Steps to Modernize Your Data Ecosystem | Mindtree

Six Steps to Modernize your Data
Ecosystem to make Collaborative
Intelligence Possible
A Mindtree White paper | 2020
Manoj Karanth
Sowjanyakumar Kothapalli

Table of Contents
1. Separation of Compute and Storage 3
2. Assess Workload Type 4
3. Data Processing and Data Store optimization 6
4. Design Consumption Landscape 7
– SQL 7
– Machine Learning 8
5. Assess Security and Governance 8
6. Data Migration 9
7. What does Mindtree bring to the table? 10
8. References 11

A Mindtree White paper 3
Organizations across the world are striving towards
being more data-driven in their decision-making. The
right mix of human and machine intelligence is crucial
for organizations to succeed in this journey. Machine
intelligence needs to be supported with the right data
infrastructure, and organizations have invested in
setting up the same with the likes of data lakes, data
warehouses etc.
At the same time, these investments have not quite
provided the outcome that organizations had expected.
A common set of challenges that organizations have
faced are:
Business value of insights: Choosing the right use cases
and KPIs which could generate valuable insights for the
business has always been a challenge.
Time to insight: While Big Data improved the capability
to process data faster, organizations have proceeded to
crunch more data. However, the availability of the data
in time remains a key aim.
Cost per insight: Once the data is gathered in the
system, a big challenge in making it available for other
teams to use is the cost. Big data environments guzzle a
lot of computing power which increase the cost of the
environment specially in the cloud.
This has led organizations to take a re-look at their data
estates and look to address these challenges. Over the
years, technologies in the Big Data landscape have
continued to change with Spark emerging as the de-fac-
to processing mechanism for data needs. These technol-
ogies alleviate the limitations in first-generation big data
systems built with Apache Hadoop-based systems with
distributions like Cloudera, Hortonworks etc.
In this paper, we highlight how one can approach this
modernization path. We have identified Databricks on
AWS as the target environment. New generation data
platforms are unified i.e. we have the same stack for
batch, streaming, machine learning. We have chosen
Databricks since it is best performing Spark engine and
is the leading player in bringing unified platforms to life
The modernization approach is composed of the
following steps
Separation of Compute and Storage
One of the founding principles in Hadoop was that for data processing to be scaled horizontally, compute had to be
moved to where the storage resided. This would reduce the load on network I/O transfer and make the systems truly
distributed. To process the data efficiently, these machines or nodes would have high memory and CPU requirements.
Separation of
Compute and
Storage
Workload Type
and
Resource Usage
Data Processing
and Data Store
Optimization
Design
Consumption
Landscape
Assess Security
and Governance
Data Migration
and Movement
When we transported the same concept to cloud-based
environments, the cost of running and scaling started
becoming increasingly high. When one is running a data
lake, most of the time one needs the storage and not the
processing capacity. In Hadoop-based data environ-
ments, compute and storage are tied together with HDFS
as the file system. E.g. In AWS, d2 is the most cost-effi-
cient storage instance type
In the cloud, object storage is durable, reliable and
cheap, while the network capabilities continue to
increase. This led to a decoupling of compute and

A Mindtree White paper
4
storage, and most cloud-native data architectures are adopting this mode with object storage as the Data Lake.
Modern data platforms like Databricks provide the elastic capability required to utilize the power of separating
storage and compute.
This has a significant impact on the cost as outlined in the example below taking AWS as an example to store 1PB of
data using a 40-node cluster. The calculation was done using the logic that a M5a.8X large cluster will run for about
12 hours a day. This instance type cluster is slightly high, since in most practical cases, due to varied loads, much
smaller instance type clusters with fewer nodes can be configured.
Hadoop
100K USD
Databricks
48 TB per node * 40 d2.8x nodes = 5.52* 40 +
15*0.78 other
Storage: S3 cost = 25K
Processing: 40 node M5a.8x large
Hosting (40 * 0.867x360 hours) = 12.5K
DBU cost (40* 0.4*4.91*360) = 28.5K
66K USD
Illustrative Cost Comparison (50% utilization)
Consumption Reports, Models
Assess Workload Type
In addition to separation of storage and compute, data processing time and cost can be optimized through an
understanding of the workloads. This impacts both time to insight and cost per insight. In Hadoop-based environments,
multiple workloads run on the same cluster to optimize the spend. Hence, it is important to assess the different
workloads and their most efficient processing environments. Following is an example of different workloads
Adhoc Query
Report and KPI Processing
Clickstream Processing
Transaction Processing
Illustrative Workload Distribution on a 24-Hour Time Scale

