[
https://issues.apache.org/jira/browse/IGNITE-20911?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aleksandr Polovtcev updated IGNITE-20911:
-----------------------------------------
Description:
Test added in IGNITE-20910 fails with a {{PrimaryReplicaMissException}}. This
happens, because after a node has been restarted, its Leases start to expire
(which is a separate bug, by the way), which leads to the following chain of
events:
# When a Tuple gets inserted, it ends up in the
{{InternalTableImpl#enlistWithRetry}} method;
# Since the Primary Replica Lease has expired,
{{InternalTableImpl#trackingInvoke}} will fail with a
{{PrimaryReplicaMissException}}, which means that we should re-enlist the
transaction again;
# When calling {{InternalTableImpl#enlist}} we acquire the new Lease by calling
{{placementDriver.awaitPrimaryReplica}};
# This information is then passed to {{ReadWriteTransactionImpl#enlist}},
which, however, effectively ignores this new information, because it calls
{{enlisted.computeIfAbsent(tablePartitionId, k -> nodeAndTerm)}}, which means
that if there has been a previously enlisted transaction (and there was, since
we are in the process of re-enlisting a transaction), then this new Lease
information will be ignored and the previous (outdated) information will be
used instead, thus making this re-trying mechanism completely useless.
> Transaction retry works incorrectly
> -----------------------------------
>
> Key: IGNITE-20911
> URL: https://issues.apache.org/jira/browse/IGNITE-20911
> Project: Ignite
> Issue Type: Bug
> Reporter: Aleksandr Polovtcev
> Priority: Critical
> Labels: ignite-3
>
> Test added in IGNITE-20910 fails with a {{PrimaryReplicaMissException}}. This
> happens, because after a node has been restarted, its Leases start to expire
> (which is a separate bug, by the way), which leads to the following chain of
> events:
> # When a Tuple gets inserted, it ends up in the
> {{InternalTableImpl#enlistWithRetry}} method;
> # Since the Primary Replica Lease has expired,
> {{InternalTableImpl#trackingInvoke}} will fail with a
> {{PrimaryReplicaMissException}}, which means that we should re-enlist the
> transaction again;
> # When calling {{InternalTableImpl#enlist}} we acquire the new Lease by
> calling {{placementDriver.awaitPrimaryReplica}};
> # This information is then passed to {{ReadWriteTransactionImpl#enlist}},
> which, however, effectively ignores this new information, because it calls
> {{enlisted.computeIfAbsent(tablePartitionId, k -> nodeAndTerm)}}, which means
> that if there has been a previously enlisted transaction (and there was,
> since we are in the process of re-enlisting a transaction), then this new
> Lease information will be ignored and the previous (outdated) information
> will be used instead, thus making this re-trying mechanism completely useless.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
