoW block storage backend.

Where Block Storage Volume Actually Wins for Your Workflow

The real advantage is persistence and reattachability:

Droplet Snapshot workflow (current):
  destroy droplet → snapshot stored → create new droplet → lazy hydrate

Block Storage Volume workflow:
  destroy droplet → volume persists → create new droplet → attach volume → done

With a volume, no snapshot needed at all between sessions:

# End of session
doctl compute droplet delete amd-gpu --force
# Volume still exists, data intact, no snapshot creation time

# Next session
doctl compute droplet create amd-gpu \
  --image <base-image-id> \
  --size gpu-h100x80-1 \
  --region ... \
  --ssh-keys ... 
  # then attach volume
doctl compute volume-action attach <volume-id> <new-droplet-id>

What to Put Where

Root disk (small, cheap droplet snapshot):
  - OS
  - conda/pip environments
  - dotfiles, SSH keys, tmux config
  - your CLI tools (ww, zz, iclaw)

Block storage volume (persistent, never snapshotted):
  - Model weights (70B models, checkpoints)
  - Datasets
  - Training outputs
  - Git repos with large artifacts

The Optimal Workflow for You

Session start (~2 min total):
  1. Create droplet from small root snapshot (fast, lazy CoW)
  2. Attach persistent volume (instant, just a mount)
  3. mount /dev/sda /mnt/data
  4. All your models/datasets immediately available

Session end:
  1. Save any new model checkpoints to volume
  2. Destroy droplet (no snapshot needed)
  3. Volume persists, billed at ~$0.10/GB/month at rest

Cost Reality Check

For 130GB of model weights on a volume:

130GB × $0.10/GB/month = $13/month
vs
130GB snapshot × $0.06/GB/month = $7.80/month

Volume is slightly more expensive but saves you 20+ min snapshot creation time every session and gives you instant attach without hydration lag for the heavy data.

Bottom Line

• Snapshot speed: no real difference between root disk vs volume snapshot
• Real win: use volume as persistent storage so you skip snapshotting entirely for your large model weights
• Best setup: thin root disk snapshot (OS + env) + persistent volume (models + data)

This is essentially the same pattern as cloud ML workflows on AWS (EBS volume for datasets, AMI snapshot just for the base environment).

---

Back Donate