Ah, thanks for the great explanation. Any particular reason that the job(s) you described should not be Samza jobs?
We're started experimenting with such jobs for Druid and Elasticsearch. For Elasticsearch, the Samza job containers join the Elasticsearch cluster as transport nodes and use the Java API to push ES data nodes. Likewise for Druid, the Samza job uses the Tranquility API to schedule jobs ( https://github.com/metamx/tranquility/tree/master/src/main/scala/com/metamx/tranquility/samza ). The nice part about push versus pull is that the downstream system does not need plugins (like ES rivers) that may complicate it's configuration or destabilize the system. Cheers, Roger On Tue, Mar 31, 2015 at 10:56 AM, Felix GV <fville...@linkedin.com.invalid> wrote: > Thanks for your reply Roger! Very insightful (: > > > 6. If there was a highly-optimized and reliable way of ingesting > > partitioned streams quickly into your online serving system, would that > > help you leverage Samza more effectively? > > >> 6. Can you elaborate please? > > Sure. The feature set I have in mind is the following: > > * Provide a thinly-wrapped Kafka producer which does appropriate > partitioning and includes useful metadata (such as production timestamp, > etc.) alongside the payload. This producer would be used in the last step > of processing of a Samza topology, in order to emit to Kafka some > processed/joined/enriched data which is destined for online serving. > * Provide a consumer process which can be co-located on the same hosts > as your data serving system. This process consumes from the appropriate > partitions and checkpoints its offsets on its own. It leverages Kafka > batching and compression to make consumption very efficient. > * For each records the consumer process issues a put/insert locally to > the co-located serving process. Since this is a local operation, it is also > very cheap and efficient. > * The consumer process can also optionally throttle its insertion rate > by monitoring some performance metrics of the co-located data serving > process. For example, if the data serving process exposes a p99 latency via > JMX or other means, this can be used in a tight feedback loop to back off > if read latency degrades beyond a certain threshold. > * This ingestion platform should be easy to integrate with any > consistently-routed data serving system, by implementing some simple > interfaces to let the ingestion system understand the key-to-partition > assignment strategy, as well as the partition-to-node assignment strategy. > Optionally, a hook to access performance metrics could also be implemented > if throttling is deemed important (as described in the previous point). > * Since the consumer process lives in a separate process, the system > benefits from good isolation guarantees. The consumer process can be capped > to a low amount of heap, and its GC is inconsequential for the serving > platform. It's also possible to bounce the consumer and data serving > processes independently of each other, if need be. > > There are some more nuances and additional features which could be nice to > have, but that's the general idea. > > > It seems to me like such system would be valuable, but I'm wondering what > other people in the open-source community think, hence why I was interested > in starting this thread... > > > Thanks for your feedback! > > -F >