 Batch-based data ingestion workloads: These are
usually ingested in fixed time intervals. Here too, the
workloads are varied.
 Input data like clickstream usually have a lot of
JSON which requires memory bound processing.
 Batch processing of typical structured data have
more crunching which is usually more CPU bound.
 Continuous data ingestion in small data streams.
Some examples are live transaction data which again
could vary between memory and CPU bound processing.
However, the required capacity is much lesser.
 The real reason behind the data ingestion is to derive
insights. Therefore, we have a lot of processing for
which we require data aggregation and summary
calculations. This has a lot of memory processing.
 Along with this, we have workloads for report
calculations, predictive algorithm data processing, ad
hoc queries etc.
While one could optimize it at the time of production,
over time, the usage patterns change. Data ingestion
increases with addition of new data sources. This puts
additional pressure on the platform. The increased data
processing also requires more time for KPI calculations,
leading to contention in resources, in turn also limiting
the time and resource available for ad hoc queries. E.g.
One Hive query could hog the entire cluster.
As a result, customers experience both under-utilized
capacity and a capacity-crunch because of fluctuating
demand. It is also not easy to scale up or down on
demand due to the specific machine requirements (e.g.
EC2 instance types with ephemeral storage as an
additional cost). This means that new environments
cannot be brought up quickly.
One of the promises of new data technology on the
cloud is serverless and on-demand data infrastructure.
Separating the workload types and processing times
helps us plan for the same. If we look at Databricks,
which is designed for running spark effectively in the
cloud, we see a built-in cluster manager, which provides
features like auto scaling and auto termination.
Databricks also provides connectors with cloud storage
and integrated notebook environment which helps with
development immensely.
Given the power of auto scaling and auto termination,
one could design the environment more optimally for
the workload types as outlined in the example.
Workload Types and Cluster Configurations
Workload Type Cluster Type Cost Benefit
Memory Bound Data
Ingestion Jobs
Spot instances help keep the
cost down. No issues with noisy
neighbours
Having auto-scaling instead of
a fixed cluster helps keep the
cost down and maintain the
balance between on-demand
and spot instances
This provides more control
based on the job. E.g. One can
run a complete cluster based
on spot instances for a data
exploration workload
Same as above
Run these workloads in their own cluster which is a
combination of on-demand and spot instances. Use
instances which are memory optimized. Shut down
once the job is done
The workload requests are varied across different
departments. Using a shared cluster designed for
auto-scaling helps manage the varied demand
Depending on the usage being time-bound, one can
design a specific cluster with pre-built libraries for
easier configuration
Similar to the earlier case, but on a cluster with
compute optimized instances
CPU-intensive Data
Ingestion Jobs
Ad hoc Queries
Reports, KPI Crunching,
Statistical Models

