Parquet File format - Storage details

By



Columnar storage has revolutionized Big data processing, since its inception. Its power can be realized from the fact that Google Big Query, Hbase, Amazon Redshift, Azure SQL Data Warehouse and many more, all utilize columnar storage.

As you are here today reading this article, it is obvious that you are curious to learn how Columnar storage internally works. Well, just stay tuned for next few minutes, as I'll explain all the details of the most popular columnar file format called Parquet. 

We'll learn the details by actually opening a parquet file and by going deep, at the level of how things work under the hood at disc level. We'll also learn why Parquet would save cost, when it is used as underlying file format with AWS services such as Athena.

When to use Parquet format - 

Just check this less than 30 seconds conversation, going on between a Data Engineer(male) and a BI Analyst(female).


When to use parquet - Whenever our dataset has large number of attributes(columns) and our processing on that data (lets say ETL code or BI reporting queries), is focused only on very small number of attributes and we want this data processing to execute quickly. 

For instance, say we have a table with 500 columns, but most of our SQL queries execute on just 10 out of those 500 columns. Here Parquet, would bring back results very swiftly, when compared to any row format storage(such as Text, AVRO etc). 

Well what is so different about Parquet which brings such efficiency?


What is inside a Parquet file

Like any file format, parquet file consists of data. But along with the data, it also stores metadata about how the data has been stored on disc. Metadata is written after the data is successfully written. 

The main units of Parquet file format which contribute to its efficiency are, Row groups, Column chunks and Page. You might be thinking "What the hell are these row groups, column chunks, blah blah?" Don't worry if these 3 terms does not make sense, as these would make complete sense, once you have read this article till the end. 

Without going into the documentation details of Parquet, which you can anytime find here, lets pick a business requirement and open a parquet file and practically understand the reason for efficiency.

Our Parquet Data

We'll use the data of a hypothetical vehicle valuation company, such as Kelly Blue Book(aka KBB). Basically kbb.com provides a platform to know the value of any used as well as new cars. To earn money, KBB also let different car companies such as Toyota, Honda etc, post the ads of their cars on kbb.com. And to post ads, these car companies pay money to KBB. 

So our parquet data is the data of advertisements of cars, as posted by different car companies(such as Honda, Toyota etc) on kbb.com. 

Our data is composed of one single parquet file, with approximately 40 million rows. Our parquet file consists of the following 6 columns:- 

1. make  - Make of the car for which ad is posted(eg Honda).
2. model  - Model of the car for which ad is posted(eg Accord).
3. advertiser_nbr  - a unique id of the car company which posted the ad. 
4. order_nbr - order number of ad(as every ad posted by different car companies is an order for kbb).
5. creative_nbr  - Unique number of ad.
6. creative_sz  - Size of the ad slot on kbb.com.

The business is just interested in the details of the advertising companies, as they are the customers who generate the revenue. And hence we need to just fetch one column advertiser_nbr out of the 6 columns. Using this data we will learn, when we want to fetch just one column advertiser_nbr out of the 6 columns, how parquet format would return results quickly

Important Note - Parquet is a binary format and a parquet file cannot be just opened and read like a text file. Hence in the next section, I will explain how we can open and read a parquet file. If you are more interested in learning just about the parquet storage details, please feel free to skip the next section and directly jump to the Metadata of our parquet file section, where I have opened and explained the storage details.

Parquet data table 

I am using Hive Datawarehouse for storing this kbb's parquet data file. I have created a Hive table add_details and stored it as Parquet, choosing my home directory /user/akshay/add_details/ as data file location on HDFS.


DDL





The data of this table is in one single parquet file 000000_0 in HDFS. To read the file, we will pull the data file 000000_0 to local, using traditional hadoop get command as follows:




Count of rows in this table - Our parquet dataset has close to 39.5 million records as shown below:
















Lets see what is inside the Parquet file:

Parquet format is a binary format and a parquet file cannot be just opened and read like a text file. Hence we'll use parquet-tools(a utility built to read parquet files) and read the metadata of our file 000000_0, to see how parquet stores the columns of these 39.5 million records.

