Storage Engine Considerations for Your Apache Spark Applications with Mladen Kovacevic
6 likes2,110 views
The document discusses various storage solutions for Apache Spark applications, including HDFS, HBase, Kudu, and Solr, outlining their unique capabilities and ideal use cases. It emphasizes the importance of understanding ingestion and consumption requirements, asking the right questions, and carefully choosing a storage system based on specific use-case needs. Additionally, it provides design patterns and implementation strategies for integrating these storage systems with Spark.
Storage Engine Considerations for Your Apache Spark Applications with Mladen Kovacevic
- 1. Mladen Kovacevic, Senior Solutions Architect Cloudera Inc. Storage Engine Considerations for your Apache Spark Applications #EUdev10
- 2. Outline ⢠Motivation â store your data â where exactly? ⢠Storage Capabilities: â HDFS â HBase â Kudu â Solr ⢠Asking the right questions ⢠Decide on right storage solution 2#EUdev10
- 3. Motivation ⢠Spark, SparkStreaming, SparkSQL â great for processing â need a place to store content ⢠Integration with variety of storage systems ⢠Ingest and consumption requirements â use case! 3#EUdev10
- 4. Design patterns 4#EUdev10 Choosing the right storage for the use case 2006 2007 2016 2008
- 5. HDFS ⢠Distributed file system â cheap, scalable, storage ⢠Immutable â ârecordâ changes are painful ⢠Columnar file formats - ideal for analytics ⢠SQL overlays (SparkSQL, Hive Metastore, more) to define schema Highlights Very high throughput, painful random IO, batch oriented, coding overhead (ie. dealing with small files problems), any file 5#EUdev10
- 6. HDFS design pattern df.write.parquet(â/data/person_tableâ) 6#EUdev10 ⢠Small files accumulate ⢠External processes, or additional application logic to manage these files ⢠Partition management ⢠Manage metadata carefully (depends on ecosystem) ⢠Considerations- changing dimensions (fast/slow) ⢠Late arriving data
- 7. HBase ⢠NoSQL engine, manages files on HDFS ⢠Key-value, distributed storage engine ⢠No data types â just binary fields ⢠Thousands to millions of columns ⢠Store entity data (profiles of people, devices, accounts) Highlights Very fast random IO, low throughput, NRT oriented, challenging BI, no strict data types 7#EUdev10
- 8. HBase design pattern 8#EUdev10 ⢠HBase Connection anywhere in Spark/SparkStreaming app ⢠SparkSQL/DataFrames, Bulk Load ⢠Primary storage for ingestion or complementary preserving state ⢠NoSQL store vs. structured ⢠Near-real-time CDH hbase-spark: https://github.com/cloudera/hbase/tree/cdh5-1.2.0_5.13.0/hbase-spark CDH HBase and Spark docs: http://archive.cloudera.com/cdh5/cdh/5/hbase-1.2.0-cdh5.13.0/book.html#spark Upstream hbase-spark (watch for updates in HBase 2.x release): https://github.com/apache/hbase/tree/master/hbase-spark/ Upstream HBase and Spark book: http://hbase.apache.org/book.html#spark
- 10. Kudu ⢠Storage system for tables of structured data ⢠Bring-your-own-SQL (SparkSQL, Impala), NoSQL-like API, integration with Spark, MapReduce, more.. ⢠Columnar, key partitioning by range and/or hash ⢠Limited number of columns (strongly typed) Highlights Fast random IO, fast throughput, NRT oriented, terrific for BI, structured data 10#EUdev10
- 11. Kudu design pattern df.write.options(kuduOptions).mode(âappendâ).kudu OR kuduContext.insertRows() 11#EUdev10 ⢠DataFrame perfect match for Kudu (structured) ⢠Data available immediately to SQL engines (Impala, SparkSQL) ⢠Ideal case is append with moderate updates Kudu Integration with Spark: http://kudu.apache.org/docs/developing.html#_kudu_integration_with_spark Up and running with Apache Spark on Apache Kudu: https://blog.cloudera.com/blog/2017/02/up-and-running-with-apache-spark-on-apache-kudu/
- 13. Solr ⢠Distributed index enabling search capabilities (Lucene) ⢠Typed, REST API based, search index query processing ⢠Search interface, faceting, integration with HBase storing content (typically) in HDFS Highlights High random IO, low throughput, multi-faceted use cases, NRT oriented, terrific for BI with the right tools (non-SQL), loose schema, data types 13#EUdev10
- 14. Solr design pattern 14#EUdev10 ⢠Prepare Solr document, add to SolrCloud directly OR ⢠Write to HBase, leverage Lily HBase Indexer service to update Solr ⢠Store complete record in HBase, while indexed fields for search in Solr ⢠NRT availability (short soft commits)
- 15. Questions we ask (1) ⢠How many voters have cast their ballots by city thus far in the election, by the second? â streaming data into âvoterâ table, aggregate query, immediate data availability : Kudu ⢠How many people watched last nights game compared to the night before? â daily batch, aggregate query : HDFS parquet 15#EUdev10
- 16. Questions we ask (2) ⢠What version is my device running and how many dropped packets do I have? â streaming entity profile data, metrics may change per release, many updates, specific device, NRT: HBase ⢠Which tweets talk to the housing market, in the 21-30 age group? â streaming, keyword search, facet filtering : Solr 16#EUdev10
- 17. Use case questionnaire ⢠Consumption interface: SQL (JDBC/ODBC) vs. API ⢠Near-real-time requirement for consumers ⢠Ingestion rate (can we keep up?) ⢠Entity vs. Events (time-based) ⢠Append-only vs moderate updates vs many updates ⢠Distinct values in dataset 17#EUdev10
- 18. Storage considerations (1) 18#EUdev10 Criteria SQL interface API interface Near-real-time ingestion Append-only + available for query Appends with moderate updates Mostly updates
- 19. Storage considerations (2) 19#EUdev10 Criteria Entity based data Event based data (time-series) High distinct values Many and unknown attributes Binary data (Images, PDFs, etc) Analytics
- 20. Wrap-up ⢠Review entire use-case end-to-end early ⢠Understand storage capabilities ⢠Ask the right questions (upstream/consumers) ⢠Consider security, architecture and development costs ⢠Decide on the right storage solution 20#EUdev10