kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jay Kreps <jay.kr...@gmail.com>
Subject Re: Consumer throughput imbalance
Date Mon, 26 Aug 2013 18:37:54 GMT
Yes exactly.

Lowering queuedchunks.max shouldn't help if the problem is what I
described. That options controls how many chunks the consumer has ready in
memory for processing. But we are hypothesisizing that your problem is
actually that the individual chunks are just too large leading to the
consumer spending a long time processing from one partition before it gets
the next chunk.

-Jay


On Mon, Aug 26, 2013 at 11:18 AM, Ian Friedman <ian@flurry.com> wrote:

> Just to make sure i have this right, on the producer side we'd set
> max.message.size and then on the consumer side we'd set fetch.size? I
> admittedly didn't research how all the tuning options would affect us,
> thank you for the info. Would queuedchunks.max have any effect?
>
> --
> Ian Friedman
>
>
> On Monday, August 26, 2013 at 1:26 PM, Jay Kreps wrote:
>
> > Yeah it is always equal to the fetch size. The fetch size needs to be at
> > least equal to the max message size you have allowed on the server,
> though.
> >
> > -Jay
> >
> >
> > On Sun, Aug 25, 2013 at 10:00 PM, Ian Friedman <ian@flurry.com (mailto:
> ian@flurry.com)> wrote:
> >
> > > Jay - is there any way to control the size of the interleaved chunks?
> The
> > > performance hit would likely be negligible for us at the moment.
> > >
> > > --
> > > Ian Friedman
> > >
> > >
> > > On Sunday, August 25, 2013 at 3:11 PM, Jay Kreps wrote:
> > >
> > > > I'm still a little confused by your description of the problem. It
> might
> > > be
> > > > easier to understand if you listed out the exact things you have
> > >
> > > measured,
> > > > what you saw, and what you expected to see.
> > > >
> > > > Since you mentioned the consumer I can give a little info on how that
> > > > works. The consumer consumes from all the partitions it owns
> > > > simultaneously. The behavior is that we interleve fetched data
> chunks of
> > > > messages from each partition the consumer is processing. The chunk
> size
> > > >
> > >
> > > is
> > > > controlled by the fetch size set in the consumer. So the behavior you
> > >
> > > would
> > > > expect is that you would get a bunch of messages from one partition
> > > > followed by a bunch from another partition. The reason for doing this
> > > > instead of, say, interleving individual messages is that it is a big
> > > > performance boost--making every message an entry in a blocking queue
> > > >
> > >
> > > gives
> > > > a 5x performance hit in high-throughput cases. Perhaps this
> interleaving
> > >
> > > is
> > > > the problem?
> > > >
> > > > -Jay
> > > >
> > > >
> > > > On Sun, Aug 25, 2013 at 10:22 AM, Ian Friedman <ian@flurry.com(mailto:
> ian@flurry.com) (mailto:
> > > ian@flurry.com (mailto:ian@flurry.com))> wrote:
> > > >
> > > > > Sorry I reread what I've written so far and found that it doesn't
> state
> > > > > the actual problem very well. Let me clarify once again:
> > > > >
> > > > > The problem we're trying to solve is that we can't let messages go
> for
> > > > > unbounded amounts of time without getting processed, and it seems
> that
> > > > > something about what we're doing (which I suspect is the fact that
> > > > > consumers own several partitions but only consume from one of them
> at a
> > > > > time until it's caught up) is causing a small number of them to sit
> > > > >
> > > >
> > > >
> > >
> > > around
> > > > > for hours and hours. This is despite some consumers idling due to
> being
> > > > > fully caught up on the partitions they own. We've found that
> > > > >
> > > >
> > >
> > > requeueing the
> > > > > oldest messages (consumers ignore messages that have already been
> > > > > processed) is fairly effective in getting them to go away, but I'm
> > > > >
> > > >
> > >
> > > looking
> > > > > for a more stable solution.
> > > > >
> > > > > --
> > > > > Ian Friedman
> > > > >
> > > > >
> > > > > On Sunday, August 25, 2013 at 1:15 PM, Ian Friedman wrote:
> > > > >
> > > > > > When I said "some messages take longer than others" that may
have
> > > been
> > > > > misleading. What I meant there is that the performance of the
> entire
> > > > > application is inconsistent, mostly due to pressure from other
> > > > >
> > > >
> > >
> > > applications
> > > > > (mapreduce) on our HBase and MySQL backends. On top of that, some
> > > >
> > >
> > > messages
> > > > > just contain more data. Now I suppose what you're suggesting is
> that I
> > > > > segment my messages by the average or expected time it takes the
> > > > >
> > > >
> > >
> > > payloads
> > > > > to process, but I suspect what will happen if I do that is I will
> have
> > > > > several consumers doing nothing most of the time, and the rest of
> them
> > > > > backlogged inconsistently the same way they are now. The problem
> isn't
> > > > >
> > > >
> > >
> > > so
> > > > > much the size of the payloads but the fact that we're seeing some
> > > >
> > >
> > > messages,
> > > > > which i suspect are in partitions with lots of longer running
> > > >
> > >
> > > processing
> > > > > tasks, sit around for hours without getting consumed. That's what
> I'm
> > > > > trying to solve.
> > > > > >
> > > > > > Is there any way to "add more consumers" without actually adding
> more
> > > > > consumer JVM processes? We've hit something of a saturation point
> for
> > > > >
> > > >
> > >
> > > our
> > > > > MySQL database. Is this maybe where having multiple consumer
> threads
> > > >
> > >
> > > would
> > > > > help? If so, given that I have a singular shared processing queue
> in
> > > >
> > >
> > > each
