Steven Schlansker created KAFKA-14942:
-----------------------------------------
Summary: CopyOnWriteMap implements ConcurrentMap but does not
implement required default methods
Key: KAFKA-14942
URL: https://issues.apache.org/jira/browse/KAFKA-14942
Project: Kafka
Issue Type: Bug
Components: clients
Affects Versions: 3.4.0
Reporter: Steven Schlansker
Hi Kafka team,
I was reading through the kafka-clients CopyOnWriteMap while investigating a
problem in a different library, and I think it is declaring that it is a
ConcurrentMap but does not completely implement that interface.
In particular, it inherits e.g. computeIfAbsent as a default method from Map,
which is noted to be a non-atomic implementation, and is not synchronized in
any way. I think this can lead to a reader experiencing a map whose contents
are not consistent with any serial execution of write ops.
Consider a thread T1 which calls computeIfAbsent("a", _ -> "1")
T1 computeIfAbsent calls get("a") and observes null, and is then pre-empted
T2 calls put("a", "2"), which copies the (empty) backing map and stores \{"a":
"2"}
T1 computeIfAbsent then wakes up, still thinking the value is null, and calls
put("a", "1").
This leads to the map finishing with the contents \{"a":"1"}, while any serial
execution of these two operations should always finish with \{"a":"2"}.
I think CopyOnWriteMap should either re-implement all mutating default methods
at least as synchronized. If this is a special internal map and we know those
will never be called, perhaps they should throw UnsupportedOperationException
or at least document the class as not a complete and proper implementation.
Thank you for your consideration.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
