That's the whole problem from my perspective: it stays CONNECTED.  It never
becomes DISCONNECTED.  You can't delete it from the API in 1.2.0.

That's why I said it was a single point of failure.  The exact semantics of
calling it a single point of failure might be debatable, but the fact that
the cluster can't be modified and/or gracefully shutdown (afaik) is what I
was referring to.

On Fri, May 19, 2017 at 12:40 PM, Joe Witt <[email protected]> wrote:

> I believe at the state you describe that down node is now considered
> disconnected.  The cluster behavior prohibits you from making changes when
> it knows not all members of the cluster cannot honor the change.  If you
> are sure you want to make the changes anyway and move on without that node
> you should be able to remove it/delete it from the cluster.  Now you have a
> cluster of two connected nodes and you can make changes.
>
> On May 19, 2017 12:23 PM, "Neil Derraugh" <neil.derraugh@
> intellifylearning.com> wrote:
>
>> That's fair.  But for the sake of total clarity on my own part, after one
>> of these disaster scenarios with a newly quorum-elected primary things
>> cannot be driven through the UI and at least through parts the REST API.
>>
>> I just ran through the following.  We have 3 nodes A, B, C with A
>> primary, and A becomes unreachable without first disconnecting.  Then B and
>> C may (I haven't verified) continue operating the flow they had in the
>> clusters' last "good" state.  But they do elect a new primary, as per the
>> REST nifi-api/controller/cluster response.  But now the flow can't be
>> changed, and in some cases it can't be reported on either, i.e. some GETs
>> fail, like nifi-api/flow/process-groups/root.
>>
>> Are we describing the same behavior?
>>
>> On Fri, May 19, 2017 at 11:12 AM, Joe Witt <[email protected]> wrote:
>>
>>> If there is no longer a quorum then we cannot drive things from the UI
>>> but the cluster remaining is in tact from a functioning point of view
>>> other than being able to assign a primary to handle the one-off items.
>>>
>>> On Fri, May 19, 2017 at 11:04 AM, Neil Derraugh
>>> <[email protected]> wrote:
>>> > Hi Joe,
>>> >
>>> > Maybe I'm missing something, but if the primary node suffers a network
>>> > partition or container/vm/machine loss or becomes otherwise
>>> unreachable then
>>> > the cluster is unusable, at least from the UI.
>>> >
>>> > If that's not so please correct me.
>>> >
>>> > Thanks,
>>> > Neil
>>> >
>>> > On Thu, May 18, 2017 at 9:56 PM, Joe Witt <[email protected]> wrote:
>>> >>
>>> >> Neil,
>>> >>
>>> >> Want to make sure I understand what you're saying.  What are stating
>>> >> is a single point of failure?
>>> >>
>>> >> Thanks
>>> >> Joe
>>> >>
>>> >> On Thu, May 18, 2017 at 5:27 PM, Neil Derraugh
>>> >> <[email protected]> wrote:
>>> >> > Thanks for the insight Matt.
>>> >> >
>>> >> > It's a disaster recovery issue.  It's not something I plan on doing
>>> on
>>> >> > purpose.  It seems it is a single point of failure unfortunately.
>>> I can
>>> >> > see
>>> >> > no other way to resolve the issue other than to blow everything
>>> away and
>>> >> > start a new cluster.
>>> >> >
>>> >> > On Thu, May 18, 2017 at 2:49 PM, Matt Gilman <
>>> [email protected]>
>>> >> > wrote:
>>> >> >>
>>> >> >> Neil,
>>> >> >>
>>> >> >> Disconnecting a node prior to removal is the correct process. It
>>> >> >> appears
>>> >> >> that the check was lost going from 0.x to 1.x. Folks reported this
>>> JIRA
>>> >> >> [1]
>>> >> >> indicating that deleting a connected node did not work. This
>>> process
>>> >> >> does
>>> >> >> not work because the node needs to be disconnected first. The JIRA
>>> was
>>> >> >> addressed by restoring the check that a node is disconnected prior
>>> to
>>> >> >> deletion.
>>> >> >>
>>> >> >> Hopefully the JIRA I filed earlier today [2] will address the
>>> phantom
>>> >> >> node
>>> >> >> you were seeing. Until then, can you update your workaround to
>>> >> >> disconnect
>>> >> >> the node in question prior to deletion?
>>> >> >>
>>> >> >> Thanks
>>> >> >>
>>> >> >> Matt
>>> >> >>
>>> >> >> [1] https://issues.apache.org/jira/browse/NIFI-3295
>>> >> >> [2] https://issues.apache.org/jira/browse/NIFI-3933
>>> >> >>
>>> >> >> On Thu, May 18, 2017 at 12:29 PM, Neil Derraugh
>>> >> >> <[email protected]> wrote:
>>> >> >>>
>>> >> >>> Pretty sure this is the problem I was describing in the "Phantom
>>> Node"
>>> >> >>> thread recently.
