The indexing topology is extremely simple. Data is ingested into kafka and sent to
/apps/metron/enrichment/indexed
Errors during indexing are sent to a kafka queue called index_errors
Default installed Metron is untuned for production deployment. By far and wide, the most likely piece to require TLC from a performance perspective is the indexing layer. An index that does not keep up will back up and you will see errors in the kafka bolt. There are a few knobs to tune to get the most out of your system.
The indexing
kafka queue is a collection point from the enrichment topology. As such, make sure that the number of partitions in the kafka topic is sufficient to handle the throughput that you expect.
The enrichment topology as started by the $METRON_HOME/bin/start_elasticsearch_topology.sh
or $METRON_HOME/bin/start_solr_topology.sh
script uses a default of one executor per bolt. In a real production system, this should be customized by modifying the flux file in $METRON_HOME/flux/indexing/remote.yaml
.
parallelism
field to the bolts to give Storm a parallelism hint for the various components. Give bolts which appear to be bottlenecks (e.g. the indexing bolt) a larger hint.parallelism
field to the kafka spout which matches the number of partitions for the enrichment kafka queue.topology.workers
field for the topology.Finally, if workers and executors are new to you or you don't know where to modify the flux file, the following might be of use to you: