Sunday, January 29, 2017

Handling data with changing schema on Hadoop

User data often is unpredictable, even when we can predict a change coming our way we need to prepare for that. Make changes in our environment to accept the incoming change, accommodate and absorb. 

With that being a fact of life, our design should allow for those changes to happen.  With the structured data warehouses gone (well, mostly), Hadoop based data stores are the norm these days, and would be default in future not so distant.  By definition, Hadoop allows all kinds of data to be stored on the cluster and happily provides various tools to process such data.

However, with different types of data out there on cluster, any given tool would need to "know" what data that is? How to read it, interpret it and process it. That would have been relatively straightforward with the data schema being static, as in, a file comes in from source A, with a certain format, and that’s about it. You can define programs/structures to work with it. 

What if the source system decides to alter the data set schema, add a few columns, and change some column ordering, perhaps data type too, what then? Would our program reading that dataset on Hadoop cluster still work as it? No.. It won’t .. Because it would also require change to reflect the same changes sent from source.

Assume that you did those changes too. So, now our program is version 2, able to handle the dataset sent by source (v2 as well). What about the older data lying around on Hadoop cluster (v1, with old schema)? How is that dataset going to used/usable?  What if that data was used by hive tables, would you have two (or more) set of tables? 

What if that data was exposed to some analytical programs, through R or similar? I am sure we won’t want to rewrite/test the whole solutions just because source system decided to change something?

And this scenario is not purely hypothetical, it happens around us, more often than not, just that we deal with it in our own patchwork styles, or solutions varying on a case to case basis.

Here I propose a somewhat generic solution, one that might work for all, well, nearly all use cases.

Avro is a data serialization system, wherein the schema is built into the file itself. As a result, programs reading the data-file doesn’t have to bother with interpreting the data file through external schema.

Avro was designed by Doug Cutting for similar reasons, as a tool that is language and schema independent and therefore unrelated tools/technologies can deal with data files.

That being said, is not a silver bullet in itself, but it handles the schema change fairly nicely.

Basically, you store the data on hdfs in Avro formatted files. For that, Avro schema needs to be created, which is fairly straightforward json structured file, indicating the record structure. Take a look at an example here - 

{"namespace": "myExample.avro",  "type": "record",  "name": "User",  "fields": [      {"name": "name", "type": "string"},      {"name": "age",  "type": ["int", "null"]},      {"name": "address", "type": ["string", "null"]}  ] }
create external  table ex_users (name     String ,  age     int , address     String) stored as AVRO location '/apps/data/hive/ex_users.tbl';

When files are saved as Avro formatted files, information from the schema is built into them, in plain text, and is parse-able for programs like hive and others to interpret the rest of the file.

For this to work, the tool has to rely on Avro. Look at this hive table definition - 

With the clause "Stored as Avro" we are basically telling hive to match the columns defined in hive with columns in Avro fi
les in that hdfs location and use whatever columns map to hive definition.
Avro is compressible, split-able and stores data in binary format, providing some additional features.

When this Avro formatted file is exposed to programs like hive, they don’t have to worry about changes in schema, since they rely on the schema information in the file header. If the hive / R schema asks for 5 columns, but the file only has 3 of them (with names as matching criteria and not order/position of columns), those matching columns are project to the consuming tool, and the rest is ignored. 

Similarly, if the tool asks for data that is not present in a certain data file (remember our example v2 file with additional column as against v1), the column that is not present is reported as null.

Now, look at the following data file (new avro schema - after source system decided to change)

{"namespace": "myExample.avro", "type": "record", "name": "User", "fields": [     {"name": "name", "type": "string"},     {"name": "age",  "type": ["int", "null"]},       {"name": "salary",  "type": ["float", "null"]},     {"name": "address", "type": ["string", "null"]} ]}

And the corresponding hive table definition - 

create external  table ex_users (name     String ,  age     int , address     String, salary float)stored as AVROlocation '/apps/data/hive/ex_users.tbl';

For this table definition, its not a problem to query and report a data file in old schema (v1) wherein the column salary is not present. it would simply know from the avro schema definition in that header that the file doenst have that column, and it would gladly report that as null.

On the other hand, in the new structure has a different order of columns, which is also fine for hive/avro combination since the schema definition helps match up the columns and report the right content in right columns. 

Similar analogies can be derived for other data consumers who need to interpret the schemas for data files on hdfs.

In comparison to some other storage formats, we do lose some points on the performance scale, but for the use case, where schema is changing and we want to be able to handle the changes with little or no effort, this is a better fit than many others.

Tuesday, December 20, 2016

Hadoop - Small Files vs Big Files


One of the frequently overlooked yet essential best practices for Hadoop is to prefer fewer, bigger files over more, smaller files. How small is too small and how many is too many? How do you stitch together all those small Internet of Things files into files "big enough" for Hadoop to process efficiently?
The Problem
One performance best practice for Hadoop is to have fewer large files as opposed to large numbers of small files. A related best practice is to not partition “too much”. Part of the reason for not over-partitioning is that it generally leads to larger numbers of smaller files.
Too small is smaller than HDFS block size (chunk size), or realistically small is something less than several times larger than chunk size. A very, very rough rule of thumb is files should be at least 1GB each and no more than maybe around 10,000-ish files per table. These numbers, especially the maximum total number of files per table, vary depending on many factors. However, it gives you a reference point. The 1GB is based on multiples of the chunk size while the 2nd is honestly a bit of a guess based on a typical small cluster.
Why Is It Important?
One reason for this recommendation is that Hadoop’s name node service keep track of all the files and where the internal chunks of the individual files are. The more files it has to track the more memory it needs on the head node and the longer it takes to build a job execution plan. The number and size of files also affects how memory is used on each node.
smallpiebigpieLet’s say your chunk size is 256MB. That’s the maximum size of each piece of the file that Hadoop will store per node. So if you have 10 nodes and a single 1GB file it would be split into 4 chunks of 256MB each and stored on 4 of those nodes (I’m ignoring the replication factor for this discussion). If you have 1000 files that are 1MB each (still a total data size of ~1GB) then every one of those files is a separate chunk and 1000 chunks are spread across those 10 nodes. NOTE: In Azure and WASB this happens somewhat differently behind the scenes – the data isn’t physically chunked up when initially stored but rather chunked up at the time a job runs.
With the single 1GB file the name node has 5 things to keep track of – the logical file plus the 4 physical chunks and their associated physical locations. With 1000 smaller files the name node has to track the logical file plus 1000 physical chunks and their physical locations. That uses more memory and results in more work when the head node service uses the file location information to build out the plan for how it will split out any Hadoop job into tasks across the many nodes. When we’re talking about systems that often have TBs or PBs of data the difference between small and large files can add up quickly.
The other problem comes at the time that the data is read by a Hadoop job. When the job runs on each node it loads the files the task tracker identified for it to work with into memory on that local node (in WASB the chunking is done at this point). When there are more files to be read for the same amount of data it results in more work and slower execution time for each task within each job. Sometimes you will see hard errors when operating system limits are hit related to the number of open files. There is also more internal work involved in reading the larger number of files and combining the data.
There are several options for stitching files together.
  • Combine the files as they land using the code that moves the files. This is the most performant and efficient method in most cases.
  • INSERT into new Hive tables (directories) which creates larger files under the covers. The output file size can be controlled with settings like hive.merge.smallfiles.avgsize and hive.merge.size.per.task.
  • Use a combiner in Pig to load the many small files into bigger splits.
  • Use the HDFS FileSystem Concat API
  • Write custom stitching code and make it a JAR.
  • Enable the Hadoop Archive (HAR). This is not very efficient for this scenario but I am including it for completeness.
There are several writeups out there that address the details of each of these methods so I won’t repeat them.
The key here is to work with fewer, larger files as much as possible in Hadoop. The exact steps to get there will vary depending on your specific scenario.

