Detecting failures in distributed systems

Falcon is a fast and practical reliable failure detector. Falcon has three goals: to be fast (sub-second detection time in the common case), to be reliable (never report an alive process as down), and to be minimally disruptive.
Falcon achieves these three goals by deploying a network of spies across various layers of the system. Spies use inside information local to a layer to determine aliveness. In order to provide reliability, spies sometimes kill. However, to achieve minimal disruption, spies limit their use of lethal force. Here is an abstract view of Falcon:



FALCON stands for Fast And Lethal Component Observation Network

• Detecting failures in distributed systems with the FALCON spy network
  Joshua B. Leners, Hao Wu, Wei-Lun Hung, Marcos K. Aguilera, and Michael Walfish
  23rd ACM Symposium on Operating Systems Principles
  SOSP 2011, Cascais, Portugal, October 2011   [pdf] [talk]
  (on the design, implementation, and evaluation of Falcon)
• No Time for Asynchrony
  Marcos K. Aguilera and Michael Walfish
  12th Workshop on Hot Topics in Operating Systems
  HotOS 2009, Monte Verità, Switzerland, May 2009   [pdf]
  (a philosophical argument against designing distributed systems under the asynchronous model)
  
• Improving availability in distributed systems with failure informers
  Joshua B. Leners, Trinabh Gupta, Marcos K. Aguilera, and Michael Walfish
  10th USENIX Symposium on Networked Systems Design and Implementation
  NSDI 2013, Lombard, IL, April 2013   [pdf] [talk]
  (describes Pigeon, an alternative to Falcon that handles network failures and does not kill)
• Taming uncertainty in distributed systems with help from the network
  Joshua B. Leners, Trinabh Gupta, Marcos K. Aguilera, and Michael Walfish
  ACM European Conference on Computer Systems
  EUROSYS 2015, Bordeaux, France, April 2015   [pdf]
  (describes Albatross, which virtualizes network partitions as an alternative to killing)

• Marcos K. Aguilera
• Trinabh Gupta
• Wei-Lun Hung
• Joshua B. Leners
• Michael Walfish
• Hao Wu

This project is supported in part by AFOSR grant FA9550-10-1-0073 and NSF grants 1055057 and 1040083.

• image.tar.xz: 15 GB QEMU raw x86 VM image (~3.1 GB compressed) with code for Falcon, Pigeon, Albatross, our modified versions of ZooKeeper and libvirt, and other dependencies. To use this image, you will need to install Qemu/KVM (e.g., with

  apt-get install qemu-system

  ). After extracting raw image (endhost.img) with

  tar xJf image.tar.xz

  , you can then execute the virtual machine with the command

  qemu-system-x86_64 -hda endhost.img

  . If you have issues with any of these steps, do not hesitate to contact us.
• eval_driver.tbz: automated eval scripts
• configs.tbz: configuration notes and settings, for advance users wishing to use Falcon in new environments

If you would like to contribute to Falcon, you can check out our GitHub Falcon Project

leners at cs.utexas.edu