No, I don't think we ever did get a definitive reproduction of #36431. We thought we did at one point, but it turned out to be #40600.

We see in results.edn that timeline 204 detected a linearizability violation. We see in independent/204/results.edn that this timeline failed with the following invalid sequence of operations:

:final-paths
([{:op
   {:process 24,
    :type :ok,
    :f :txn,
    :value
    [[:write 2 4]
     [:write 4 8]
     [:write 3 9]
     [:write 0 5]
     [:write 1 3]],
    :index 12332,
    :time 29902845586},
   :model {2 4, 4 8, 3 9, 0 5, 1 3}}
  {:op
   {:process 25,
    :type :ok,
    :f :txn,
    :value [[:read 4 1]],
    :index 12346,
    :time 29927838442},
   :model {:msg "8≠1"}}]
 [{:op
   {:process 24,
    :type :ok,
    :f :txn,
    :value
    [[:write 2 4]
     [:write 4 8]
     [:write 3 9]
     [:write 0 5]
     [:write 1 3]],
    :index 12332,
    :time 29902845586},
   :model {2 4, 4 8, 3 9, 0 5, 1 3}}
  {:op
   {:process 24,
    :type :invoke,
    :f :txn,
    :value [[:write 2 7] [:write 4 6]],
    :index 12339,
    :time 29915064692},
   :model {2 7, 4 6, 3 9, 0 5, 1 3}}
  {:op
   {:process 25,
    :type :ok,
    :f :txn,
    :value [[:read 4 1]],
    :index 12346,
    :time 29927838442},
   :model {:msg "6≠1"}}]),

What we see here is that we have a read txn in process 25 that is observing a value of 1 in register 4. We have a write that happens before this read in process 24 and writes a value of 8 to register 4. We also have a write that happens concurrently to this read in process 24 and writes a value of 6 to register 4. So it would be valid for the read in process 25 to see 8 or 6 in register 4, but not 1. So this is a stale read.

We can see this more easily in independent/204/timeline.html:

Circled in red, we see the two writes, either of which would be valid. Circled in green, we see the read. Circled in purple, we see the write that the read observes.

While we can't say definitively that this is #37394, there's also nothing that proves that it's not that issue. We're running the store-skews, which we expect will help trigger this bug. We also see that both ignored writes were second in their batch, meaning that they would have undergone async intent resolution (a necessary condition of #37394).

roachtest: jepsen/multi-register/strobe-skews failed [likely #36431] #49360

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions