You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ratis.apache.org by "Duong (Jira)" <ji...@apache.org> on 2024/04/17 01:57:00 UTC
[jira] [Updated] (RATIS-2059) Missing reference count when putting log entries to cache on follower
[ https://issues.apache.org/jira/browse/RATIS-2059?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Duong updated RATIS-2059:
-------------------------
Description:
From RATIS-1979, when stateMachineCacheEnabled, the log entry is copied and decoupled from the original log entry reference counter.
{code:java}
final LogEntryProto removedStateMachineData = LogProtoUtils.removeStateMachineData(entry);
....
if (stateMachineCachingEnabled) {
// The stateMachineData will be cached inside the StateMachine itself.
cache.appendEntry(LogSegment.Op.WRITE_CACHE_WITH_STATE_MACHINE_CACHE,
ReferenceCountedObject.wrap(removedStateMachineData));
} else {
cache.appendEntry(LogSegment.Op.WRITE_CACHE_WITHOUT_STATE_MACHINE_CACHE, entryRef
);
}
{code}
This opens a problem because in the below cases, *removedStateMachineData* is not a copy but the original entry itself.
# When the entry is a configuration or a metadata entry.
# When the entry is a stateMachine and is replicated from leader to follower. Somehow, the state machine data ends up in *LogEntryProto.stateMachineLogEntry.logData* (and not {*}LogEntryProto.stateMachineLogEntry.stateMachineEntry.stateMachineData{*}). This makes the logic of removeStateMachineData completely skip the removal of state machine data.
#2 is very concerning and I'm not sure it is a regression from another change. It makes it inconsistent that state machine data is available applyTransaction for followers (not the case for leader).
was:
From RATIS-1979, when stateMachineCacheEnabled, the log entry is copied and decoupled from the original log entry reference counter.
# tProtos.LogEntryProto removedStateMachineData = LogProt
# oUtils.removeStateMachineData(entry);
...
if (this.stateMachineCachingEnabled) \{ this.cache.appendEntry(Op.WRITE_CACHE_WITH_STATE_MACHINE_CACHE, ReferenceCountedObject.wrap(removedStateMachineData)); } else \{ this.cache.appendEntry(Op.WRITE_CACHE_WITHOUT_STATE_MACHINE_CACHE, entryRef); }
This opens a problem because in the below cases, *removedStateMachineData* is not a copy but the original entry itself.
# When the entry is a configuration or a metadata entry.
# When the entry is a stateMachine and is replicated from leader to follower. Somehow, the state machine data ends up in *LogEntryProto.stateMachineLogEntry.logData* (and not {*}LogEntryProto.stateMachineLogEntry.stateMachineEntry.stateMachineData{*}). This makes the logic of removeStateMachineData completely skip the removal of state machine data.
#2 is very concerning and I'm not sure it is a regression from another change. It makes it inconsistent that state machine data is available applyTransaction for followers (not the case for leader).
> Missing reference count when putting log entries to cache on follower
> ---------------------------------------------------------------------
>
> Key: RATIS-2059
> URL: https://issues.apache.org/jira/browse/RATIS-2059
> Project: Ratis
> Issue Type: Sub-task
> Affects Versions: 3.1.0
> Reporter: Duong
> Priority: Major
>
> From RATIS-1979, when stateMachineCacheEnabled, the log entry is copied and decoupled from the original log entry reference counter.
> {code:java}
> final LogEntryProto removedStateMachineData = LogProtoUtils.removeStateMachineData(entry);
> ....
> if (stateMachineCachingEnabled) {
> // The stateMachineData will be cached inside the StateMachine itself.
> cache.appendEntry(LogSegment.Op.WRITE_CACHE_WITH_STATE_MACHINE_CACHE,
> ReferenceCountedObject.wrap(removedStateMachineData));
> } else {
> cache.appendEntry(LogSegment.Op.WRITE_CACHE_WITHOUT_STATE_MACHINE_CACHE, entryRef
> );
> }
> {code}
> This opens a problem because in the below cases, *removedStateMachineData* is not a copy but the original entry itself.
> # When the entry is a configuration or a metadata entry.
> # When the entry is a stateMachine and is replicated from leader to follower. Somehow, the state machine data ends up in *LogEntryProto.stateMachineLogEntry.logData* (and not {*}LogEntryProto.stateMachineLogEntry.stateMachineEntry.stateMachineData{*}). This makes the logic of removeStateMachineData completely skip the removal of state machine data.
> #2 is very concerning and I'm not sure it is a regression from another change. It makes it inconsistent that state machine data is available applyTransaction for followers (not the case for leader).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)