ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-11-07 01:12

Snorkel

3 steps1 pieces4 Sources

AI数据公司Snorkel通过采用Modal平台的弹性突发(Bursting)机制,使其Agent模拟沙箱的总账单降低了约20%。这一举措旨在解决Agent工作负载中资源使用尖峰化与不可预知性导致的静态配置浪费或运行失败问题。目前,Snorkel约有10-15%的沙箱启用了该功能。

在更广泛的行业背景下,长时运行Agent的基础设施正面临从“预留资源”向“按需/活跃资源计费”的转变。例如,Amazon在其内部系统Sandman中处理数百万级并发沙箱以支持强化学习训练;而Sailboxes和Upstash等平台则通过暂停空闲状态或仅对活跃CPU计费来进一步压缩成本。争议的核心在于如何在保证Agent执行环境隔离性与稳定性的前提下,最大化底层硬件的利用率。

How it got here

  1. 2026-11-07

    Snorkel高级AI平台经理Rustem Feyzkhanov表示,通过Modal的弹性突发机制,Snorkel的沙箱总账单降低了约20%,且运行可靠性提高。

    modal.com ↗
  2. 2026-10-30

    Upstash发布分析指出,传统按墙钟时间计费的沙箱在Agent等待LLM响应时产生大量闲置成本,建议转向按活跃CPU计费或自动暂停模式。

    upstash.com ↗
  3. 2026-10-01

    Amazon团队披露其内部沙箱服务Sandman在2026年10月的四个生产级SWE-bench训练活动中创建了310万个沙箱,峰值并发达11,858个。

    arxiv.org ↗

What is still unsettled

  • 弹性突发机制在超大规模并发下是否会导致资源争用从而增加P99延迟?
  • 随着更多厂商推出基于“活跃CPU”计费的方案,行业是否会形成新的成本基准标准?
  • Snorkel计划何时将弹性突发功能的覆盖率从目前的10-15%进一步提升?

Sources

Sign in / Sign up
ReadingResearchRadarInvestment framework
modal.com
Modal Blog: How Snorkel runs realistic agent simulations with Modal
↗
  • arxiv.orgArXiv: Sandman: Building and Operating Sandboxes for Agentic Reinforcement Learning on Shared Kubernetes↗
  • upstash.comUpstash KB: What an Idle Agent Sandbox Actually Costs↗
  • docs.sailresearch.comSail Research Docs: Sailboxes↗
  • Coverage in this topic

    • AI Data Company Snorkel Says Elastic Bursting Mechanism Reduced Agent Simulation Sandbox Costs by Approximately 20%2026-11-07News

      Snorkel's agent simulations face dual challenges of resource usage spikes and unpredictable workloads, where static configurations lead to failures or waste. The company adopted Modal's elastic bursting mechanism to scale resources on demand; the Senior Manager of AI Platform stated that total sandbox bills were thereby reduced by approximately 20%, with about 10-15% of sandboxes currently enabled for this feature.