False positive on lost update

[bsbds]

Hi, I recently use jepsen to test our database, and I suspect that elle sometimes generates false-positive lost update anomalies during history analyses.

To reproduce, consider the following jepsen history:

(def h (h/history
        [{:index 0 :time 0 :type :invoke :process 0 :f :txn :value [[:r 1 nil] [:w 1 2]]}
         {:index 1 :time 1 :type :invoke :process 1 :f :txn :value [[:r 1 nil] [:w 1 3]]}
         {:index 2 :time 2 :type :ok :process 0 :f :txn :value [[:r 1 2] [:w 1 2]]}
         {:index 3 :time 3 :type :ok :process 1 :f :txn :value [[:r 1 2] [:w 1 3]]}]))

1. The first txn writes the exact same value as the initial value.
2. The second txn writes a different value to the key.

It appears that elle would interpret the second txn as having read the initial value and thereby generating a false-positive lost update anomaly.

[aphyr]

Can I ask how you got this history? Elle initial values are always supposed to be nil, not 2. If you're getting a 2 out of thin air, something is seriously wrong with this database and Elle should be complaining about it.

[bsbds]

Can I ask how you got this history? Elle initial values are always supposed to be nil, not 2. If you're getting a 2 out of thin air, something is seriously wrong with this database and Elle should be complaining about it.

Sorry, I realize I made a mistake. I assumed the key had an initial value, but the same value shouldn't be written by more than one operation. Thanks for the clarification!