[
https://issues.apache.org/jira/browse/KAFKA-513?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Swapnil Ghike updated KAFKA-513:
--------------------------------
Attachment: kafka-513-v4.patch
Uploading patch v4. The comments on the merge tool are written above. Remaining
comments -
A. Logging in kafka -
11. tried to make the log format more consistent - used [%s,%d] for printing
topicAndPartition, logged correlationId when the logging statement was
concerned with leaderAndIsrRequest, printed controllerId and controllerEpoch
wherever needed. I have tried to ensure that there are no mistakes in the order
of parameters to .format() etc, but it would be helpful if you could also
scrutinize them as well.
12. Changed state change log logging level to trace, except for certain errors.
13. As you mentioned, it's an excellent idea to mention in
ControllerChannelManager whether the leaderAndIsr request is become-leader or
become-follower request. Added a change to log that. Added a change to do that
on the broker in ReplicaManager.{makeLeader,makeFollower}.
14. Included correlationId in LeaderAndIsrRequest.toString
15. Changed PartitionStateChangeLogger to StateChangeLogger everywhere.
16. Fixed the mentioned typos.
17. Added the state change log entries that you suggested.
18. As discussed offline, kept the wrapper class around Utils.Logging.
Providing a logIdent to this class will save us the trouble to specify the
broker id in every state change log entry, and it will keep the logging
consistent with the regular server logging.
19. The successful lifecycle of a state change request will look like the
following (error/discard/abort messages can be included in this sequence in
case of failures) -
On the controller - Controller %d, epoch %d sending become-leader/follower
LeaderAndIsr request with correlationId %d to broker %d for partition [%s,%d]
On a broker - [Replica Manager on Broker %d]: Handling LeaderAndIsr request
correlationId %d received from controller %d epoch %d for partition [%s,%d]
On the same broker - [Replica Manager on Broker %d]: LeaderAndIsr request
correlationId %d received from controller %d epoch %d starting the
become-leader/follower transition for partition [%s,%d]
On the same broker - [Replica Manager on Broker %d]: Completed
become-leader/follower transition for partition [%s,%d]
On the same broker - [Replica Manager on Broker %d]: Handled LeaderAndIsr
request correlationId %d received from controller %d epoch %d for partition
[%s,%d]
On the controller - Controller %d received response correlationId %d for a
request sent to broker %d
I think KafkaApis is not the right place to include a "received leaderAndIsr
request" log entry. since KafkaApis handles all types of requests. We should
rather expect to directly see a "Handling LeaderAndIsr request" log entry. On
the broker, we don't know whether the received leaderAndIsr request is a
become-leader or become-follower request, but I guess it's ok since we log that
information in the second statement as seen above and while doing that we log
correlationId.
B. I included correlationId in the abstract class RequestOrResponse, probably
all its derived classes should include a correlationId.
C. config/log4j.properties now uses a separate controller.log and a separate
state-change.log.
D. Our kafka-run-class.sh script removes quotes passed to the command line
arguments. Changed it so that the quotes can be passed as such. It's useful for
passing values containing whitespaces like "2013-03-01 16:03:43,093".
> Add state change log to Kafka brokers
> -------------------------------------
>
> Key: KAFKA-513
> URL: https://issues.apache.org/jira/browse/KAFKA-513
> Project: Kafka
> Issue Type: Sub-task
> Affects Versions: 0.8
> Reporter: Neha Narkhede
> Assignee: Swapnil Ghike
> Priority: Blocker
> Labels: p1, replication, tools
> Fix For: 0.8
>
> Attachments: kafka-513-v1.patch, kafka-513-v2.patch,
> kafka-513-v3.patch, kafka-513-v4.patch
>
> Original Estimate: 96h
> Remaining Estimate: 96h
>
> Once KAFKA-499 is checked in, every controller to broker communication can be
> modelled as a state change for one or more partitions. Every state change
> request will carry the controller epoch. If there is a problem with the state
> of some partitions, it will be good to have a tool that can create a timeline
> of requested and completed state changes. This will require each broker to
> output a state change log that has entries like
> [2012-09-10 10:06:17,280] broker 1 received request LeaderAndIsr() for
> partition [foo, 0] from controller 2, epoch 1
> [2012-09-10 10:06:17,350] broker 1 completed request LeaderAndIsr() for
> partition [foo, 0] from controller 2, epoch 1
> On controller, this will look like -
> [2012-09-10 10:06:17,198] controller 2, epoch 1, initiated state change
> request LeaderAndIsr() for partition [foo, 0]
> We need a tool that can collect the state change log from all brokers and
> create a per-partition timeline of state changes -
> [foo, 0]
> [2012-09-10 10:06:17,198] controller 2, epoch 1 initiated state change
> request LeaderAndIsr()
> [2012-09-10 10:06:17,280] broker 1 received request LeaderAndIsr() from
> controller 2, epoch 1
> [2012-09-10 10:06:17,350] broker 1 completed request LeaderAndIsr() from
> controller 2, epoch 1
> This JIRA involves adding the state change log to each broker and adding the
> tool to create the timeline
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