> > > > > consumer, how would I leverage that to solve this problem?
> > > > > >
> > > > > > --
> > > > > > Ian Friedman
> > > > > >
> > > > > >
> > > > > > On Sunday, August 25, 2013 at 12:13 PM, Mark wrote:
> > > > > >
> > > > > > > I don't think it would matter as long as you separate the
> types of
> > > > > message in different topics. Then just add more consumers to the
> ones
> > > > >
> > > >
> > >
> > > that
> > > > > are slow. Am I missing something?
> > > > > > >
> > > > > > > On Aug 25, 2013, at 8:59 AM, Ian Friedman <ian@flurry.com(mailto:
> ian@flurry.com) (mailto:
> > > ian@flurry.com (mailto:ian@flurry.com)) (mailto:
> > > > > ian@flurry.com (mailto:ian@flurry.com))> wrote:
> > > > > > >
> > > > > > > > What if you don't know ahead of time how long a message
will
> > > take to
> > > > > consume?
> > > > > > > >
> > > > > > > > --
> > > > > > > > Ian Friedman
> > > > > > > >
> > > > > > > >
> > > > > > > > On Sunday, August 25, 2013 at 10:45 AM, Neha Narkhede
wrote:
> > > > > > > >
> > > > > > > > > Making producer side partitioning depend on consumer
> behavior
> > > > > might not be
> > > > > > > > > such a good idea. If consumption is a bottleneck,
changing
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > producer side
> > > > > > > > > partitioning may not help. To relieve consumption
> bottleneck,
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > you
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > may need
> > > > > > > > > to increase the number of partitions for those
topics and
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > increase
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > the
> > > > > > > > > number of consumer instances.
> > > > > > > > >
> > > > > > > > > You mentioned that the consumers take longer
to process
> certain
> > > > > kinds of
> > > > > > > > > messages. What you can do is place the messages
that
> require
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > slower
> > > > > > > > > processing in separate topics, so that you can
scale the
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> > > number of
> > > > > > > > > partitions and number of consumer instances,
for those
> messages
> > > > > > > > > independently.
> > > > > > > > >
> > > > > > > > > Thanks,
> > > > > > > > > Neha
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > On Sat, Aug 24, 2013 at 9:57 AM, Ian Friedman
<
> ian@flurry.com (mailto:ian@flurry.com)(mailto:
> > > ian@flurry.com (mailto:ian@flurry.com))(mailto:
> > > > > ian@flurry.com) (mailto:ian@flurry.com)> wrote:
> > > > > > > > >
> > > > > > > > > > Hey guys! We recently deployed our kafka
data pipeline
> > > > > application over
> > > > > > > > > > the weekend and it is working out quite
well once we
> ironed
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > out
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > all the
> > > > > > > > > > issues. There is one behavior that we've
noticed that is
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > mildly
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > troubling,
> > > > > > > > > > though not a deal breaker. We're using a
single topic
> with
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > many
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > partitions
> > > > > > > > > > (1200 total) to load balance our 300 consumers,
but what
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > seems
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > to happen is
> > > > > > > > > > that some partitions end up more backed
up than others.
> This
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > is
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > probably
> > > > > > > > > > due more to the specifics of the application
since some
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > messages
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > take much
> > > > > > > > > > longer than others to process.
> > > > > > > > > >
> > > > > > > > > > I'm thinking that the random partitioning
in the
> producer is
> > > > > unsuited to
> > > > > > > > > > our specific needs. One option I was considering
was to
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > write an
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > alternate
> > > > > > > > > > partitioner that looks at the consumer offsets
from
> zookeeper
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > (as in the
> > > > > > > > > > ConsumerOffsetChecker) and probabilistically
weights the
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > partitions by
> > > > > > > > > > their lag. Does this sound like a good idea
to anyone
> else?
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > Is
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > there a
> > > > > > > > > > better or preferably already built solution?
If anyone
> has
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > > any
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > ideas or
> > > > > > > > > > feedback I'd sincerely appreciate it.
> > > > > > > > > >
> > > > > > > > > > Thanks so much in advance.
> > > > > > > > > >
> > > > > > > > > > P.S. thanks especially to everyone who's
answered my dumb
> > > > > questions on
> > > > > > > > > > this mailing list over the past few months,
we couldn't
> have
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > > > done it
> > > > > > > > > > without you!
> > > > > > > > > >
> > > > > > > > > > --
> > > > > > > > > > Ian Friedman
> > > > > > > > > >
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > > >
> > > >
> > >
> > >
> >
> >
> >
>
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message