functional-tester: txn stresser

[heyitsanthony]

Bugs in the txn code can lead to inconsistencies among members distinct from bugs caused by only puts/gets, so it should be tested in the functional tester. Although the etcd-runner tests use txns in some places, the throughput/concurrency/txn complexity isn't enough.

Flat txns could be stressed via:

1. Have a set of keys K (e.g., "/k0", "/k1", ..., "/k9")
2. Choose random keys kA...kB from K
3. Submit a txn If(kA.Version > 0, ..., kB.Version > 0).Then(Delete kA...kB).Else(Put kA...kB).

Nested txns stressing could assign a key per level:

If (kA.Version > 0) {
   Delete(kA)
   If (kB.Version > 0) {
      Delete(kB)
   } else {
      Put(kB)
   }
} else {
    Put(kA)
    If (kB.Version > 0) { ... }
}

Related: #8214

[xiang90]

1. All client write requests are executed in one single thread non-concurrently.
2. The only thing might happen concurrently with client writes is compaction.

We either violates 1, or there is a bug in 2.

Is there anything else missed?

[xiang90]

oh. or there is a bug in the recovery/restore logic.