Performance Difference between Data Environments
In addition to cost, using the right cluster sizes and types leads to a decrease in processing time. The following
example highlights the same.
Data Processing and
Data Store optimization
While data processing was substantially improved
through cluster and workload optimization, further
improvements can be made by looking at the data stores
and intermediate data processing. Typical Big Data tools
have limitations with respect to source/sinks & type of
data processing - whether batch or streaming. Hadoop
does not integrate well with multiple cloud source/sinks
e.g. Data Warehouse, NoSQL databases. This leads to
multiple tools being used with an additional workflow
(Oozie, Step Functions, Data Pipeline etc.) on top. This
can cause non-optimized code, as data needs to be
written to an intermediate storage multiple times. Also,
there may be delays as developers with disparate skill
sets need to collaborate.
This is best highlighted by the dual use of HBase and
Hive as the data store formats. HBase is used primarily
for updating dimensions and Hive tables for appending
transactions. While HBase is write-optimized, it isn’t as
query-friendly as Hive, which is read-optimized. Therefore,
most systems have both transaction and report stores.
Typically, this leads to data in HBase getting converted
to Hive through a complex intermediate staging layer.
Additionally, hive tables stored on top of parquet files
perform very badly if they need to read many small files.
Hence, ingesting data from streaming applications needs
an additional administrative task of merging small files.
This increases the administrative complexity (e.g.
merging small files), while increasing the data process-
ing time.
Having a common data store and processing (data
management system) can greatly alleviate this pain of
multiple processing technologies and data stores. This
starts with agreeing on a common open file format for
data storage. Today, Parquet has emerged as the most
common used format, since it is better optimized for fast
query and data compression.
With Databricks, we have delta which is a unified data
management system that fits this need. HBase and Hive
external tables can be replaced with a unified table,
Delta (Parquet-based), which is read-optimized. The
ACID merge features ensure that the performance of
read on HBase tables is comparable to Hive tables. Most
importantly, we don’t have to convert the HBase tables
into Hive tables for downstream analysis.
The solution also enables RDBMS features such as ACID
transactions, UPDATE/Merge and DELETE. Elegant
OPTIMIZE/VACUUM is available for the consolidation/
update of small part files. This makes it easier to clean up
or correct bad data at a record level. This also means that
we don’t have to run additional administrative tasks like
merging small files which translates to redirecting the
available resources for this purpose. Additionally, data
cleanup of Hive tables is easy.
6

From a performance perspective in our experience, we
have seen significant improvements in read and write
speeds across Hive and HBase.
 Improvement in Hive table reads by about 30-40%
on Databricks delta post tuning with techniques like
ZOrdering (co-locate related information in the same
set of files) while reads on HBase tables saw
70-80% improvement.
 In Hive, 60 -70% improvement in inserts was
observed while updates are almost same speed
as HBase
An additional complexity that sometimes arises is when
both batch and stream data sets need to be processed
together. Often, this needs to be done using different
tools which bring their own challenges. Spark and more
specifically Databricks can provide a unified API which
can handle both batch & streaming data sources.
Taken together, this makes the data processing pipeline
more performant and much easier to develop and
maintain. Along the way, there are a few nice touches
that Databricks provides to further improve the processing
speed and improve productivity. Some of these are
 Data caching to improve query and processing speeds
 Schema enforcement and Schema Evolution, which
help manage data changes and evolutions more
effectively.
 Time-travel: Databricks Delta automatically versions
the Big Data that is stored in the data lake allowing
one to access any version of that data. This allows for
audit and roll back data in case of accidental bad
writes or deletes to its original value. Multiple
versions of data can be accessed using either the
timestamp or version number. This is similar to
temporal tables in Amazon RDS for SQL Server, and
was missing from Big Data systems.
Design Consumption
Landscape
The reason for the data infrastructure is to drive insights.
Therefore, the consumption layer which includes the
analytical and reporting store becomes extremely
important. This feeds the reporting layer, data APIs,
machine learning APIs among others. There are two
primary modes of consumption we need to focus on.
These are SQL engine performance and Machine
Learning workspaces.
SQL
In traditional Hadoop-based architectures, Hive is
largely the analytical layer with queries, followed by
HBase in some scenarios. As the performance improve-
ment need increases, we also need the presence of
data-marts and data-warehouses.
It is in this context that we need to view the evolution
of Spark SQL. While Spark was initially only a compute
platform, today, it is a fully functioning SQL engine
capable of interfacing with the JDBC engine. As stated
by Spark, Spark SQL is designed to be compatible with
the Hive Metastore, SerDes and UDFs.
We saw an increased usage of Cloud Datawarehouse
like Snowflake, Azure Synapse, AWS Redshift among
others. A common theme among these architectures is
an increased focus on the outcome as against the
design. i.e. The focus is on SQL query performance. e.g.
AWS Redshift now has Aqua which focuses on query
performance, so does serverless querying on Azure
Synapse and Google Big Query. However, for the
purpose of this paper, the key insight is that SQL contin-
ues to be strong and has emerged as the most important
language for insight generation.
Databricks Delta, in addition to providing ACID compli-
ant tables, also provides faster query execution with
indexing, statistics, and auto-caching support. Like other
organizations, query performance continues to be a key
theme, with the recent introduction of dynamic file
pruning, which increases query performance by 2x to 8x.
This will be followed by a faster JDBC driver going by
the Databricks public roadmap.
Therefore, since the choices have evolved, any moderni-
zation of the data environment will require the evaluation
of these platforms, based on the consumption needs.

