[
https://issues.apache.org/jira/browse/IGNITE-23391?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Roman Puchkovskiy updated IGNITE-23391:
---------------------------------------
Description:
Let's imagine the following scenario:
# Cluster consists of nodes A, B; MG is on A, CMG is on B
# A is turned off
# User initiates MG repair via node B
# A starts
# 3 and 4 happen concurrently; at step 3, B does not see A in the physical
topology, so A does not participate in the repair, but A's JoinRequest is
processed before node B stops
# As a result, A has started the Metastorage and thinks it's in the old cluster
# At the same time, B restarts, gets CMG reset and MG repaired and it now also
has a functioning Metastorage
There is a split brain in the Metastorage, even though both leaders of the new
Metastorage used the same CMG for join.
We should make it impossible for node B to both allow a connection with A
(allowing it to join) and not see it in the physical topology when initiating a
cluster reset.
> Linearize addition to physical topology wrt cluster reset initiation
> --------------------------------------------------------------------
>
> Key: IGNITE-23391
> URL: https://issues.apache.org/jira/browse/IGNITE-23391
> Project: Ignite
> Issue Type: Improvement
> Reporter: Roman Puchkovskiy
> Assignee: Roman Puchkovskiy
> Priority: Major
> Labels: iep-128, ignite-3
>
> Let's imagine the following scenario:
> # Cluster consists of nodes A, B; MG is on A, CMG is on B
> # A is turned off
> # User initiates MG repair via node B
> # A starts
> # 3 and 4 happen concurrently; at step 3, B does not see A in the physical
> topology, so A does not participate in the repair, but A's JoinRequest is
> processed before node B stops
> # As a result, A has started the Metastorage and thinks it's in the old
> cluster
> # At the same time, B restarts, gets CMG reset and MG repaired and it now
> also has a functioning Metastorage
> There is a split brain in the Metastorage, even though both leaders of the
> new Metastorage used the same CMG for join.
> We should make it impossible for node B to both allow a connection with A
> (allowing it to join) and not see it in the physical topology when initiating
> a cluster reset.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
