[
https://issues.apache.org/jira/browse/IGNITE-13016?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vladimir Steshin updated IGNITE-13016:
--------------------------------------
Description:
We should fix several drawbacks in the backward checking of failed node. They
prolongs node failure detection upto:
ServerImpl.CON_CHECK_INTERVAL + 2 * IgniteConfiguretion.failureDetectionTimeout
+ 300ms.
See:
* ‘_NodeFailureResearch.patch_'. It creates test 'FailureDetectionResearch'
which emulates long answears on a failed node and measures failure detection
delays.
* '_NodeFailureResearch.txt_' - results of the test.
* 'NodeFailureResearch_fixed.txt' - results of the test after this fix.
* '_WostCaseStepByStep.txt_' - description how the worst case happens.
*Suggestions:*
1) We should replace hardcoded timeout 100ms with a parameter like
failureDetectionTimeout:
{code:java}
private boolean ServerImpl.isConnectionRefused(SocketAddress addr) {
...
sock.connect(addr, 100); // Make it rely on failureDetectionTimeout.
...
}
{code}
2) Any negative result of the connection checking should be considered as node
failed. Currently, we look only at refused connection. Any other exceptions,
including a timeout, are treated as living connection:
{code:java}
private boolean ServerImpl.isConnectionRefused(SocketAddress addr) {
...
catch (ConnectException e) {
return true;
}
catch (IOException e) {
return false; // Make any error mean lost connection.
}
return false;
}
{code}
3) Maximal interval to check previous node should rely on actual failure
detection timeout:
{code:java}
TcpDiscoveryHandshakeResponse res = new TcpDiscoveryHandshakeResponse(...);
...
// We got message from previous in less than double connection check
interval.
boolean ok = rcvdTime + CON_CHECK_INTERVAL * 2 >= now; // Here should be a
timeout of failure detection.
if (ok) {
// Check case when previous node suddenly died. This will speed up
// node failing.
...
}
res.previousNodeAlive(ok);
{code}
was:
We should fix several drawbacks in the backward checking of failed node. They
prolongs node failure detection upto:
ServerImpl.CON_CHECK_INTERVAL + 2 * IgniteConfiguretion.failureDetectionTimeout
+ 300ms.
See:
* ‘_NodeFailureResearch.patch_'. It creates test 'FailureDetectionResearch'
which emulates long answears on a failed node and measures failure detection
delays.
* '_NodeFailureResearch.txt_' - results of the test.
* 'NodeFailureResearch_fixed.txt' - results of the test after this fix.
* '_WostCaseStepByStep.txt_' - description how the worst case happens.
*Suggestions:*
1) We should replace hardcoded timeout 100ms with a parameter like
failureDetectionTimeout:
{code:java}
private boolean ServerImpl.isConnectionRefused(SocketAddress addr) {
...
sock.connect(addr, 100); // Make it rely on failureDetectionTimeout.
...
}
{code}
2) Any negative result of the connection checking should be considered as node
failed. Currently, we look only at refused connection. Any other exceptions,
including a timeout, are treated as living connection:
{code:java}
private boolean ServerImpl.isConnectionRefused(SocketAddress addr) {
...
catch (ConnectException e) {
return true;
}
catch (IOException e) {
return false; // Make any error mean lost connection.
}
return false;
}
{code}
3) Maximal interval to check previous node should rely on actual failure
detection timeout:
{code:java}
TcpDiscoveryHandshakeResponse res = new TcpDiscoveryHandshakeResponse(...);
...
// We got message from previous in less than double connection check
interval.
boolean ok = rcvdTime + CON_CHECK_INTERVAL * 2 >= now; // Here should be a
timeout of failure detection.
if (ok) {
// Check case when previous node suddenly died. This will speed up
// node failing.
...
}
res.previousNodeAlive(ok);
{code}
4) Remove hardcoded sleep of 200ms when marking previous node alive:
{code:java}
ServerImpl.CrossRingMessageSendState.markLastFailedNodeAlive(){
...
try {
Thread.sleep(200);
}
catch (InterruptedException e) {
Thread.currentThread().interrupt();
}
...
}
{code}
> Fix backward checking of failed node.
> -------------------------------------
>
> Key: IGNITE-13016
> URL: https://issues.apache.org/jira/browse/IGNITE-13016
> Project: Ignite
> Issue Type: Sub-task
> Reporter: Vladimir Steshin
> Assignee: Vladimir Steshin
> Priority: Major
> Labels: iep-45
> Attachments: FailureDetectionResearch.txt,
> FailureDetectionResearch_fixed.txt, NodeFailureResearch.patch,
> WostCaseStepByStep.txt
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> We should fix several drawbacks in the backward checking of failed node. They
> prolongs node failure detection upto:
> ServerImpl.CON_CHECK_INTERVAL + 2 *
> IgniteConfiguretion.failureDetectionTimeout + 300ms.
> See:
> * ‘_NodeFailureResearch.patch_'. It creates test 'FailureDetectionResearch'
> which emulates long answears on a failed node and measures failure detection
> delays.
> * '_NodeFailureResearch.txt_' - results of the test.
> * 'NodeFailureResearch_fixed.txt' - results of the test after this fix.
> * '_WostCaseStepByStep.txt_' - description how the worst case happens.
> *Suggestions:*
> 1) We should replace hardcoded timeout 100ms with a parameter like
> failureDetectionTimeout:
> {code:java}
> private boolean ServerImpl.isConnectionRefused(SocketAddress addr) {
> ...
> sock.connect(addr, 100); // Make it rely on failureDetectionTimeout.
> ...
> }
> {code}
> 2) Any negative result of the connection checking should be considered as
> node failed. Currently, we look only at refused connection. Any other
> exceptions, including a timeout, are treated as living connection:
> {code:java}
> private boolean ServerImpl.isConnectionRefused(SocketAddress addr) {
> ...
> catch (ConnectException e) {
> return true;
> }
> catch (IOException e) {
> return false; // Make any error mean lost connection.
> }
> return false;
> }
> {code}
> 3) Maximal interval to check previous node should rely on actual failure
> detection timeout:
> {code:java}
> TcpDiscoveryHandshakeResponse res = new TcpDiscoveryHandshakeResponse(...);
> ...
> // We got message from previous in less than double connection check
> interval.
> boolean ok = rcvdTime + CON_CHECK_INTERVAL * 2 >= now; // Here should be a
> timeout of failure detection.
> if (ok) {
> // Check case when previous node suddenly died. This will speed up
> // node failing.
> ...
> }
> res.previousNodeAlive(ok);
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
