Sandbox filesystem layering
- tmpfs (Writable Layer)RAM Speed (Write)
- EROFS (Base Image Layer)Read-Only Base
01from blaxel.core import SandboxInstance02 03sandbox = await SandboxInstance.create({04 "name": "agent-workspace",05 "image": "blaxel/base-image:latest",06 "memory": 4096, # about half backs files07 "region": "us-pdx-1",08})09 10# Writes land in the RAM layer11await sandbox.fs.write("/app/plan.md", "...")[DISCOVER OUR SANDBOXES STORAGE]
A RAM filesystem at the speed of your agents
Blaxel keeps the writable layer of every sandbox in RAM. Your image sits in a read-only base, an in-memory layer takes every write, and the two are stitched together transparently, so reads and writes land at memory speed. The agent sees one ordinary filesystem and never has to think about where the bytes live.
Writes hit RAM
A tmpfs layer in the sandbox's memory takes every write, so file operations run at memory speed.
Read-only base
Your image is stored on the host as a compact, read-only EROFS layer underneath.
One seamless filesystem
OverlayFS merges the two layers, sending reads to the base and writes to RAM. The agent sees one filesystem.
Configurable memory split
By default, about half of a sandbox's memory backs its filesystem. Change the ratio to give files or processes more room.
[SUSPEND WITHOUT LOSING A BYTE]
Snapshots persist for as long as needed
When a sandbox goes idle, Blaxel snapshots the entire state, the full in-memory filesystem and every running process, and parks it. Reconnect and it restores in under 25 milliseconds, with files and processes exactly as they were. The snapshot stays until you delete the sandbox, or until an expiration policy you set removes it.
Full memory snapshots
The snapshot captures the whole filesystem and running work, not just data at rest.
Perpetual sandboxes
Snapshots persist until you choose to remove them.
Sub-25ms resume
A suspended sandbox comes back almost instantly.
Expiration policies
Delete sandboxes automatically after a TTL, a period of inactivity so you stop storing snapshots that will never resume.
When to add a volume
In-memory storage
Default- Near-zero latency with RAM performance
- Snapshot entire process state automatically
- Half of sandbox memory by default, configurable
Persistent volumes
Optional- Scalable storage bounds
- Outlives the lifecycle of the sandbox container
- High performance block storage
