Sandbox Lifecycle Management: Pause, Resume, Snapshot, Hibernation

🌏 中文版

What This Post Is

This post is the "business" post (order 28) of the Sandbox Series, also serving as the series appendix. It focuses on sandbox lifecycle management:

1. Pause & Resume — Release compute resources, retain state
2. Snapshot — Point-in-time state copy, create new sandbox
3. Hibernation — Long-term idle, zero compute cost state retention
4. Long-running task scenarios — When lifecycle management is needed
5. Cost optimization — Reducing costs through lifecycle management

---

Lifecycle Stages

The complete sandbox lifecycle has six stages:

Stage	Description	Resource Usage
Create	Start sandbox from image	Full
Execute	Run agent code	Full
Pause	Release CPU, retain memory & process	Medium (memory)
Snapshot	Create state checkpoint	Low (storage)
Hibernate	Long-term idle, zero compute cost	Very low (storage)
Destroy	Release all resources	Zero

Create → Execute → [Pause ↔ Resume] → [Snapshot → Create from snapshot] → [Hibernate → Wake] → Destroy

---

Pause & Resume

Concept

Pause temporarily freezes sandbox execution, releases CPU resources, but retains memory and process state. Resume continues execution from the paused state.

Applicable Scenarios

• Interactive agent workflows: User-agent back-and-forth conversation, sandbox can pause between messages
• Long tasks waiting for input: Task waiting for external input, sandbox can pause during wait
• Cost optimization: No CPU charges during idle periods

Implementation Differences

Provider	Pause/Resume API	Resume Speed	Retained Content
E2B	pause/resume API	~1 second	Memory + process + files
Modal	pause/resume API	sub-second	Memory + process + files
Vercel	pause API	3–3 seconds	Files (memory rebuilt)
Cloudflare	pause API	2–3 seconds	Files (memory rebuilt)
Azure Container Apps	pause/resume API	Depends on scale	Memory + process + files

Caveats

• Memory retention cost: Memory still occupies resources during pause; some providers charge by memory size
• Network disconnection: Network connections drop during pause, need to reconnect after resume
• External dependency expiration: Tokens, leases, locks may expire during pause

---

Snapshot

Concept

Snapshot is a point-in-time copy of sandbox state, used for creating new sandboxes, backup, and restore.

Difference from Pause

Feature	Pause	Snapshot
Continue execution	Original sandbox continues	Create new sandbox
Memory state	Retained	Not retained
Process state	Retained	Not retained
File state	Retained	Retained
Applicable scenario	Short interruption	Long-term storage, cross-sandbox sharing

When to Use

• Create checkpoints: Create snapshots before important task steps, rollback on failure
• Cross-sandbox sharing: Copy one sandbox state to multiple new sandboxes
• Long-term storage: When pause cost is too high, convert to snapshot storage

---

Hibernation

Concept

Hibernation is a long-term idle state where the complete sandbox state (including memory) is stored to storage media, compute resources fully released, compute cost reduced to near zero.

Applicable Scenarios

• Long-term idle: Sandbox not used for days or weeks
• Cost-sensitive: Zero compute cost state retention
• Periodic workflows: Agent workflows running weekly or monthly

Implementation Differences

Provider	Hibernation Support	Retention Time	Wake Speed
Modal	No	N/A	N/A
E2B	Yes	Unlimited	Seconds
Blaxel	Yes (standby)	Unlimited	25 ms
Azure Container Apps	Yes	Depends on scale	Depends on scale
Freestyle	Yes	Unlimited	Seconds

---

Long-Running Task Scenarios

Sandbox lifecycle management is especially suitable for the following long-running task scenarios:

Scenario 1: Multi-step Agent Workflow

Step 1: Analyze data (sandbox executes)
    ↓ Pause
Wait for user confirmation
    ↓ Resume
Step 2: Generate report (sandbox continues executing)

Scenario 2: Parallel Evaluation

Create multiple snapshots from the same baseline state, launch multiple parallel evaluation sandboxes, compare different agent performance.

Scenario 3: Long Training & Evaluation

ML model training or evaluation may take hours or days; sandbox can pause and resume during the process, avoiding restart from scratch after interruption.

---

Cost Optimization Strategies

Strategy	Description	Applicable Scenario
Auto-pause on idle	Sandbox auto-pauses after no activity	Interactive agents
Periodic snapshots	Periodic snapshots of important state, avoid recomputation	Long tasks
Hibernate long idle	Hibernate when not used for long period	Periodic workflows
Destroy on task completion	Destroy sandbox immediately after task completion	Short-lived tasks
Choose right billing model	Bursty → active-CPU, Long-running → wall-clock	All scenarios

---

What You Should Understand After Reading

1. Six sandbox lifecycle stages: create, execute, pause, snapshot, hibernate, destroy.
2. Difference between pause and snapshot, and applicable scenarios for each.
3. Hibernation concept and implementation differences (provider support varies).
4. Three long-running task scenarios: multi-step workflow, parallel evaluation, long training.
5. Five cost optimization strategies: auto-pause on idle, periodic snapshots, hibernate long idle, destroy on task completion, choose right billing model.

---

References

• Azure Container Apps Sandboxes Documentation
• E2B Sandbox Persistence
• Modal Sandbox Snapshots
• Blaxel Sub-second Sandbox Startup 2026
• Beam: Best Stateful Sandboxes for Code Execution in 2026
• Kubernetes-sigs/agent-sandbox: Guest Agent Proposal
• 本站：沙箱系列導讀（order 0）
• 本站：雲端產品全景與選型矩陣（order 7）
• 本站：沙箱成本模型（order 26）