Good luck on your test! As for the batching within Avro and by Kafka Producer, here are my thoughts without any empirical proof. There is a certain amount of overhead in terms of execution AND bytes in converting a request record into Avro and producing (generating) a Kafka message out of it. For requests of size 100-200 bytes, that can be a substantial amount - especially the fact that you will be bundling the Avro schema for each request in its Kafka message.
By batching the requests, you are significantly amortizing that overhead across many rows. From: Matt Daum <m...@setfive.com> Date: Monday, March 5, 2018 at 5:54 AM To: "Thakrar, Jayesh" <jthak...@conversantmedia.com> Cc: "users@kafka.apache.org" <users@kafka.apache.org> Subject: Re: Kafka Setup for Daily counts on wide array of keys Thanks for the suggestions! It does look like it's using local RocksDB stores for the state info by default. Will look into using an external one. As for the "millions of different values per grouped attribute" an example would be assume on each requests there is a parameters "X" which at the end of each day I want to know the counts per unique value, it could have 100's of millions of possible values. I'll start to hopefully work this week on an initial test of everything and will report back. A few last questions if you have the time: - For the batching of the AVRO files, would this be different than the Producer batching? - Any other things you'd suggest looking out for as gotcha's or configurations that probably will be good to tweak further? Thanks! Matt On Sun, Mar 4, 2018 at 11:23 PM, Thakrar, Jayesh <jthak...@conversantmedia.com<mailto:jthak...@conversantmedia.com>> wrote: BTW - I did not mean to rule-out Aerospike as a possible datastore. Its just that I am not familiar with it, but surely looks like a good candidate to store the raw and/or aggregated data, given that it also has a Kafka Connect module. From: "Thakrar, Jayesh" <jthak...@conversantmedia.com<mailto:jthak...@conversantmedia.com>> Date: Sunday, March 4, 2018 at 9:25 PM To: Matt Daum <m...@setfive.com<mailto:m...@setfive.com>> Cc: "users@kafka.apache.org<mailto:users@kafka.apache.org>" <users@kafka.apache.org<mailto:users@kafka.apache.org>> Subject: Re: Kafka Setup for Daily counts on wide array of keys I don’t have any experience/knowledge on the Kafka inbuilt datastore, but believe thatfor some portions of streaming Kafka uses (used?) RocksDB to locally store some state info in the brokers. Personally I would use an external datastore. There's a wide choice out there - regular key-value stores like Cassandra, ScyllaDB, RocksDB, timeseries key-value stores like InfluxDB to regular RDBMSes. If you have hadoop in the picture, its even possible to bypass a datastore completely (if appropriate) and store the raw data on HDFS organized by (say) date+hour by using periodic (minute to hourly) extract jobs and store data in hive-compatible directory structure using ORC or Parquet. The reason for shying away from NoSQL datastores is their tendency to do compaction on data which leads to unnecessary reads and writes (referred to as write-amplification). With periodic jobs in Hadoop, you (usually) write your data once only. Ofcourse with that approach you loose the "random/keyed access" to the data, but if you are only interested in the aggregations across various dimensions, those can be stored in a SQL/NoSQL datastore. As for "having millions of different values per grouped attribute" - not sure what you mean by them. Is it that each record has some fields that represent different kinds of attributes and that their domain can have millions to hundreds of millions of values? I don't think that should matter. From: Matt Daum <m...@setfive.com<mailto:m...@setfive.com>> Date: Sunday, March 4, 2018 at 2:39 PM To: "Thakrar, Jayesh" <jthak...@conversantmedia.com<mailto:jthak...@conversantmedia.com>> Cc: "users@kafka.apache.org<mailto:users@kafka.apache.org>" <users@kafka.apache.org<mailto:users@kafka.apache.org>> Subject: Re: Kafka Setup for Daily counts on wide array of keys Thanks! For the counts I'd need to use a global table to make sure it's across all the data right? Also having millions of different values per grouped attribute will scale ok? On Mar 4, 2018 8:45 AM, "Thakrar, Jayesh" <jthak...@conversantmedia.com<mailto:jthak...