The Data Load Process We looked at the performance that these engines have in the last article, now it’s time to look at how the data got loaded in. There are trade-offs here to be aware of when loading data into each of these engines, as they use different mechanisms to accomplish the task. Hive Load There is no immediate need for a Schema-on-Write data load when you are using Hive with your native file format. Your only “load” operation is copying the files from your local file system to HDFS. With schema-on-read functionality, Hive can instantly access data as soon as its’ underlying file is loaded into HDFS. In our case the real data load step was in converting this Schema-on-Read external Hive table data into optimized ORC format, therefore loading it from an external table to a Hive-managed table. This was a relatively short process, coming in much under an hour. HBase Load Contrast that with HBase, where a bulk data load for our sample data set of 200M rows (around 30GB of disk size in CSV format) took 4+ hours using a single-threaded Java application running in the cluster. In this case, HBase went through a process of taking several columns of the CSV data and concatenating them together to come up with a composite key. This, along with the fact that the inserts were causing hot-spotting within the Region Servers, slowed things down. One way to improve this performance would be to pre-split the regions so your inserts aren’t all going to one region to start with. We could have parallelized the data load as well to improve the performance, writing a MapReduce job to distribute the work. Druid Load Let’s also contrast that with the Druid load, which took about 2 hours. Druid bulk loads data using a MapReduce job; this is a fairly efficient way of doing things since it distributes the work across the cluster and is why we’re seeing a lower time relative to HBase. Druid still has to do the work of adding its own indexes on top of the data and optionally pre-aggregate the data to a certain user-defined level, so it doesn’t have a trivial path to getting the data in either. Although we didn’t choose to pre-aggregate this data, this is what allows Druid to save a lot of space; instead of storing the raw data, Druid can roll the data up to a minute-level granularity if you think your users will not query deeper than that. But remember - Once you aggregate the data, you no longer have the raw data. Space Considerations Another interesting way to slice this data is by how much space it takes up in each of the 3 columnar formats. Engine Size on Disk with Replication Hive - ORC w/ Zlib 28.4GB HBase - Snappy compression 89.5GB Druid 31.5GB Hive and Druid have compressed the data very efficiently considering the initial data size was 90GB with replication, but HBase is sitting right around the raw data size. At this point, we've covered both relative loading times for the three engines as well as data storage space requirements across the three. These may change as you use different compression formats or load different kinds of data into the engines, but this is intended as a general reference to understand relative strengths between the three. ... View more

