Permalink
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
57 lines (48 sloc) 3.38 KB

Testing

Approaches

  • Test driven development. Relevant etcd raft tests have been ported to dragonboat to ensure all corner cases identified by the etcd project have been handled.
  • High test coverage. Extensively tested by unit testing and monkey testing.
  • Linearizability checkers. Jepsen's Knossos and porcupine are utilized to check whether IOs are linearizable.
  • Fuzz testing using go-fuzz.
  • I/O error injection tests. charybdefs from scylladb is employed to inject I/O errors to the underlying file-system to ensure that Dragonboat handle them correctly.
  • Power loss tests. We test the system to see what actually happens after power loss.

Monkey Testing

Setup

  • 5 NodeNosts and 3 Drummer servers per process
  • hundreds of Raft clusters per process
  • randomly kill and restart NodeHosts and Drummer servers, each NodeHost usually stay online for a few minutes
  • randomly delete all data owned by a certain NodeHost to emulate permanent disk failure
  • randomly drop and re-order messages exchanged between NodeHosts
  • randomly partition NodeHosts from rest of the network
  • for selected instances, snapshotting and log compaction happen all the time in the background
  • committed entries are applied with random delays
  • snapshots are captured and applied with random delays
  • a list of background workers keep writing to/reading from random Raft clusters with stale read checks
  • client activity history files are verified by linearizability checkers such as Jepsen's Knossos
  • run hundreds of above described processes concurrently on each test server, 30 minutes each iteration, many iterations every night
  • run concurrently on many servers every night

Checks

  • no linearizability violation
  • no cluster is permanently stuck
  • state machines must be in sync
  • cluster membership must be consistent
  • raft log saved in LogDB must be consistent
  • no zombie cluster node

Results

Some history files in Jepsen's Knossos edn format have been made publicly available.

Benchmark

Setup

  • Three servers each with a single 22-core Intel XEON E5-2696v4 processor, all cores can boost to 2.8Ghz
  • 40GE Mellanox NIC
  • Intel 900P for storing the RocksDB's WAL and Intel P3700 1.6T for storing all other data
  • Ubuntu 16.04 with Spectre and Meltdown patches, ext4 file-system

Benchmark method

  • 48 Raft clusters on three NodeHost instances across three servers
  • Each Raft node is backed by a in-memory Key-Value data store as RSM
  • Mostly update operations in the Key-Value store
  • All I/O requests are launched from local processes
  • Each request is handled in its own goroutine, simple threading model & easy for application debugging
  • fsync is strictly honored
  • MutualTLS is disabled

Intel Optane SSD

Compared with enterprise NVME SSDs such as Intel P3700, Optane based SSD doesn't increase throughput when payload is 16/128 bytes. It does slightly increase the throughput when the payload size is 1024 byte each. It also improves write latency when the payload size is 1024.