In my previous post about Hadoop and Impala I benchmarked performance of analytical queries in Impala.

This time I’ve tried InfiniDB for Hadoop (open-source version) on the modern hardware with an 8-node Hadoop cluster. One of the main advantages (at least for me) of InifiniDB for Hadoop is that it stores the data inside the Hadoop cluster but uses the MySQL server to execute queries. This allows for an easy “migration” of existing analytical tools. The results are quite interesting and promising.

Quick How-To

The InfiniDB documentation is not very clear on step-by-step instructions so I’ve created this quick guide:

  1. Install Hadoop cluster (minimum install will work). I’ve used Cloudera Manager (CDH5) to compare the speed of InfiniDB to Cloudera Impala. Install the tools in the “Pre-requirements” sections of InfiniDB for Hadoop Manual
  2. Install the InfiniDB for Hadoop binaries on 1 Hadoop node (you can choose any node).  This will install InfiniDB and its version of MySQL (based on MySQL 5.1).
  3. After installation it will tell you the variables to set and run the postConfigure script. Example:
  4. The postConfigure script will ask the questions. Couple imfortant notes:
  • Make sure to use HDFS as a “type of Data Storage”.
  • The performance module 1 (pm1) should point to the host (hostname and IP) you are running the postConfigure script on. Other pm(s) should point to other Hadoop nodes

When installation is finished you will be able to login into MySQL server, it uses script called ibdmysql which will call mysql cli with the correct socket and port. Check that the infiniDB is enabled by running “show engines”, InfiniDB should be in the list.

The next step will be importing data.

Data import

First we will need to create a MySQL table with “engine=InfiniDB”:

Second,  I’ve used the cpimport to load the data. It turned out it is much more efficient and easier to load 1 big file rather than 20×12 smaller files (original “ontime” data is 1 file per month), so I’ve exported the “Ontime” data from MySQL table and created 1 big file “ontime.psv”.

I used the following command to export data into InfiniDB:

The data is stored in Hadoop:

The total size of the data is 8×1.4G = 11.2G (compressed). To compare the size of the same dataset in Impala Parquet format is 3.6G. Original size was ~60G.

Now we can run the 2 queries I’ve tested before:

1. Simple group-by

2. The complex query from my original post:

The same query in impala (on the same hardware) runs for 7.18 seconds:

Conclusion and charts

To summaries I’ve created the following charts:

Simple query:

As we can see InfiniDB looks pretty good here. It also uses MySQL protocol, so existing application which uses MySQL will be able to work here without any additional “connectors”.

One note regarding my query example: the “complex” query is designed in a way that will make it hard to use any particular set of index; this query will have to scan the >70% of the table to generate the resultset. That is why it is so slow in MySQL compared to columnar databases. Another “issue” is that the table is very wide and most of the columns are declared as varchar (table is not normalized), which makes it large in MySQL. All this will make it ideal for columnar storage and compression. Other cases may not show that huge of a difference.

So far I was testing with small data (60G), I will plan to run big data benchmark next.

4 Comments
Oldest
Newest Most Voted
Inline Feedbacks
View all comments
Dipti Joshi

Thank you Alex for sharing your experience with InfiniDB! Let us know as you test on larger data set.

0
Ebot Tabi

Hey nice piece, thanks for sharing your experience with Infinidb.

0
Vadim Tkachenko

Alexander,

Now it would be interesting to compare with Spark http://spark.apache.org/

0
Krishan

Dear Alex, Do you still consider InfiniDB-Hadoop to be a superior option for SQL-on-Hadoop for BI analytics?

0