Kinesis max shard reads/sec and multiple consumers

问题

So I have a AWS Kinesis stream where I publish events for multiple consumers. It is important for most of them to receive hot data - which means that many of them will possibly poll and read the latest data at the same time. According to the AWS documentation increasing the number of shards will increase the level of parallelism while the number of reads/sec can be max 5/sec per shard. My question is whether (and how?) would adding more shards help the situation where all my consumers are up-to-date and attempt to read new incoming data from the same shard? It seems to be that this reads per sec limitation automatically introduces a limitation on the number of consumers you can have (at least when they need to be updated at all times), or am I missing something?

回答1:

Yes you are right.

In the consumers, I assume you'll use Amazon Kinesis Client (or KCL: amazon-kinesis-client) as API helper; and please take a look that there is a parameter "idleTimeBetweenReadsInMillis" in the consumer logic. That defines how much your application will poll the stream (the lower this value, more frequent your apps will poll).

Whether your stream contains 1 shard or 100 shards, you cannot make more than 5 "GetRecords" requests per second for each shard. That is;

• If you have 1 application, you can make poll interval 200ms at best (theoretically).
• If you have 2 applications, it can be 400 ms minimum.
• If you have 3 applications, it can be 600 ms minimum.
• Or with your 3 applications, two of them can poll at 1000 ms and final one can poll at 333ms rate.

You can also create a Kafka cluster for yourself and benchmark their performance. Kafka may give higher throughput.

See this answer for sample comparison between Kafka and Kinesis concepts: Kafka like offset on Kinesis Stream?

回答2:

Another alternative architecture is for you to have one kinesis consumer application which pushes the messages from the kinesis stream to an SNS topic. Of course, if your consumers need to 'look back' over past messages for processing, this may not work, but just wanted to throw that out as an option.

来源：https://stackoverflow.com/questions/44136758/kinesis-max-shard-reads-sec-and-multiple-consumers