This article series is an expansion into the technical details behind the Big Data Processing Engines blog post: https://hortonworks.com/blog/big-data-processing-engines-which-one-do-i-use-part-1/ Intro to performance analysis Here, we will be deep diving into the performance side of the three Big Data Processing Engines discussed in the above blog post: Druid, HBase, and Hive. I ran a number of query types to represent the various workloads generally executed on these processing engines, and measured their performance in a consistent manner. I ran a few tests on each of the three engines to showcase their different strengths, and show where some are less effective (but could still fill the gap in a pinch). We will be executing the following queries: Simple count of all records in a table, highlighting aggregation capabilities A select with a where clause, highlighting drill-down and "needle in the haystack" OLTP queries A join, showcasing ad-hoc analysis across the dataset Updates, showcasing scenarios in which data is constantly changing and our dataset needs to stay up to date Performance Analysis An aggregation much like an analyst would do, such as summing data on a column Performance Analysis A few notes about setup: Data size: 200 million rows, 30GB on disk (90GB after replication) Cluster size: 8 nodes, broken down into 2 masters and 6 workers Node size: 8 core, 16 GB RAM machines in a virtualized environment Method of querying: Hive on Tez+LLAP was used to query Hive-managed and Druid-managed data. Phoenix was used to query HBase-managed data. A reasonable configuration was used for each of the engines Cache was taken out of the picture in order to get accurate estimates for initial query execution. Query execution and re-execution times will be much faster with cache in place for each of these engines Note that this is not an ideal setup for Hive+HBase+Druid. Dedicated nodes for each of these services would yield better numbers but we decided to keep it approachable so you can reproduce these results on your own small cluster. As I laid out, the three processing engines performed about how you would expect given their relative strengths and weaknesses. Take a look at the table below. Query HBase/Phoenix (seconds) Hive (seconds) Druid (seconds) Count(*) 281.44 4.72 0.71 Select with filter 1.35 8.71 0.34 Select with join and filter 365.41 9.16 N/A Update with filter 1.52 9.75 N/A Aggregation with filter 353.07 8.66 1.72 Here is that same information in a graph format, with HBase capped at 15s to keep the scale readable. As expected, HBase outshined the other two when it came to ACID operations, with an average of 1.5 seconds on the updates. Druid is not capable of them and Hive took a bit longer. HBase however is not great at aggregation queries, as seen in the ~6 minute query times. Druid is extremely efficient at everything it does, giving no results above 2 seconds and mostly under 1 second. Lastly, Hive with its latest updates has become a real-time database and serviced all queries thrown at it in under 10 seconds. Queries Here are all of the queries that were run, multiple times each, to arrive at the results above. --queries select count(*) from transactions; select count(*) from transactions_hbase; select count(*) from transactions_druid; select trxn_amt,rep_id from transactions_partitioned where trxn_date="2018-10-09" and trxn_hour=0 and trxn_time="2018-10-09 00:33:59.07"; select * from transactions_hbase_simple where row_key>="2018-09-11 12:03:05.860" and row_key<"2018-09-11 12:03:05.861"; select * from transactions_druid where `__time`='2018-09-11 12:03:05.85 UTC'; select distinct(b.name) from transactions_partitioned a join rep b on a.rep_id=b.id where rep_id in (1,2,3) and trxn_amt > 180; select distinct b."name" from "transactions_hbase" a join "rep_hbase" b on a."rep_id"=b."ROW_KEY" where b."ROW_KEY" in (1,2,3) and a."trxn_amt">180.0; update transactions_partitioned set qty=10 where trxn_date="2018-10-09" and trxn_hour=0 and trxn_time>="2018-10-09 00:33:59.07" and trxn_time<"2018-10-09 00:33:59.073"; insert into table transactions_hbase_simple values ('2018-09-11 12:03:05.860~xxx-xxx~xxx xxx~1~2017-02-09', null,null,null,10,null,null,null); select sum(trxn_amt),rep_id from transactions_partitioned group by rep_id; select sum("trxn_amt"),"rep_id" from "transactions_hbase" group by "rep_id"; select sum(trxn_amt),rep_id from transactions_druid group by rep_id; ... View more

With the arrival of the Hive3Streaming processors, performance has never been better between NiFi and Hive3. Below, we'll be taking a look at the PutHive3Streaming (separate from the PutHiveStreaming processor) processor and how it can fit into a basic Change Data Capture workflow. We will be performing only inserts on the source Hive table and then carrying over those inserts into a destination Hive table. Updates are also supported through this process with the addition of a 'last_updated_datetime' timestamp, but that is out of scope for this article. This is meant to simulate copying data from one HDP cluster to another - we're using the same cluster as the source and destination here however. Here is the completed flow. Take note that most of this flow is just keeping track of the latest ID that we've seen, so that we can pull that back out of HDFS periodically and query Hive for records that were added beyond that latest ID. These last two processors, SelectHive3QL and PutHive3Streaming, are the ones doing the heavy lifting. They are the ones first getting the data from the source table (based on the pre-determined latest ID) and then inserting that retrieved data into the destination Hive table. Here's the configuration for the SelectHive3QL processor. Note the ${curr_id} variable used in the HiveQL Select Query field. That ensures our query will be dynamic. This is the configuration for the PutHive3Streaming processor. Nothing special here - we've configured the Hive Configuration Resources with the hive-site.xml file and used the Avro format (above) to retrieve data and (below) to write it back out. Here's the state of the table as we first check the destination table: And then insert a record into the source table: Here's the new state of the table: Here is the full NiFi flow: puthive3streaming-cdc-flow.xml In conclusion, we've shown one of the ways you can utilize the new PutHive3Streaming processors to perform quick inserts into Hive and at a broader level perform Change Data Capture. ... View more

