[ 
https://issues.apache.org/jira/browse/KAFKA-6603?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16383572#comment-16383572
 ] 

Igor Calabria commented on KAFKA-6603:
--------------------------------------

Hey, thanks for the quick reply. I just tested your code in production and it 
didn't reduce memory usage for simple aggregations. What really helped me was 
your pull request, I had some code that used the same rocksDB iterator and 
replaced it with something similar to what you did on the aggregations. code. 
The improvement was significant(especially for throughput), I'm attributing the 
extra memory usage to this inefficient iterator but I still need to do more 
testing.

> Kafka streams off heap memory usage does not match expected values from 
> configuration
> -------------------------------------------------------------------------------------
>
>                 Key: KAFKA-6603
>                 URL: https://issues.apache.org/jira/browse/KAFKA-6603
>             Project: Kafka
>          Issue Type: Bug
>          Components: streams
>    Affects Versions: 1.0.0
>            Reporter: Igor Calabria
>            Priority: Minor
>
> Hi, I have a simple aggregation pipeline that's backed by the default state 
> store(rocksdb). The pipeline works fine except that off heap the memory usage 
> is way higher than expected. Following the 
> [documention|https://docs.confluent.io/current/streams/developer-guide/config-streams.html#streams-developer-guide-rocksdb-config]
>  has some effect(memory usage is reduced) but the values don't match at all. 
> The java process is set to run with just `-Xmx300m -Xms300m`  and rocksdb 
> config looks like this
> {code:java}
> tableConfig.setCacheIndexAndFilterBlocks(true);
> tableConfig.setBlockCacheSize(1048576); //1MB
> tableConfig.setBlockSize(16 * 1024); // 16KB
> options.setTableFormatConfig(tableConfig);
> options.setMaxWriteBufferNumber(2);
> options.setWriteBufferSize(8 * 1024); // 8KB{code}
> To estimate memory usage, I'm using this formula  
> {noformat}
> (block_cache_size + write_buffer_size * write_buffer_number) * segments * 
> partitions{noformat}
> Since my topic has 25 partitions with 3 segments each(it's a windowed store), 
> off heap memory usage should be about 76MB. What I'm seeing in production is 
> upwards of 300MB, even taking in consideration  extra overhead from rocksdb 
> compaction threads, this seems a bit high (especially when the disk usage for 
> all files is just 1GB) 



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to