@conversantmedia.com>> wrote: Yes, that's the general design pattern. Another thing to look into is to compress the data. Now Kafka consumer/producer can already do it for you, but we choose to compress in the applications due to a historic issue that drgraded performance, although it has been resolved now. Also, just keep in mind that while you do your batching, kafka producer also tries to batch msgs to Kafka, and you will need to ensure you have enough buffer memory. However that's all configurable. Finally ensure you have the latest java updates and have kafka 0.10.2 or higher. Jayesh ________________________________ From: Matt Daum <m...@setfive.com<mailto:m...@setfive.com>> Sent: Sunday, March 4, 2018 7:06:19 AM To: Thakrar, Jayesh Cc: users@kafka.apache.org<mailto:users@kafka.apache.org> Subject: Re: Kafka Setup for Daily counts on wide array of keys We actually don't have a kafka cluster setup yet at all. Right now just have 8 of our application servers. We currently sample some impressions and then dedupe/count outside at a different DC, but are looking to try to analyze all impressions for some overall analytics. Our requests are around 100-200 bytes each. If we lost some of them due to network jitter etc. it would be fine we're trying to just get overall a rough count of each attribute. Creating batched messages definitely makes sense and will also cut down on the network IO. We're trying to determine the required setup for Kafka to do what we're looking to do as these are physical servers so we'll most likely need to buy new hardware. For the first run I think we'll try it out on one of our application clusters that get a smaller amount traffic (300-400k req/sec) and run the kafka cluster on the same machines as the applications. So would the best route here be something like each application server batches requests, send it to kafka, have a stream consumer that then tallies up the totals per attribute that we want to track, output that to a new topic, which then goes to a sink to either a DB or something like S3 which then we read into our external DBs? Thanks! On Sun, Mar 4, 2018 at 12:31 AM, Thakrar, Jayesh <jthak...@conversantmedia.com<mailto:jthak...@conversantmedia.com>> wrote: Matt, If I understand correctly, you have an 8 node Kafka cluster and need to support about 1 million requests/sec into the cluster from source servers and expect to consume that for aggregation. How big are your msgs? I would suggest looking into batching multiple requests per single Kafka msg to achieve desired throughput. So e.g. on the request receiving systems, I would suggest creating a logical avro file (byte buffer) of say N requests and then making that into one Kafka msg payload. We have a similar situation (https://www.slideshare.net/JayeshThakrar/apacheconflumekafka2016) and found anything from 4x to 10x better throughput with batching as compared to one request per msg. We have different kinds of msgs/topics and the individual "request" size varies from about 100 bytes to 1+ KB. On 3/2/18, 8:24 AM, "Matt Daum" <m...@setfive.com<mailto:m...@setfive.com>> wrote: I am new to Kafka but I think I have a good use case for it. I am trying to build daily counts of requests based on a number of different attributes in a high throughput system (~1 million requests/sec. across all 8 servers). The different attributes are unbounded in terms of values, and some will spread across 100's of millions values. This is my current through process, let me know where I could be more efficient or if there is a better way to do it. I'll create an AVRO object "Impression" which has all the attributes of the inbound request. My application servers then will on each request create and send this to a single kafka topic. I'll then have a consumer which creates a stream from the topic. From there I'll use the windowed timeframes and groupBy to group by the attributes on each given day. At the end of the day I'd need to read out the data store to an external system for storage. Since I won't know all the values I'd need something similar to the KVStore.all() but for WindowedKV Stores. This appears that it'd be possible in 1.1 with this commit: https://github.com/apache/kafka/commit/1d1c8575961bf6bce7decb049be7f10ca76bd0c5 . Is this the best approach to doing this? Or would I be better using the stream to listen and then an external DB like Aerospike to store the counts and read out of it directly end of day. Thanks for the help! Daum
