Thanks for responding, Liquan. Just so I understand better, where does
the coordinator reside? On the broker side? And the coordinator knows
the heart beat interval configured on each consumer that's connected to
the broker, I am guessing. For the usecase we have, having a higher
value for heart beat works out fine for us.
I haven't yet fully thought of what I'm proposing, but would it be
wise/feasible to introduce an option which disables heart beats
altogether and instead relies on regular consumption/communication
between the consumer and broker as a heart beat? That way, the usual
consumption and offset commits from the consumer can be used as an
inference of whether the consumer is alive or dead. That way, there
won't be a need for regular heart beats being passed by the consumer.
This helps in the case where you have many consumers connected to the
broker and practically idle except for this heart beats being passed
around. One implication of this would be that if a message isn't
produced in the topics that the consumers listen to, for a while, then
there won't be any communication (isn't it?) from the consumer and hence
it might be considered dead and a re-balance triggered. Is that a
bad/worse thing though?
-Jaikiran
On Tuesday 26 April 2016 02:10 PM, Liquan Pei wrote:
Hi Jaikiran,
Thanks for the email and the detailed analysis. One reason for setting the
heartbeat interval to a lower value is for faster failure detection. On
every received heartbeat, the coordinator starts (or resets) a timer. If no
heartbeat is received when the timer expires, the coordinator marks the
member dead and signals the rest of the group that they should rejoin so
that partitions can be reassigned.
I think the trade off here is the CPU usage and how fast you want to detect
the consumer failure. Faster failure detection makes the partitions
assigned to dead consumers to assign to other consumers.
Best,
Liquan
On Tue, Apr 26, 2016 at 1:16 AM, Jaikiran Pai <jai.forums2...@gmail.com>
wrote:
We have been investigating an unreasonably high CPU usage of the Kafka
process when there's no _real_ activity going on between the consumers and
the broker. We had this issue in 0.8.x days and is exactly the same as
what's being tracked in this JIRA
https://issues.apache.org/jira/browse/KAFKA-493. We now use 0.9.0.1 (both
client libraries, new consumer APIs and the broker). However, we still see
some CPU usage which looks a bit on the higher side when there's no real
message production or consumption going on. Just connecting around 10-20
consumers on different topics of a single broker Kafka instance shows up
this issue.
All our debugging so far points to the Processor thread on the broker side
which has a high CPU usage. There are N such Processor threads, which
always are in the RUNNABLE state doing this:
"kafka-network-thread-0-PLAINTEXT-0" #21 prio=5 os_prio=0
tid=0x00007f1858c4a800 nid=0xc81 runnable [0x00007f18106cb000]
java.lang.Thread.State: RUNNABLE
at sun.nio.ch.EPollArrayWrapper.epollWait(Native Method)
at sun.nio.ch.EPollArrayWrapper.poll(EPollArrayWrapper.java:269)
at sun.nio.ch.EPollSelectorImpl.doSelect(EPollSelectorImpl.java:79)
at sun.nio.ch.SelectorImpl.lockAndDoSelect(SelectorImpl.java:86)
- locked <0x00000006c0046128> (a sun.nio.ch.Util$2)
- locked <0x00000006c0046118> (a java.util.Collections$UnmodifiableSet)
- locked <0x00000006c0046068> (a sun.nio.ch.EPollSelectorImpl)
at sun.nio.ch.SelectorImpl.select(SelectorImpl.java:97)
at org.apache.kafka.common.network.Selector.select(Selector.java:425)
at org.apache.kafka.common.network.Selector.poll(Selector.java:254)
at kafka.network.Processor.run(SocketServer.scala:413)
at java.lang.Thread.run(Thread.java:745)
From what we have narrowed down so far, this thread in itself isn't a
"culprit", since when they are no consumers connected, the CPU isn't high.
However when a consumer connects to this and just waits for messages, these
threads start playing a role in the high CPU usage. Our debugging shows
that each of these X number of consumers that connect to the broker keep
doing 2 things when they are "idle":
1) A delayed operation every Y seconds which does the auto commit of
offsets.
2) Sending heartbeats every 3 seconds to the broker
We disabled auto commits of offsets since that's the semantic we wanted.
So #1 isn't really an issue. However, #2 is. It looks like the default
heartbeat interval is 3 seconds which is too low, IMO. This translates to a
network socket operation every 3 seconds which then has to be processed by
the broker side Processor thread. If there's just a single consumer, this
doesn't make much of a difference. As soon as you add more consumers, the
Processor on the broker side has to be start processing each of these
incoming heartbeats which become too frequent. Even though the interval is
3 seconds, the incoming heartbeats to the broker can be much more frequent
when more consumers are involved since the 3 second interval is just per
consumer. So in practice there can be a heartbeat coming every second or
few milli seconds from the X consumers to this broker which can contribute
to this high CPU usage when the system is practically idle.
So coming to the real question - why is the default heart beat interval so
low - 3 seconds? We increased it to 29 seconds (just 1 second less than the
session timeout) per consumer (via consumer configs) and in addition to
disabling auto commit, these changes have improved noticeably the CPU usage.
Ideally, what would be a better value for the heart beat interval that
doesn't unnecessary flood these messages and cause the broker to continuous
process them?
-Jaikiran
