Skip to content

Latest commit

 

History

History
46 lines (27 loc) · 3.79 KB

README.markdown

File metadata and controls

46 lines (27 loc) · 3.79 KB

Extending the Yahoo Streaming Benchmarks

Background

This code is a fork of the original Yahoo! Streaming benchmark code.

We have added to this the following additional benchmarking programs. They are all variations of the original Yahoo! benchmarks.

For additional background please read "Extending the Yahoo! Streaming Benchmark" and "Benchmarking Streaming Computation Engines at Yahoo!"

flink.benchmark.AdvertisingTopologyFlinkWindows

This Flink program does the same operation in Flink as in the original benchmarks but uses Flink's native windowing and trigger support to accomplish the same things. It computes 10 seconds windows, emitting them to Redis when they're complete and also emits early updates every second.

flink.benchmark.AdvertisingTopologyRedisDirect

This Flink program builds 60 minute windows directly in Redis. It's purpose is to illustrate the key-value store as the bottleneck when you don't have fault-tolerant state in the streaming system. Flink provides alternatives to this approach. The window time is configurable.

flink.benchmark.state.AdvertisingTopologyFlinkState

This Flink program is an example of using the Flink nodes themselves as the key-value store to eliminate a remote store as the limit to throughput. By default it computes 60 minute windows but this is configurable.

flink.benchmark.state.AdvertisingTopologyFlinkStateHighKeyCard

This is the same as above but in addition it is designed for a very high key cardinality.

flink.benchmark.state.AkkaStateQuery

This program is used to query the key-value state directly in the Flink nodes. This in combination with either of the two -FlinkState programs above is what allows you to eliminate the key-value store altogether.

storm.benchmark.AdvertisingTopologyHighKeyCard

This Storm program is derivative of Yahoo's original AdvertisingTopology code. The difference is that it is designed for very high #'s of campaigns and builds the output windows by directly updating them in Redis. This benchmark represents any computation where it's neccessary to build state outside of the streaming system such that it's fault tolerant. Many applications in production work exactly this way and typically their throughput is limited by the remote key-value store. The windows in this case are set to 60 minutes by default but this is configurable.

flink.benchmark.genertor.AdImpressionGenerator

A data generator for Flink compatible with the original Yahoo data generator. This generator runs via Flink and writes data to Kafka. It can also be embedded directly in your Flink program as a source if Kafka becomes a bottleneck. This is the generator we used in the benchmarks as it was difficult to get the Clojure based generator to run at a high enough throughput for our experiments.

flink.benchmark.generator.AdImpressionsGeneratorHighKeyCardinality

Same as the generator described above but updated to generate a much higher number of ad campaigns. The number of ad campaigns is configurable.

Configuration

We've continued to use the benchmarkConf.yaml for all configuration. You'll need to edit this file for your environment and also to control various parameters both for the load generators and the benchmarking programs.

Results

For most of these benchmarks the results are collected using Yahoo's original scripts. See that benchmark for more details. In the case of the -FlinkState programs use the AkkaStateQuery program to retrieve the results.

Other notes

This code was built and tested against Flink 1.0-SNAPSHOT (master on Jan 27, 2016) and Storm 2.0.0-SNAPSHOT (SHA a8d253).