It's per broker. Usually you run with 4-6GB of java heap. The rest is used as disk cache and it's more that 64GB seems like a sweet spot between memory cost and performance.
/svante 2018-03-01 18:30 GMT+01:00 Michal Michalski <[email protected]>: > I'm quite sure it's per broker (it's a standard way to provide > recommendation on node sizes in systems like Kafka), but you should > definitely read it in the context of the data size and traffic the cluster > has to handle. I didn't read the presentation, so not sure if it contains > such information (if it doesn't, maybe the video does?), but this context > is necessary to size Kafka properly (that includes const efficiency). To > put that in context: I've been running small Kafka cluster on AWS' > m4.xlarge instances in the past with no issues (low number of terabytes > stored in total, low single-digit thousands of messages produced per second > in peak) - I actually think it was oversized for that use case. > > On 1 March 2018 at 17:09, adrien ruffie <[email protected]> wrote: > > > Hi all, > > > > > > on the slide 5 in the following link: > > > > https://fr.slideshare.net/HadoopSummit/apache-kafka-best-practices/1 > > > > > > > > The "Memory" mentions that "24GB+ (for small) and 64GB+ (for large)" > Kafka > > Brokers > > > > but is it 24 or 64 GB spread over all brokers ? Or 24 GB for example for > > each broker ? > > > > > > Thank you very much, > > > > > > and best regards, > > > > > > Adrien > > >
