Saboteur is a network fault injection tool that aims to simplify resilience and stability testing. Its core component is an agent that accepts commands over HTTP and configures its host's network stack for various common fault scenarios.
Currently these include:
- Total network partition
- Remote service dead (not listening on the expected port)
- Delays
- Packet loss
- TCP connection timeout (as often happens when two systems are separated by a stateful firewall)
Currently it is Linux only, but support may be added for OSX in future.
RPM and DEB packages are available for the agent. See Releases.
A CLI is also available as an RPM under releases, or you can just download and use it directly from source https://raw.github.com/tomakehurst/saboteur/master/bin/sab.
Saboteur's only dependency is Python 2.6.6+.
To install from the RPM:
$ rpm -ivh https://github.com/tomakehurst/saboteur/releases/download/v0.7/saboteur-0.7-1.noarch.rpm
And from the DEB package:
$ wget -P /tmp https://github.com/tomakehurst/saboteur/releases/download/v0.7/saboteur_0.7_all.deb && dpkg --install /tmp/saboteur_0.7_all.deb
After installation, start the agent daemon in the usual way:
$ service saboteur-agent start
So far I've verified that Saboteur works on Centos 6.4, Red Hat 6.4 and Ubuntu 12.10. It currently has some issues with the tc defaults on Fedora 18, so only the NETWORK_FAILURE, SERVICE_FAILURE and FIREWALL_TIMEOUT fault types work on that OS.
I'd welcome any feedback about successes/failures on other distros or versions.
The Saboteur package installs a CLI executable which can be used to configure and reset faults both locally and remotely.
Typing sab without arguments (or with -h) will display usage information.
To create a new fault:
$ sab add --fault_type NETWORK_FAILURE --to_port 8818 --direction IN --hosts myhost1,myhost2
To remove all faults:
$ sab reset --hosts myhost1,myhost2
Note: omitting --hosts will cause commands to be sent to localhost.
Saboteur has a simple JSON over HTTP API. To add a new rule, simply POST to the agent process on port 6660 e.g.
$ curl -X POST -d '{ "name": "packet-loss-to-app-server", "type": "PACKET_LOSS", "direction": "IN", "to_port": 8080, "probability": 0.2, "correlation": 25 }' http://192.168.2.11:6660/
See the tests for more examples of valid commands.
To reset all faults, send a DELETE request to the root path e.g.
$ curl -X DELETE http://192.168.2.11:6660/
Crash Lab is a Java library for automating resilience/stability type tests (or any scenario involving network faults), which has a Saboteur client as part of it, and some examples.
The Saboteur agent currently implements its reset feature by deleting all iptables and tc rules, so it currently won't play nicely on systems that have rules configured via other means. This will be improved in future, probably by having Saboteur keep track of rules it has created so that these can be targeted for reset.
It's also likely that adding multiple DELAY and/or PACKET_LOSS rules will not work correctly.
Fixes are coming for both of these issues.
Finally, please don't use this in production (yet)! It's in no way secure, and would make an ideal attack vector for anyone wanting to DoS your site.
I'm considering trying to make the agent secure enough to be used as a building block for a chaos monkey. Again, any feedback on the usefulness of this would be appreciated!