Konstantin Osipov c43736c53b test: remove a flaky test case ...

When we append an entry to a list with the same user-defined
timestamp, the behaviour is actually undefined. If the append
is processed by the same coordinator as the one that accepted
the existing entry, then it gets the same timeuuid as the list key,
and replaces (potentially) the existing list valiue. Then it
gets a timeuuid which maybe both larger and smaller than the existing
key's timeuuid, and then turns either to an append or a prepend.

The part of the timestamp responsible for the result is the shard
id's spoof node address implemented in scope of fixing Scylla's
timeuuid uniqueness. When the test was implemented all spoof node ids
where 0 on all shards and all coordinators. Later the difference
in behaviour was dormant because cql_repl would always execute
the append on the same shard.

We could fix Scylla to use a zero spoof node address in case a user
timestamp is supplied, but the purpose of this is unclear, it
may actually be to the contrary of the user's intent.

2022-05-25 20:26:42 +03:00

3.7 KiB

Raw Blame History

View Raw