Machine Learning
Commercial distributions of Hadoop ship their own
Machine Learning workbench, which allow for secure
and collaborative data science workloads. However,
these collaboration mechanisms are proprietary and not
based on open standards.
The dominant standard today in MLflow: MLflow brings in
the discipline of DevOps to the Machine Learning world.
This helps us track the experiments, code and model
repositories, and experimentation to deployment along
with an integrated notebook environment. Databricks and
AWS provide a couple of options to integrate MLflow in
the Machine Learning workflow.
Databricks Machine Learning Runtime (MLR) provides
scalable clusters that supports popular frameworks like
Keras, Tensorflow, PyTorch, SparkML and Scikit-learn.
MLR enables data scientists and ML practitioners to
rapidly build models using its Auto-ML capabilities.
Managed MLflow can help manage MDLC (Model Devel-
opment Life Cycle) like experimentation, deployment,
and model repository. MLR also supports MLeap and the
portability of models across platforms and flexible
deployments on docker containers, SageMaker and
other cloud provider platforms.
Assess Security and
Governance
This is an often-overlooked part of the assessment
process. The security and operational fitness for Hadoop
environments was designed without the cloud in mind.
Cloud-based Data Lake offerings have evolved from
HDFS-compatible cloud distributions to native cloud
Data Lakes Should be built on proven object storage.
This allows for data organization based on finer grained
time scale partitions, and richer retention and control
policies with seamless identity and role propagation for
data zones.
Current data platforms like Databricks on Cloud there-
fore use the security and operational harness provided
by the cloud providers. With respect to AWS, Databricks
provides controls like IAM credential pass-through to
integrate with the AWS ecosystem. Other AWS principles
like VPC peering, PrivateLink and Policy enforcement
only add to this.
8
Along with this, from a data governance stand-point,
we need an integration with data catalogues. On the
AWS platform, the natural integration is with AWS Glue.
Databricks can leverage Glue as the meta-store, even
across multiple workspaces. All the metadata can reside
in one data catalog, easily accessible across their data
lake which can be accessed from entire data lake. One
advantage of keeping all the metadata in Glue is that it
can be leveraged by other tools in the AWS stack, e.g.
Athena & CloudWatch etc. Having a single meta-store
across all AWS resources brings in significant operational
efficiencies while designing enterprise ETL and reporting,
as one doesn’t have to sync multiple meta-stores,
and can query AWS Glue using powerful built in APIs
additionally.

Data Migration
Based on our experience, data migration from on-prem-
ise Hadoop-backed Data Lake to Databricks on Cloud
needs to be planned and executed across multiple
areas. The data estate includes HDFS Files, Hive and
HBase tables etc.
Feature and control structure mapping, rationalization of
data sets, choice of right migration strategy among
one-time full refresh, incremental copy, parallel run and
optional sync are key blocks in migration planning. A
well-defined and battle-tested Audit-Balance-Control
framework and associated task lists provide guidance for
clean data migration execution. Following is a detail of
the two main approaches.
 One-time full refresh:In this approach, parquet files
for Hive tables can be moved as is into S3 (Object
Storage). We can create external tables on this data
and load them into Databricks delta. However, if you
have dimensional type of data in HBase, you have to
first convert it into Hive and then move the data into
S3 for loading data into delta tables.
 Incremental loads: Incremental loads can be
achieved by using the timestamp of the record
creation date. Using this timestamp, we should get
the data as of that day and write it into parquet files
on S3. Subsequently, the steps outlined above will
remain the same.
Preserving DDLs and Schema is a best practice as
Databricks and Delta use Hive meta-store to persist
table metadata. This not only makes the migration easier
specially for Hive tables, but also helps in the migration
of security policies associated with a particular
column/table. The processing for data loading into Delta
itself can be made faster using multiple clusters, thereby
increasing the parallelism.
The data pipelines can be ported or rewritten depending
on the tools used. Any RDBMS data ingestion pipelines
created using Sqoop can be replaced easily by Spark
jobs, as Spark can ingest from a JDBC source and offer
similar scaling benefits. Any ‘Legacy’ MapReduce job
should be rewritten to take numerous advantages
offered by Spark. Cloud-managed orchestration tools
are recommended for complex workflow management,
while Databricks Jobs API can be used for simple
workflows.
Hive workloads can be migrated to SparkSQL with
minimal changes, thanks to SparkSQL’s high affinity to
Hive and support for UDFs like Hive. Serving Layer (BI
Tool Landscape) can be provided very well by built-in
connections as well as JDBC/ODBC connectors.