Wednesday, November 16, 2016

Eclipse - installing Scala plugin manually?

I have been playing around with Scala for some time, and was always using the Scala IDE ( which is based on a relatively older version of Eclipse (Luna).

I recently discovered this, wherein you could install the scala plug-in on a regular Eclipse installation.

Just add the following url as a new update site in your local eclipse installation and you'd be able to install the scala plugin just like that -

Sunday, July 24, 2016

Links to free big-data-sets

Many people who are starting their journey with big data and analytics find it hard to get their hands on the right kind of data to play or experiment with.

Most of the time, people have enthusiasm, they are learning the skill too, but they just don't have the right kind of dataset to apply their newly acquired skills.

Democratising data has been at the forefront of discussions for many data pioneers. Through their efforts and with some re-alignment of technology priorities, some government bodies have opened up their datasets to the public.

As a result, here is a set of links (reproduced) to some of the free sources.
  1. The US Government pledged last year to make all government data available freely online. This site is the first stage and acts as a portal to all sorts of amazing information on everything from climate to crime. 
  2. US Census Bureau A wealth of information on the lives of US citizens covering population data, geographic data and education. 
  3. Socrata is another interesting place to explore government-related data, with some visualisation tools built-in. 
  4. European Union Open Data Portal As the above, but based on data from European Union institutions. 
  5. Data from the UK Government, including the British National Bibliography – metadata on all UK books and publications since 1950. 
  6. Canada Open Data is a pilot project with many government and geospatial datasets. 
  7. offers open government data from US, EU, Canada, CKAN, and more. 
  8. The CIA World Factbook on history, population, economy, government, infrastructure and military of 267 countries. 
  9. 125 years of US healthcare data including claim-level Medicare data, epidemiology and population statistics. 
  10. NHS Health and Social Care Information Centre Health data sets from the UK National Health Service. 
  11. UNICEF offers statistics on the situation of women and children worldwide. 
  12. World Health Organization offers world hunger, health, and disease statistics. 
  13. Amazon Web Services public datasets Huge resource of public data, including the 1000 Genome Project, an attempt to build the most comprehensive database of human genetic information and NASA ’s database of satellite imagery of Earth. 
  14. Facebook FB +0.32% Graph Although much of the information on users’ Facebook profile is private, a lot isn’t – Facebook provide the Graph API as a way of querying the huge amount of information that its users are happy to share with the world (or can’t hide because they haven’t worked out how the privacy settings work). 
  15. A fascinating tool for facial recognition data. 
  16. UCLA makes some of the data from its courses public. 
  17. Data Market is a place to check out data related to economics, healthcare, food and agriculture, and the automotive industry. 
  18. Google Public data explorer includes data from world development indicators, OECD, and human development indicators, mostly related to economics data and the world. 
  19. Junar is a data scraping service that also includes data feeds. 
  20. Buzzdata is a social data sharing service that allows you to upload your own data and connect with others who are uploading their data. 
  21. Gapminder Compilation of data from sources including the World Health Organization and World Bank covering economic, medical and social statistics from around the world. 
  22. Google GOOGL +0.66% Trends Statistics on search volume (as a proportion of total search) for any given term, since 2004. 
  23. Google Finance 40 years’ worth of stock market data, updated in real time. 
  24. Google Books Ngrams and analyze the full text of any of the millions of books digitised as part of the Google Books project. 
  25. National Climatic Data Center Huge collection of environmental, meteorological and climate data sets from the US National Climatic Data Center. The world’s largest archive of weather data. 
  26. DBPedia Wikipedia is comprised of millions of pieces of data, structured and unstructured on every subject under the sun. DBPedia is an ambitious project to catalogue and create a public, freely distributable database allowing anyone to analyze this data. 
  27. New York Times  Searchable, indexed archive of news articles going back to 1851. 
  28. Freebase A community-compiled database of structured data about people, places and things, with over 45 million entries. 
  29. Million Song Data Set Metadata on over a million songs and pieces of music. Part of Amazon Web Services. 
  30. UCI Machine Learning Repository is a dataset specifically pre-processed for machine learning. 
  31. Financial Data Finder at OSU offers a large catalog of financial data sets. 
  32. Pew Research Center offers its raw data from its fascinating research into American life. 
  33. The BROAD Institute offers a number of cancer-related datasets. 