Steps to read parquet file:

parquet-tools project was specifically created, to support easy readability of parquet files. I have the parquet-tools utility script at /opt/cloudera/parcels/CDH-5.4.4-1.cdh5.4.4.p1406.1253/bin/ in my Cloudera Hadoop cluster. 

Note - I am keeping the details of parquet-tools out of the scope of this article and will be publishing a separate article on same.

To read parquet file's metadata on a Cloudera Hadoop cluster, you can use the following command:




To view the parquet file present in local directory(not in HDFS), follow the following steps:




As we want to view just the metadata of parquet file and not the actual data. Hence just execute the parquet-tools script with meta argument as follows:



Metadata of our parquet file:


On opening the metadata file, this is what we see


 


Click on image to enlarge














Our Parquet file stores the 39.5 million rows, by dividing them into 3 row groups(highlighted in picture in blue box). 
A row group is a logical partitioning of the data into group of rows.  

The 3 row groups have the following row count (highlighted in picture in green box): 
row group 1 - 18.6 million rows
row group 2 - 18.8 million rows
row group 3 - 2 million rows

Each row group consists of column chunks. A column chunk is a chunk of data values for any specific column. 
Each row group contains exactly one column chunk per column of the data schema. Since we have 6 columns in our data, hence each row group contains 6 column chunks. What is important to remember, is that a column chunk is guaranteed to be stored contiguously in the physical parquet file. 

A column chunk is further subdivided into pages. And a page is an indivisible unit. 

Now the business is just interested in getting the advertising company details and hence we need to just fetch one column advertiser_nbr.  So it is time to understand why fetching just this column out of the 6 columns, would fetch us results quickly.


Row group 1: 

On focussing on advertiser_nbr of row group 1, we see the following storage details(highlighted in picture in orange boxes):

1. BINARY - simply means it is binary data and is stored as byte array.

2. SNAPPY(in orange) - It is the compression algorithm. Data is compressed using SNAPPY algorithm.

3. FPO(in orange) - First Page offset. This is the exact first starting storage point(offset) on disc, from where the values of column advertiser_nbr in row group 1 are stored. 
FPO:4 means that the values for advertiser_nbr are stored starting from offset 4 on disc. 
 
4. SZ(in orange) - SZ:19022560. This is the size in bytes, of 18.6 million values of the  column advertiser_nbr. This means that 19022560 bytes of data for advertiser_nbr column's 18.6 million values, are stored on disc starting from offset 4. This is exactly where the efficiency comes from. 18.6 million values of one column advertiser_nbr are stored colocated on disc. Yes all stored contiguously at one place.  Hence when we run any SQL query or transformation code on just column advertiser_nbr, then the disc head seeking this column, gets all the 18.6 million values in single seek from one region on disc. Thus no need of several disc seeks. 

In SZ:19022560/24391353/1.28

19022560 is compressed size
24391353 is uncompressed size
1.28 is compression ratio

5. VC(in orange) - Value count, which is equal to the number of values of a column in the row group. It is 18.6 million for column advertiser_nbr in row group 1.

6. ENC(in orange) - Encoding used to store the values of this column in parquet file.


Time to answer key questions we asked in the beginning:

                          

There are 6 columns in our table. But if we write a select query fetching only one column advertiser_nbr out of the 6 columns, then:

Question 1 - Why parquet format would return result faster, compared to a row format(such as text)?

Answer - To read just the column advertiser_nbr from this parquet file, the disc head seeking this column, needs to only seek to page offset 4 and traverse till offset 19022563 (similarly seek other advertiser_nbr column chunks in Row group 2 and 3). 
And guess what - 18.6 million values (row group 1) for this column are stored colocated on disk at one place and also stored in encoded format. 
This reduces disk seeks, as the disc head(image below) seeking this one column gets all the 18.6 million values from one region on disc. Thus no need of several disc seeks. Lesser seeks means lesser processing time.
Where as in a row format, these 18 million values would be stored in row fashion, in separate rows at several distributed places on disc(in the worst case 18 million places). And to fetch these 18 million values from a row format file, the disc head would have to scan all the rows, which would require many more disk seeks and hence more processing time. 