This article is based on the following Kaggle competition: https://www.kaggle.com/arjunjoshua/predicting-fraud-in-financial-payment-services It is a Scala-based implementation of the data science exploration written in Python. In addition to training a model, we also have the ability to batch-evaluate a set of data stored in a file through the trained model. Full configuration, build, and installation instructions can be found at the GitHub repo: https://github.com/anarasimham/anomaly-detection When you execute the model training, you'll get various lines of output as the data is cleaned and the model is built. To view this output, use the link provided by the Spark job console output. This will look like the following: 18/05/10 14:33:58 INFO Client: client token: N/A diagnostics: N/A ApplicationMaster host: <HOST_IP> ApplicationMaster RPC port: 0 queue: default start time: 1525962717635 final status: SUCCEEDED tracking URL: http://<SPARK_SERVER_HOST_NAME>:8088/proxy/application_1525369563028_0053/ user: root The last few lines, which show the trained model, look like this: +-----+-------------------------------------------------------------------------------------------------------------+-----------------------------------------+----------+ |label|features |probabilities |prediction| +-----+-------------------------------------------------------------------------------------------------------------+-----------------------------------------+----------+ |0 |(10,[0,2,5,8,9],[1.0,1950.77,106511.31,-1950.77,104560.54]) |[0.9777384996414185,0.02226151153445244] |0.0 | |0 |(10,[0,2,5,8,9],[1.0,3942.44,25716.56,-3942.44,21774.120000000003]) |[0.9777384996414185,0.02226151153445244] |0.0 | |0 |[1.0,0.0,7276.69,93.0,0.0,1463.0,0.0,0.0,-7183.69,-5813.69] |[0.9777384996414185,0.02226151153445244] |0.0 | |0 |(10,[0,2,5,8,9],[1.0,13614.91,30195.0,-13614.91,16580.09]) |[0.9777384996414185,0.02226151153445244] |0.0 | |0 |[1.0,0.0,17488.56,14180.0,0.0,182385.22,199873.79,0.0,-3308.5600000000013,-34977.130000000005] |[0.9777384996414185,0.02226151153445244] |0.0 | |0 |[1.0,0.0,19772.53,0.0,0.0,44486.99,64259.52,0.0,-19772.53,-39545.06] |[0.9777384996414185,0.02226151153445244] |0.0 | |1 |(10,[0,2,3,7,9],[1.0,20128.0,20128.0,1.0,-20128.0]) |[0.022419333457946777,0.9775806665420532]|1.0 | |0 |[1.0,0.0,33782.98,0.0,0.0,39134.79,16896.7,0.0,-33782.98,-11544.890000000003] |[0.9777384996414185,0.02226151153445244] |0.0 | |0 |[1.0,0.0,34115.82,32043.0,0.0,245.56,34361.39,0.0,-2072.8199999999997,-68231.65] |[0.9777384996414185,0.02226151153445244] |0.0 | The original data is split into training data and test data, and the above is the results of running the test data through the model. The "label" column denotes which label (0=legitimate, 1=fraudulent) the row of data truly falls into The "features" column is all the data that went into training the model, in vectorized format because that is the way the model understands the data The "probabilities" column denotes how likely the model thinks each of the labels is (first number being 0, second number being 1), and the "prediction" column is what the model thinks the data falls into. You can add additional print statements and re-run the training to explore When you execute the evaluation portion of this project (instructions in the GitHub repo), you will re-load the model from disk and use test data from a file to see if the model is predicting correctly. Note that it is a bad practice to use test data from the training set (like I have) but for simplicity I have done that. You can go to the Spark UI as above to view the output. And there you have it, a straightforward approach to building a Gradient Boosted Decision Tree Machine Learning model based off of financial data. This approach can be applied not only to Finance but can be used to train a whole variety of use cases in other industries. ... View more

