Skip to content

testing strategies

Jump to bottom
Tyler Neely edited this page Nov 30, 2018 · 2 revisions

We believe operators of stateful systems should get as much sleep as they want. We take testing seriously, and we take pains to avoid the pesticide paradox wherever possible.

sled uses the following testing strategies, and is eager to expand their use:

  • quickcheck-based model testing on the Tree, PageCache, and Log
  • proptest-based model testing on the PageTable using the model testing library
  • linearizability testing on the PageTable using the model testing library
  • deterministic concurrent model testing using linux realtime priorities, approaching the utility of the PULSE system available for the Erlang ecosystem
  • ThreadSanitizer on a concurrent workload
  • LeakSanitizer on a concurrent workload
  • failpoints with model testing: at every IO operation, a test can cause the system to simulate a crash
  • crash testing: processes are quickly spun up and then kill -9'd while recovering and writing. the recovered data is verified to recover the log in-order, stopping at the first torn log message or incomplete segment
  • fuzzing: libfuzzer is used to generate sequences of operations on the Tree
  • TLA+ has been used to model some of the concurrent algorithms, but much more is necessary
Clone this wiki locally