What a disc seek looks like
What a disc seek looks like


Question 2 - How can we save cost by using Parquet with AWS Athena?

Answer - As millions of values of individual columns are stored colocated on disc, when data for specific column is fetched, we do not need to scan data for other columns unnecessarily. Thus data scan gets reduced. Whereas in a row format we would have scanned all the rows to fetch just one column, leading to more data scan. And since the pricing of Athena works on $5/TB scan(as of the date of publishing this article), Parquet by reducing scan, reduces the cost. Thus it proves to be really cost efficient.

Thank you for taking the time to read this article. I hope through this article, you have gained deeper understanding of the storage details of Parquet format and columnar storage in entirety. 
Location: California, USA
Big Data, Cloud(AWS), Graph and Machine Learning professional.

Comments

New comments are not allowed.

Popular posts from this blog

UPSERT in Hive(3 Step Process)

By
Image
In this post we'll learn an efficient 3 step process, for performing UPSERT in hive on a large size table containing entire history. Just for the audience not aware of UPSERT - It is a combination of UPDATE and INSERT. If on a table containing history data, we receive new data which needs to be inserted as well as some data which is an UPDATE to the existing data, then we have to perform an UPSERT operation to achieve this. Prerequisite  – The table containing history being very large in size should be partitioned, which is also a best practice for efficient data storage, when storing large volume of data in any Big Data warehouse. Business scenario  – Lets take an example of click stream data of a website, as gathered from different browsers of visitors who visited the website. The site_view_hist table contains the clicks and page impressions counts from different browsers and the table is partitioned on hit_date(the date on which the visitor hitted or visited the website).
Read more

Parquet tools utility

By
Image
It is a well known fact, that parquet format is a binary format. And for humans, reading binary data is not as easy as opening a text file and understanding its contents. But no worries, today we'll learn how to read a parquet file . If you would like to understand all the internal details of a parquet file , you can please refer my previous article here . In this article I have tried to explain all the inner storage details of Parquet and answered most of the questions around its usage, efficiency and performance.                                                                                Parquet tools - As we know parquet is a binary format, we need an interpreter to translate its contents so that we can understand them. And this is what parquet-tools  utility project does for us. It lets us read parquet files. Note - At the time of writing this blog, I used parquet-tools utility already installed on my Hadoop cluster. But you can always install parquet-tools on you
Read more

Hive - Merging small files into bigger files.

By
Image
Since Hadoop's inception, we have learnt that it is good for processing small number of large files, instead of large number of small files. Part of the reason pertaining to this behavior is, its MapReduce Paradigm which works on splits of input data.  In our day to day data processing, there can be various situations where our processes can produce large number of small files. Assume a process generated output data of 300 MB, but this data is distributed in approximately 3000 files of 99 KB each. Today we'll learn how to merge such large number of small files into bigger files. Before that let us quickly understand the problems, such large number of small files can cause. Problems posed by large number of small files -   1.   I f we store the data of this table in GZIP format(3000 files each of 99KB), then a downstream process running a hive query(MapReduce job) on this data, would spin up 3000 mappers to process individual splits for each file, as the GZIP format is non
Read more

Skew Join Optimization in Hive

By
Image
Skew is a very common issue which most of the data engineers come across.  What is Skew - When in our data we have very large number of records associated with one(or more) particular key, then this data is said to be skewed on that key.  We'll understand this with a very simple example. Think about world's population data table, where each record is of one person and the number of records equals to total world population. As we all know China and India, each accounts for around 17-18% of world's population, hence such population table is referred to as skewed on keys country=China and country=India. If you focus on the following pie chart of world's population distribution by country, then this would make exact sense. Problem statement  - Assume we have a table which tracks the online visits to different online websites. In this table amazon.com  being very popular,  has 1 Billion rows whereas other not very popular sites ( www.abc.com , www.xyz.com..etc) i
Read more

