[
https://issues.apache.org/jira/browse/KAFKA-6753?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Lucas Wang updated KAFKA-6753:
------------------------------
Description:
The existing controller code updates metrics after processing every event. This
can slow down event processing on the controller tremendously. In one profiling
we see that updating metrics takes nearly 100% of the CPU for the controller
event processing thread. Specifically the slowness can be attributed to two
factors:
1. Each invocation to update the metrics is expensive. Specifically trying to
calculate the offline partitions count requires iterating through all the
partitions in the cluster to check if the partition is offline; and calculating
the preferred replica imbalance count requires iterating through all the
partitions in the cluster to check if a partition has a leader other than the
preferred leader. In a large cluster, the number of partitions can be quite
large, all seen by the controller. Even if the time spent to check a single
partition is small, the accumulation effect of so many partitions in the
cluster can make the invocation to update metrics quite expensive. One might
argue that maybe the logic for processing each single partition is not
optimized, we checked the CPU percentage of leaf nodes in the profiling result,
and found that inside the loops of collection objects, e.g. the set of all
partitions, no single function dominates the processing. Hence the large number
of the partitions in a cluster is the main contributor to the slowness of one
invocation to update the metrics.
2. The invocation to update metrics is called many times when the is a high
number of events to be processed by the controller, one invocation after
processing any event.
was:
The existing controller code updates metrics after processing every event. This
can slow down event processing on the controller tremendously. In one profiling
we see that updating metrics takes nearly 100% of the CPU for the controller
event processing thread. Specifically the slowness can be attributed to two
factors:
1. Each invocation to update the metrics is expensive. Specifically trying to
calculate the offline partitions count requires iterating through all the
partitions in the cluster to check if the partition is offline; and calculating
the preferred replica imbalance count requires iterating through all the
partitions in the cluster to check if a partition has a leader other than the
preferred leader. In a large cluster, the number of partitions can be quite
large, all seen by the controller. Even if the time spent to check a single
partition is small, the accumulation effect of so many partitions in the
cluster can make the invocation to update metrics quite expensive. One might
argue that maybe the logic for processing each single partition is not
optimized, we checked the CPU percentage of leaf nodes in the profiling result,
and found that inside the loops of collection objects, e.g. the set of all
partitions, no single function dominates the processing. Hence the large number
of the partitions in a cluster is the main contributor to the slowness of one
invocation to update the metrics.
2. The invocation to update metrics is called many times when the is a high
number of events to be processed by the controller, one invocation after
processing any event.
The patch that will be submitted tries to fix bullet 2 above, i.e. reducing the
number of invocations to update metrics. Instead of updating the metrics after
processing any event, we only periodically check if the metrics needs to be
updated, i.e. once every second.
* If after the previous invocation to update metrics, there are other types of
events that changed the controller’s state, then one second later the metrics
will be updated.
* If after the previous invocation, there has been no other types of events,
then the call to update metrics can be bypassed.
> Speed up event processing on the controller
> --------------------------------------------
>
> Key: KAFKA-6753
> URL: https://issues.apache.org/jira/browse/KAFKA-6753
> Project: Kafka
> Issue Type: Improvement
> Reporter: Lucas Wang
> Assignee: Lucas Wang
> Priority: Minor
> Fix For: 2.1.0
>
> Attachments: Screen Shot 2018-04-04 at 7.08.55 PM.png
>
>
> The existing controller code updates metrics after processing every event.
> This can slow down event processing on the controller tremendously. In one
> profiling we see that updating metrics takes nearly 100% of the CPU for the
> controller event processing thread. Specifically the slowness can be
> attributed to two factors:
> 1. Each invocation to update the metrics is expensive. Specifically trying to
> calculate the offline partitions count requires iterating through all the
> partitions in the cluster to check if the partition is offline; and
> calculating the preferred replica imbalance count requires iterating through
> all the partitions in the cluster to check if a partition has a leader other
> than the preferred leader. In a large cluster, the number of partitions can
> be quite large, all seen by the controller. Even if the time spent to check a
> single partition is small, the accumulation effect of so many partitions in
> the cluster can make the invocation to update metrics quite expensive. One
> might argue that maybe the logic for processing each single partition is not
> optimized, we checked the CPU percentage of leaf nodes in the profiling
> result, and found that inside the loops of collection objects, e.g. the set
> of all partitions, no single function dominates the processing. Hence the
> large number of the partitions in a cluster is the main contributor to the
> slowness of one invocation to update the metrics.
> 2. The invocation to update metrics is called many times when the is a high
> number of events to be processed by the controller, one invocation after
> processing any event.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
