Skip to main content
Loading navigation
Back to challenges

Distributed Inference Fault Injection & Recovery Simulator with AutoGen

Orchestrate a multi-agent distributed inference fault simulator using AutoGen to test node failover and KV-cache recovery protocols during high-throughput serving. Maintain zero dropped user requests and recover lost state in under 500 milliseconds during simulated hardware outages.

EngineeringHosted by Vera
Sign in free to start

What you are building

Build a multi-agent cluster fault injection and recovery simulation framework using AutoGen.

How work is evaluated

Evaluation

Evaluates dropped request rate and state recovery duration during hardware fault injection.

Datasets

Public datasets, plus the private datasets you have access to.

How submissions are scored

Dimensions
2 scoring checks
Binary
2 pass or fail dimensions
Ordinal
0 scaled dimensions
Dimension 1

Zero Dropped Requests

Ensure zero active user streams are dropped during worker failover.

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Dimension 2

Recovery Duration MS

Time in milliseconds required to migrate KV cache state to backup worker. • target: 380 • range: 0-500

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Learning goals

  • Build a multi-agent resilience simulation using AutoGen framework.

  • Simulate node crash failures, network partition events, and GPU memory ECC faults.

  • Implement distributed state migration for continuous LLM generation streams.

  • Verify zero-data-loss failover boundaries during peak request throughput.

Resources

Dataset notes

Fault injection event logs and stream state snapshots across distributed worker nodes.

Challenge setup

Evaluates dropped request rate and state recovery duration during hardware fault injection.

Challenge input

JSON scenario configuration specifying failure type and active stream count

AutoGen

Microsoft framework for orchestrating conversing agents in system simulations

Evaluated output

JSON result containing dropped_requests, recovery_duration_ms, and state_consistency_score

Evaluation checks
  • Ensure zero active user streams are dropped during worker failover.
  • Time in milliseconds required to migrate KV cache state to backup worker. • target: 380 • range: 0-500
Configured evaluation evidence
  • Benchmark: Distributed System Fault Tolerance Suite
  • Recovery Duration MS target: 380
  • 1 public reference case
Configured runtime
  • Python execution harness
  • Python sandbox (unavailable on Versalist)

Shows the first task, up to three tools, and up to three evidence items.

View technical recipe

Configured tools

Action Space
  • AutoGen · Required
  • Anthropic · Optional
  • Microsoft AutoGen · Optional

Evaluation contract

  • Zero Dropped Requests · Weight 1
  • Recovery Duration MS · Weight 1

Metrics

  • Recovery Duration MS · Target 380

Benchmarks

  • Distributed System Fault Tolerance Suite
  • general

Reference cases

  • 1 public reference case

Workflow tasks

  1. Task 1

    Input: JSON scenario configuration specifying failure type and active stream count

    Output: JSON result containing dropped_requests, recovery_duration_ms, and state_consistency_score

    Execution harness: Python

Evaluation tools

  • AutoGen · Agent Framework

Recipe freeze marker

No recipe freeze time is recorded.

Frequently Asked Questions about Distributed Inference Fault Injection & Recovery Simulator with AutoGen