Credit to Forbes article at

Friday, June 19, 2015

Teradata Data type abbreviation - described

Teradata data types (as reported in DBC.Columns.ColumnType can be cryptic and not always easy to remember.  Here's a ready reckoner - 

Equivalent English :)
UDT Type 

Thursday, May 7, 2015

Hadoop Meetup on the sidelines of Strata Hadoop Conference - Part 2

Read part 1 of this here

Day 2 of the meetup was equally exciting, if not better.  Lined up were talks from Qubit and Google, William Hill (a surprise for me - more later on that) and then PostCodeAnywhere, all very exciting from the synopsis.

Google & Qubit showcased basically a stream processing engine, with pluggable components, many of them can be written in different technologies and programming languages.

Of course Google Cloud Data flow is much more than just a stream processing engine, however, for real time data ingestion perspective, that feature is pretty significant.  

A completely managed system, it woks on the publish-subscribe (pub-sub) model.  As Reza put it, “pub-sub is not just data delivery mechanism, its used as a glue to hold the complete system together”.  Pluggable components is another differentiator for Google’s offering, in today’s demo they showcased bigtable as one of the consumers at the end.

From my own knowledge of stream processing, which is not significant in anyway, i could relate to many similarities with IBM’s info sphere streams and some with apache kafka.  However, a question around comparisons with these sites remained unanswered from Google (though in very good spirit, in a chat with the speaker Reza later on, it came out as more of a philosophical question avoidance than anything else).

The william hill talk (by Peter Morgan, their head of engineering), was a genuine surprise, at least for me.  Perhaps due to my ignorance, due to which i didn't realize, their systems are far more sophisticated and load bearing than I would have imagined.  As an example, they process 160TB of data through their systems on a daily basis.

Including many complexities managed through their system are their main components, the betting engine, the settlement engine among others. 

William Hill supports an open API as well, enabling app developers to pick up data elements and innovate. However, for obvious reasons, very limited data is thrown open in the public domain.  Would that be a deterrent for app developers ? not having enough data ?   For example, if i would want to report in an app, who’s betting on a  certain game, cross referenced with geo location data .. I cant do that, since William hill doesn't publish demographic data.  I personally feel alright with it, there are possibilities that many of those data elements can be used in ways to influence the betting system itself, becoming counter-productive.

I would imagine their IT systems to be one of the top notch systems around the place, to be able to manage such data volumes, with such speeds and accuracy. Commendable job.  I would probably write exclusively on their architecture once i get my hands on the presentation slides (couple of days may be).

The talk from PostCodeAnywhere was more educative to me, personally.  Got to understand a bit about Markov Models, chains and how they can be used for machine Learning.  Very interesting stuff there too.

Apache Spark is being seen more and more as the tool to be perform analytics on the fly, specially on large volumes of data.  It would be very interesting to see how R and python analytical capabilities compare with what spark offers.

Speaking to another attendee today, it came out the people prefer to use R more and more for massaging and cleansing purposes, however, its not seen as fit for heavy lifting required for performing real analytic and/or predictive pieces. For these areas, people still prefer to use Python.

IBM’s bigR is a possible contender for the job, where they talk about having optimised R for a hadoop cluster and have enabled it to work on top of hdfs.  However, bigR is not open source and that could be its biggest challenge in adoption.