Repo Description This repository allows one to generate two types of data: Automotive parts production data, like that which comes off an assembly line POS transaction data After generating that data, you can insert it into either a relational store like MySQL or Hive. More details are on the GitHub page Repo Info Github Repo URL https://github.com/anarasimham/data-gen Github account name anarasimham Repo name data-gen ... View more

Repo Description This is a demo project that shows an end-to-end workflow of ingesting streaming data in via NiFi, putting it into a Kafka queue, reading that Kafka queue with Druid, and finally monitoring a custom Python/Flask dashboard for errors. You can additionally use Superset to visualize the data in Druid by creating your own dashboards. The project generates sample data in two possible domains: automotive part production data or POS transactions. Instructions on deployment and configuration are contained in the repo, as well as more detail on context and purpose. Repo Info Github Repo URL https://github.com/anarasimham/manf-part-production Github account name anarasimham Repo name manf-part-production ... View more

If you'd like to generate some data to test out the HDP/HDF platforms at a larger scale, you can use the following GitHub repository: https://github.com/anarasimham/data-gen This will allow you to generate two types of data: Point-of-sale (POS) transactions, containing data such as transaction amount, time stamp, store ID, employee ID, part SKU, and quantity of product. These are transactions you make at a store when you are checking out. For simplicity's sake, this assumes each shopper only buys one product (potentially greater than 1 in quantity) Automotive manufacturing parts production records, simulating the completion of parts in an assembly line. Imagine a warehouse completing different components of a car, such as the hood, front bumper, etc. at different points in time and those parts being tested for heat and vibration thresholds. This data will contain a timestamp of when the part was produced, thresholds for heat & vibration, values as tested for heat & vibration, quanity of produced part, a "short name" identifier for the part, a notes field, and a part location Full details of both schemas are documented in the code in file datagen/datagen.py at the repository above. The application is able to generate data and insert into one of two supported locations: Hive MySQL You will need to configure the table by running one of the scripts in the mysql folder after connecting to the desired server and the desired database as the desired user. Once that is done, you can copy the inserter/mysql.passwd.template file into inserter/mysql.passwd and edit it to provide the correct details. If you'd like to insert into Hive, do the same with the hive.passwd.template file. After editing, you can execute using the following command: python main_manf.py 10 mysql This will insert 10 rows of manufacturing data into the configured MySQL database table. At this point, you're ready to explore your data in greater detail. Possible next steps include using NiFi to pull the data out of MySQL and push into Druid for a dashboard-style data lookup workflow. You can also push into Hive for ad-hoc analyses. These activities are out of scope for this article but are suggestions to think about. ... View more

Assumptions: -You have a running HDP cluster with Sqoop installed -Basic knowledge of Sqoop and its parameters Ingesting SAP HANA data with Sqoop To ingest SAP HANA data, all you need is a JDBC driver. To the HDP platform, HANA is just another database - drop the JDBC driver in and you can plug & play. 1. Download the JDBC driver. This driver is not publicly available - it is only available to customers using the SAP HANA product. Find it on their members-only website and download it. 2. Drop the JDBC driver into Sqoop's lib directory. For me, this is located at /usr/hdp/current/sqoop-client/lib 3. Execute a Sqoop import. This command has many variations and many command-line parameters, but the following is one such example. sqoop import --connect "jdbc:sap://<HANA_SERVER>:30015" --driver com.sap.db.jdbc.Driver --username <YOUR_USERNAME> --password <PASSWORD> --table "<TABLE_NAME>" --target-dir=/path/to/hdfs/dir -m 1 -- --schema "<YOUR_SCHEMA_NAME>" The '-m 1' argument will limit Sqoop to using one thread, so don't use this if you want parallelism. You'll need to use the --split-by argument and give it a column name to be able to parallelize the import work. If all goes well, Sqoop should start importing the data into your target directory. Happy Sqooping! ... View more