Apache Spark - Sort Merge Join

By
Image
Apache Spark is a well known open source cluster computation system, for large scale data processing. If you are into Big Data and Machine Learning, Apache Spark may not need any introduction. Its capabilities include In-memory processing, Distributed SQL Analytics, Real Time stream processing, Graph processing, Machine Learning,  support for several programming languages  and many more.  " What happens under the hood, when we execute a Spark application? ".    thinking????? This is a very common question, which crops up in every Apache Spark user's mind. If you have the same question, I 'll answer this using Spark SQL by executing one of the most common data processing operation i.e. "Join". Spark being a vast technology, it is nearly impossible to cover all of Spark's details in a single article. And we all learn something new, every time we use a technology to solve a new challenge.  However, the details I'll cover here, should be enough to give
Read more

Apache Spark - Explicit Broadcasting Optimization

By
Image
Join operation is an indispensable part of majority of Big Data Projects. If you are using Apache Spark for your Big Data processing and using join operation, in this article I'll explain a strategy to optimize joins. The technique I'll explain is how to explicitly broadcast a table, in any Spark join operation and convert the join operation to Broadcast Hash Join. Why Broadcasting - W hen we perform a join operation in Spark SQL, b roadcasting proves very efficient in reducing data shuffling and hence serialization and deserialization of data over the network, which happens during a Sort Merge Join operation. If you haven't read my previous article on Spark's Broadcast Hash join, I highly recommend you to first go through that. This would provide you most of the deeper details of Spark's Broadcast hash Join. You can please find it here :- Broadcast Hash Join - By Akshay Agarwal Business use case  -  To explain this optimization, let's take the business
Read more

Spark Optimization - Bucketing

By
Image
Complex data computation is a part of every Big Data Project.  Do you have a business scenario requiring complex join on data having billions or trillions of values?  If you are using Apache Spark, stay tuned for next few minutes, as this article will address this expensive join problem, using Spark's optimization technique called Bucketing. Before reading this article, I recommend you to once go through my previous articles on Spark's Sort Merge Join and Broadcast Hash Join . They will set up your foundation for understanding bucketing technique. It is recommended but is not mandatory, as you would still be able to grasp the core concept. When to use Bucketing ?    For any business problem, where we need to perform a join on tables with very high cardinality on join column( I repeat very high ) in say millions, billions or even trillions and also this join needs to happen multiple times. Lets understand bucketing using the data of a hypothetical service provider company
Read more

Apache Spark - Broadcast Hash Join

By
Image
In this technical article, we'll discuss about one of Apache Spark's efficient join technique called Broadcast Hash Join. I will explain in detail how this operation is implemented internally in Spark at DAG(Directed Acyclic Graph) level and also different phases of DAG. We'll also uncover the details of division of spark application into jobs and stages.  Data  -  To explain this join concept, I'll take customer data of a hypothetical online service provider organization with international presence. Different customers across the world using its services holds accounts with this company. Moreover the customers uses different devices to access this company's website. Hence to fetch the details of different  customers(business_accounts) and the corresponding device used by them, we'll use two datasets: Business Accounts data   -  Data consisting of details of the customer's business account, created with this service providing organization.  The schema f
Read more

Graph Technology - In plain English

By
Image
Graph processi ng systems have become a prominent buzzword in computation world in recent years. Is this just a passing trend which is here today and would vanish tomorrow? There is no harm in being skeptical. But with knowledge graphs recognized as a top emerging technology trend by industry expert Gartner, it's evident that graph technology is here to stay.  So let's dive in and discover what makes graph technology special and why it's gaining momentum.   What industry experts are saying about Graph Technology: Knowledge graphs has been recognized by Gartner among the top 5 emerging technology trends, alongside IOT & Blockchain, that will blur the lines between humans and machines. I have included a snippet from Gartner's press release below. You can also find Gartner's official full press release   here  . Snippet from Gartner's Newsroom Press Release  An important note before we proceed further. If you do a simple google search about graphs, y
Read more