What does Mindtree bring
to the table?
Our core philosophy is that data by itself
does not inspire action. We need the right
mix of human and machine intelligence
for real world solutions and insights. Unless
data infrastructure allows enterprise
consumers to experiment and access the
data, this goal cannot be achieved. We have
worked with Global Top 1000 organizations
in achieving their data modernization
journey.
We bring these experiences to modernize
your data environments with certainty.
Our accelerators housed under ‘Decision
Moments,’ backed by the partnership with
the right technology partners accelerates
this journey. Together, we can decrease the
cost per insight, increase the value through
the identiﬁcation of right business use
cases and improve the time to insight.

About Mindtree
Mindtree [NSE: MINDTREE] is a global technology consulting and services company, helping enterprises marry scale with agility to achieve
competitive advantage. “Born digital,” in 1999 and now a Larsen & Toubro Group Company, Mindtree applies its deep domain knowledge to
300+ enterprise client engagements to break down silos, make sense of digital complexity and bring new initiatives to market faster. We
enable IT to move at the speed of business, leveraging emerging technologies and the efficiencies of Continuous Delivery to spur business
innovation. Operating in 18 countries and over 40 offices across the world, we’re consistently regarded as one of the best places to work,
embodied every day by our winning culture made up of over 21,000 entrepreneurial, collaborative and dedicated “Mindtree Minds.”
www.mindtree.com ©Mindtree 2020
About Authors
Manoj Karanth
Global Head- Data science and Engineering, Mindtree
With an industry experience of 20+ years, Manoj Karanth heads the Cloud, Data
Science and Engineering (Data Engineering, ML, AI) teams globally for Mindtree’s
Digital Business. He works with Global 2000 clients to accelerate their cloud-first
journey and become insight-driven businesses using Mindtree's cloud-native technology
platforms. Manoj has built the technology center of excellence with leading cloud and
data analytics partners to deliver differentiated digital solutions. He has lived cloud
and product engineering through his two decade career
Sowjanyakumar Kothapalli
Principal- Big Data Architect, Mindtree
Sowjanyakumar Kothapalli is a Program Architect in our digital practice and has
24 years of experience working with fortune 500 companies; of which 15 years were
spent in building complex data platforms across various industries like infrastructure
management, structured finance, credit risk management & retail
References
https://docs.databricks.com/administration-guide/capacity-planning/cmbp.html
https://databricks.com/blog/2017/05/31/top-5-reasons-for-choosing-s3-over-hdfs.html
https://www.datanami.com/2018/05/16/big-data-file-formats-demystified/
https://databricks.com/blog/2017/10/25/databricks-delta-a-unified-management-system-for-real-time-big-data.html
https://databricks.com/blog/2019/09/24/diving-into-delta-lake-schema-enforcement-evolution.html
https://www.knowledgelens.com/assets/migrating-to-data-lake-3_0.pdf
https://spark.apache.org/docs/latest/sql-migration-guide-hive-compatibility.html
https://databricks.com/blog/2020/04/30/faster-sql-queries-on-delta-lake-with-dynamic-file-pruning.html
https://docs.aws.amazon.com/whitepapers/latest/building-data-lakes/amazon-s3-data-lake-storage-platform.html
https://docs.azuredatabricks.net/_static/notebooks/delta/optimize-python.html
https://spark.apache.org/docs/latest/sql-migration-guide-hive-compatibility.html
https://databricks.com/blog/2020/03/16/security-that-unblocks.html

Steps to Modernize Your Data Ecosystem | Mindtree

Recommended

More Related Content

Similar to Steps to Modernize Your Data Ecosystem | Mindtree (20)

More from AnikeyRoy (11)

Recently uploaded (20)

Steps to Modernize Your Data Ecosystem | Mindtree