Vladimir Steshin created IGNITE-21603:
-----------------------------------------
Summary: Incorrect bacward connection chech with loopback address
Key: IGNITE-21603
URL: https://issues.apache.org/jira/browse/IGNITE-21603
Project: Ignite
Issue Type: Improvement
Reporter: Vladimir Steshin
We may skip backward connection check to a previous node if it has the same
loopback address as the current node.
Consider:
# Neither _IgniteConfiguration#setLocalHost()_ or
_TcpDiscoverySpi#setLocalAddress()_ is set. Or the localhost parameter is set
to "_0.0.0.0_".
# Nodes start on different hosts. All the available host addresses are resolved
and
# Among the other addresses, all nodes get the loopback address
"127.0.0.1:47500" (47500 is the default tcp discovery port).
# Cluster starts and works. But
# Some node N (A) decides the connection to node N+1 (B) is lost and tries to
connect to node N+2 (C) and sends _TcpDiscoveryHandshakeRequest_.
# Before C accepts incoming A's connection, it decides to check B and pings it
with _ServerImpl#checkConnection(List<InetSocketAddress> addrs, int timeout)_
# Around here, the network is restored, and A can now connect to B anew.
# "_127.0.0.1:47500_" is last in _List<InetSocketAddress>_ addrs by
_IgniteUtils#inetAddressesComparator(boolean sameHost)_. But the connect
attempts in _checkConnection(...)_ are parallel. "_127.0.0.1:47500_" answers
first.
# C sees it can connect to "_127.0.0.1:47500_" and chooses it as the alive
address of B. Other pings to rest of B's addresses are ignored.
# But "_127.0.0.1:47500_" is one of C's addresses. C realizes it pinged itself
and decides that B is not reachable:
{code:java}
// If local node was able to connect to previous, confirm that it's
alive.
ok = liveAddr != null && (!liveAddr.getAddress().isLoopbackAddress()
|| !locNode.socketAddresses().contains(liveAddr));
{code}
_emphasized text_C accepts connection from A and answers with
TcpDiscoveryHandshakeResponse#previousNodeAlive() == false
# But B is ok now. But A connects to C and B is kicked from the ring.
The problem is that C ping itself by B's address "_127.0.0.1:47500_"
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
