Amazon Kinesis is 20x Cheaper With the Kinesis Producer Library

DZone 's Guide to

Amazon Kinesis is 20x Cheaper With the Kinesis Producer Library

With the recent release of the Kinesis Producer Library (KPL), Amazon exposed a native capability to aggregate multiple messages/events into a single PUT unit.

· Big Data Zone ·
Free Resource

We've been looking at the tradeoffs between Kafka and Kinesis, and just recently.... Kinesis won!

Kinesis pricing is based on two dimensions: hourly price for each shard (which dictates overall concurrency/throughput), and a unit price per 1M messages.  As I alluded to in my previous post, we process a *ton* of events, and if we simply pushed those events onto Kinesis streams 1:1, we would quickly hand over all of our money to Amazon and go out of business.

However, with the recent release of the Kinesis Producer Library (KPL), Amazon exposed a native capability to aggregate multiple messages/events into a single PUT unit.  The maximum size of a PUT unit is 25Kb.  If you have a message size of 150 bytes, you can cram about 150 messages  into a single PUT unit! (and save some serious dough!)

Here is the math, straight from the Kinesis Pricing Calculator:

Without the KPL,

100K m / s * (150 bytes / m) = 100 shards and 263,520M PUT units = $4,787.28 / month

* Note: That each shard can only process 1K/s, which is why we end up with 100 shards.

Now with KPL, let's assume we can only fit 100 messages in each PUT unit.

We would reduce our required throughput (messages / second) down to 1K / s.

With 100 messages in each PUT unit,  each unit would be 15Kb in size. (100 * 150 bytes)

This gives us:

1K m / s * (15 Kb / m) = 15 shards and 2,635.2M PUT units = $201.60 / month

That is s savings of: ~20x!!

So, let's look at what that means architecturally...

First, you are going to want to look at the KPL recommended usage matrix:


My naive interpretation of that chart is: Use Java!

Ironically, the producer is actually a native built micro-service/binary.  And the KPL is a java wrapper around that native binary that uses interprocess communication to delegate work to the micro-service.

In Java, it is straightforward to integrate the KPL.  Here are the key components.

First, Configure the Producer:

KinesisProducerConfiguration config = new KinesisProducerConfiguration();
config.setCredentialsProvider(new ProfileCredentialsProvider());
producer = new KinesisProducer(config);

Notice, that the KPL actually sends the messages asynchronously, buffering/aggregating those messages internally (which makes for challenging guarantees around message delivery -- but we'll defer that problem for another day).

Also notice, that I built my own native binary, and specified its location.  You don't necessarily need to do that, the binary is actually bundled into the JAR file and extracted into a temporary location at runtime.  (Yeah, crazy eh?)

Next, You Need to Send Some Data:

This is done simply, one line of code:

ListenableFuture<UserRecordResult> f = producer.addUserRecord(streamName,
  Long.toString(System.currentTimeMillis()), data);

Finally, You Need to Find Out What Happened to Your Message:

For this one, we'll add a callback to the future:

Futures.addCallback(f, callback);

Where, callback is a method that will be called when the record is processed.  Here is an example:

final FutureCallback<UserRecordResult> callback = new FutureCallback<UserRecordResult>() {
 public void onFailure(Throwable t) {
  if (t instanceof UserRecordFailedException) {
   Attempt last = Iterables.getLast(((UserRecordFailedException) t).getResult().getAttempts());
   LOGGER.error(String.format("Record failed to put - %s : %s", last.getErrorCode(),
  LOGGER.error("Exception during put", t);

 public void onSuccess(UserRecordResult result) {
  // LOGGER.info("Successfully wrote [" + result.toString() + "]");

And there you have it, KPL in a nutshell.

For us, it was a huge win, and actually made Kinesis a viable alternative.
Now, we just need to figure out how to plumb it into everything else. =)

Kinesis Firehose to the rescue!  (stay tuned)

big data, big data and analytics solutions, big data-as-a-service bdaas, data as a service

Published at DZone with permission of Brian O' Neill , DZone MVB. See the original article here.

Opinions expressed by DZone contributors are their own.

{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}