>>> >> >>>
>>> >> >>> If I kill non-primary nodes the cluster remains healthy despite
>>> the
>>> >> >>> lost
>>> >> >>> nodes.  The terminated nodes end up with a DISCONNECTED status.
>>> >> >>>
>>> >> >>> If I kill the primary it winds up with a CONNECTED status, but a
>>> new
>>> >> >>> primary/cluster coordinator gets elected too.
>>> >> >>>
>>> >> >>> Additionally it seems in 1.2.0 that the REST API no longer support
>>> >> >>> deleting a node in a CONNECTED state (Cannot remove Node with ID
>>> >> >>> 1780fde7-c2f4-469c-9884-fe843eac5b73 because it is not
>>> disconnected,
>>> >> >>> current
>>> >> >>> state = CONNECTED).  So right now I don't have a workaround and
>>> have
>>> >> >>> to kill
>>> >> >>> all the nodes and start over.
>>> >> >>>
>>> >> >>> On Thu, May 18, 2017 at 11:20 AM, Mark Payne <
>>> [email protected]>
>>> >> >>> wrote:
>>> >> >>>>
>>> >> >>>> Hello,
>>> >> >>>>
>>> >> >>>> Just looking through this thread now. I believe that I
>>> understand the
>>> >> >>>> problem. I have updated the JIRA with details about what I think
>>> is
>>> >> >>>> the
>>> >> >>>> problem and a potential remedy for the problem.
>>> >> >>>>
>>> >> >>>> Thanks
>>> >> >>>> -Mark
>>> >> >>>>
>>> >> >>>> > On May 18, 2017, at 9:49 AM, Matt Gilman <
>>> [email protected]>
>>> >> >>>> > wrote:
>>> >> >>>> >
>>> >> >>>> > Thanks for the additional details. They will be helpful when
>>> >> >>>> > working
>>> >> >>>> > the JIRA. All nodes, including the coordinator, heartbeat to
>>> the
>>> >> >>>> > active
>>> >> >>>> > coordinator. This means that the coordinator effectively
>>> heartbeats
>>> >> >>>> > to
>>> >> >>>> > itself. It appears, based on your log messages, that this is
>>> not
>>> >> >>>> > happening.
>>> >> >>>> > Because no heartbeats were receive from any node, the lack of
>>> >> >>>> > heartbeats
>>> >> >>>> > from the terminated node is not considered.
>>> >> >>>> >
>>> >> >>>> > Matt
>>> >> >>>> >
>>> >> >>>> > Sent from my iPhone
>>> >> >>>> >
>>> >> >>>> >> On May 18, 2017, at 8:30 AM, ddewaele <[email protected]>
>>> wrote:
>>> >> >>>> >>
>>> >> >>>> >> Found something interesting in the centos-b debug logging....
>>> >> >>>> >>
>>> >> >>>> >> after centos-a (the coordinator) is killed centos-b takes
>>> over.
>>> >> >>>> >> Notice how
>>> >> >>>> >> it "Will not disconnect any nodes due to lack of heartbeat"
>>> and
>>> >> >>>> >> how
>>> >> >>>> >> it still
>>> >> >>>> >> sees centos-a as connected despite the fact that there are no
>>> >> >>>> >> heartbeats
>>> >> >>>> >> anymore.
>>> >> >>>> >>
>>> >> >>>> >> 2017-05-18 12:41:38,010 INFO [Leader Election Notification
>>> >> >>>> >> Thread-2]
>>> >> >>>> >> o.apache.nifi.controller.FlowController This node elected
>>> Active
>>> >> >>>> >> Cluster
>>> >> >>>> >> Coordinator
>>> >> >>>> >> 2017-05-18 12:41:38,010 DEBUG [Leader Election Notification
>>> >> >>>> >> Thread-2]
>>> >> >>>> >> o.a.n.c.c.h.ClusterProtocolHeartbeatMonitor Purging old
>>> heartbeats
>>> >> >>>> >> 2017-05-18 12:41:38,014 INFO [Leader Election Notification
>>> >> >>>> >> Thread-1]
>>> >> >>>> >> o.apache.nifi.controller.FlowController This node has been
>>> elected
>>> >> >>>> >> Primary
>>> >> >>>> >> Node
>>> >> >>>> >> 2017-05-18 12:41:38,353 DEBUG [Heartbeat Monitor Thread-1]
>>> >> >>>> >> o.a.n.c.c.h.AbstractHeartbeatMonitor Received no new
>>> heartbeats.
