roshi-walker walks the keyspace of a Roshi farm in semirandom order (random order of Redis instances; Redis SCAN command on each instance) and a user-defined rate. It makes Select request for each key, using the SendAllReadAll read strategy in order to perform complete read repair.
roshi-walker uses vendored dependencies and a "blessed build" process to ensure
stability over time. Users should get and build roshi-walker by cloning this
repository and running make in the roshi-walker subdirectory. A working Go
toolchain is assumed.
git clone git@github.com:soundcloud/roshi
cd roshi/roshi-walker
make
It's also possible to get roshi-walker via go get, and/or build it with a
simple go build, with the caveat that it will use your normal GOPATH to
resolve dependencies, and therefore will enforce no constraints on dependency
versions, which could introduce bugs or strange behavior.
roshi-walker is designed to be used in two situations.
One or more roshi-walker instances can be started for a given Roshi farm, set to walk at a moderate rate and run forever. That will ensure that all keys are made consistent, within some upper time bound. That in turn improves data durability, especially if normal reads are non-uniform or low volume.
It's recommended to run a few roshi-walker processes in this mode on every Roshi farm. Test against your infrastructure to determine an appropriate rate.
roshi-walker supports a -once flag, which will walk the entire keyspace once and exit. This is useful if a Redis instance has crashed and come back online with no data. In this situation, while the farm still returns correct data, it is less resilient to further node failure. After the walk is complete, the empty instance will be repopulated with relevant data via read repair, and the resiliency of the farm is returned to normal levels.