>>> >> >>>> >> Will
>>> >> >>>> >> not
>>> >> >>>> >> disconnect any nodes due to lack of heartbeat
>>> >> >>>> >> 2017-05-18 12:41:41,336 DEBUG [Process Cluster Protocol
>>> Request-3]
>>> >> >>>> >> o.a.n.c.c.h.ClusterProtocolHeartbeatMonitor Received new
>>> heartbeat
>>> >> >>>> >> from
>>> >> >>>> >> centos-b:8080
>>> >> >>>> >> 2017-05-18 12:41:41,337 DEBUG [Process Cluster Protocol
>>> Request-3]
>>> >> >>>> >> o.a.n.c.c.h.ClusterProtocolHeartbeatMonitor
>>> >> >>>> >>
>>> >> >>>> >> Calculated diff between current cluster status and node
>>> cluster
>>> >> >>>> >> status as
>>> >> >>>> >> follows:
>>> >> >>>> >> Node: [NodeConnectionStatus[nodeId=centos-b:8080,
>>> state=CONNECTED,
>>> >> >>>> >> updateId=45], NodeConnectionStatus[nodeId=centos-a:8080,
>>> >> >>>> >> state=CONNECTED,
>>> >> >>>> >> updateId=42]]
>>> >> >>>> >> Self: [NodeConnectionStatus[nodeId=centos-b:8080,
>>> state=CONNECTED,
>>> >> >>>> >> updateId=45], NodeConnectionStatus[nodeId=centos-a:8080,
>>> >> >>>> >> state=CONNECTED,
>>> >> >>>> >> updateId=42]]
>>> >> >>>> >> Difference: []
>>> >> >>>> >>
>>> >> >>>> >>
>>> >> >>>> >> 2017-05-18 12:41:41,337 INFO [Process Cluster Protocol
>>> Request-3]
>>> >> >>>> >> o.a.n.c.p.impl.SocketProtocolListener Finished processing
>>> request
>>> >> >>>> >> 410e7db5-8bb0-4f97-8ee8-fc8647c54959 (type=HEARTBEAT,
>>> length=2341
>>> >> >>>> >> bytes)
>>> >> >>>> >> from centos-b:8080 in 3 millis
>>> >> >>>> >> 2017-05-18 12:41:41,339 INFO [Clustering Tasks Thread-2]
>>> >> >>>> >> o.a.n.c.c.ClusterProtocolHeartbeater Heartbeat created at
>>> >> >>>> >> 2017-05-18
>>> >> >>>> >> 12:41:41,330 and sent to centos-b:10001 at 2017-05-18
>>> >> >>>> >> 12:41:41,339;
>>> >> >>>> >> send
>>> >> >>>> >> took 8 millis
>>> >> >>>> >> 2017-05-18 12:41:43,354 INFO [Heartbeat Monitor Thread-1]
>>> >> >>>> >> o.a.n.c.c.h.AbstractHeartbeatMonitor Finished processing 1
>>> >> >>>> >> heartbeats
>>> >> >>>> >> in
>>> >> >>>> >> 93276 nanos
>>> >> >>>> >> 2017-05-18 12:41:46,346 DEBUG [Process Cluster Protocol
>>> Request-4]
>>> >> >>>> >> o.a.n.c.c.h.ClusterProtocolHeartbeatMonitor Received new
>>> heartbeat
>>> >> >>>> >> from
>>> >> >>>> >> centos-b:8080
>>> >> >>>> >> 2017-05-18 12:41:46,346 DEBUG [Process Cluster Protocol
>>> Request-4]
>>> >> >>>> >> o.a.n.c.c.h.ClusterProtocolHeartbeatMonitor
>>> >> >>>> >>
>>> >> >>>> >> Calculated diff between current cluster status and node
>>> cluster
>>> >> >>>> >> status as
>>> >> >>>> >> follows:
>>> >> >>>> >> Node: [NodeConnectionStatus[nodeId=centos-b:8080,
>>> state=CONNECTED,
>>> >> >>>> >> updateId=45], NodeConnectionStatus[nodeId=centos-a:8080,
>>> >> >>>> >> state=CONNECTED,
>>> >> >>>> >> updateId=42]]
>>> >> >>>> >> Self: [NodeConnectionStatus[nodeId=centos-b:8080,
>>> state=CONNECTED,
>>> >> >>>> >> updateId=45], NodeConnectionStatus[nodeId=centos-a:8080,
>>> >> >>>> >> state=CONNECTED,
>>> >> >>>> >> updateId=42]]
>>> >> >>>> >> Difference: []
>>> >> >>>> >>
>>> >> >>>> >>
>>> >> >>>> >>
>>> >> >>>> >>
>>> >> >>>> >> --
>>> >> >>>> >> View this message in context:
>>> >> >>>> >>
>>> >> >>>> >> http://apache-nifi-users-list.2361937.n4.nabble.com/Nifi-Clu
>>> ster-fails-to-disconnect-node-when-node-was-killed-tp1942p1950.html
>>> >> >>>> >> Sent from the Apache NiFi Users List mailing list archive at
>>> >> >>>> >> Nabble.com.
>>> >> >>>>
>>> >> >>>
>>> >> >>
>>> >> >
>>> >
>>> >
>>>
>>
>>

Reply via email to