Introducing Apache Kudu (Incubating) - Montreal HUG May 2016
- 1. 1© Cloudera, Inc. All rights reserved. Introducing Apache Kudu (incubating) Mladen Kovacevic | Solutions Architect May 2016 Adapted from Jeremy Beard’s original presentation in New York, Dec 2015
- 2. 2© Cloudera, Inc. All rights reserved. Presenter • Mladen Kovacevic • Solutions Architect at Cloudera (Toronto based) • 2+ yrs in Big Data • 3+ yrs building workload optimized systems, exploiting hardware features to improve software performance and capability (compute, storage, networking) • 6+ yrs engine development enterprise RDBMS mladen@cloudera.com
- 3. 3© Cloudera, Inc. All rights reserved. Current storage landscape in Hadoop HDFS (GFS) excels at: • Scanning large amounts of data • Accumulating data with high throughput HBase (BigTable) excels at: • Efficiently finding and writing individual rows • Making data mutable Gaps exist when these properties are needed simultaneously
- 4. 4© Cloudera, Inc. All rights reserved. Managing the gap (today) Code Complexity • Manage flow and sync of data between HDFS and HBase Monitoring and Security • Managing consistent backups, security policies, monitoring and more is hard Performance • Significant lag between arrival of HBase data “staging” and time when data is available for analytics.
- 5. 5© Cloudera, Inc. All rights reserved. Hardware landscape is changing rethink design! • Spinning disk (HDD) -> solid state storage (SSD) • NAND flash: Up to 450k read 250k write IOPS, about 2GB/sec read and 1.5GB/sec write throughput, at a price of less than $3/GB and dropping • 3D XPoint memory – persistent storage (1000x faster than NAND, cheaper than RAM) • RAM is cheaper and more abundant: • 64->128->256GB over last few years • Takeaway 1: The next bottleneck is CPU, and current storage heavy applications weren’t designed with CPU efficiency in mind • Takeaway 2: Column stores are feasible for random access
- 6. 6© Cloudera, Inc. All rights reserved. Apache Kudu (Incubating) Storage for Fast Analytics on Fast Data • New updatable column store for Hadoop • Simplifies the architecture for building analytic applications on changing data • Designed for fast analytic performance • Natively integrated with Hadoop • Donated as incubating project at Apache Software Foundation (November 17, 2015) • Beta v0.8.0 now available STRUCTURED Sqoop UNSTRUCTURED Kafka, Flume PROCESS, ANALYZE, SERVE UNIFIED SERVICES RESOURCE MANAGEMENT YARN SECURITY Sentry, RecordService FILESYSTEM HDFS RELATIONAL Kudu NoSQL HBase STORE INTEGRATE BATCH Spark, Hive, Pig MapReduce STREAM Spark SQL Impala SEARCH Solr SDK Kite
- 7. 7© Cloudera, Inc. All rights reserved. • High throughput for big scans (columnar storage and replication) Goal: Within 2x of Parquet • Low-latency for short accesses (primary key indexes and quorum design) Goal: 1ms read/write on SSD • Database-like semantics (initially single-row ACID) • Relational data model (tables/schemas) • Easily integrate with SQL engines (Impala, Drill, Hive) • “NoSQL” style scan/insert/update (Java, C++, Python clients) Kudu design goals
- 8. 8© Cloudera, Inc. All rights reserved. What Kudu is not • Not a SQL interface • Just the storage layer • “BYOSQL” – Bring-your-own SQL • Not a file system • Data must have tabular structure • Not an application that runs on HDFS • An alternative, native Hadoop storage engine • Not a replacement for HDFS or HBase • Select the right storage for the right use case • Cloudera will continue to support and invest in all three
- 9. 9© Cloudera, Inc. All rights reserved. What Kudu is STORAGE SYSTEM for TABLES of STRUCTURED DATA No SQL engine, nor query planner, nor optimizer. Storage system that lets you get at your data quickly (random/scan). Highly scalable, with ability to provide higher level systems info to exploit its abilities (ie. locality). Data logically presented to clients as rows and finite set of columns Stored in columnar format Tables are partitioned across tablets managed by tablet servers. Tablets are replicated managed by Raft consensus. Strongly typed columns Finite number of columns Add/remove columns solely by ALTER TABLE statements
- 10. 10© Cloudera, Inc. All rights reserved. About Kudu • Apache-licensed open source software • Structured data model • Basic construct: tables • Tables broken down into tablets (roughly equivalent to partitions) • Architecture supports geographically disparate, active/active systems • Not the initial design goal
- 11. 11© Cloudera, Inc. All rights reserved. Kudu data model • Tables have a RDBMS-like schema • Finite number of columns (unlike HBase/Cassandra) • Types: BOOL, INT8/16/32/64, FLOAT, DOUBLE, STRING, BINARY, TIMESTAMP • Some subset of columns make up a primary key • Fast random reads/writes by primary key • No secondary indexes (yet) • Columnar layout on disk – Parquet-like • Lazy materialization • Encoding and compression options 11
- 12. 12© Cloudera, Inc. All rights reserved. Table partitioning • Hash bucketing • Distribute records by hash of partition column(s) • N buckets leads to N tablets • Range partitioning • Distribute records by ranges of the partition column(s) • N split keys leads to N tablets • Can be a mix for different columns of the primary key
- 13. 13© Cloudera, Inc. All rights reserved. Kudu tables are partitioned into “tablets” (regions) Partitioning based on primary key (PK) • Native support for range partitioning and/or hash partitioning (salting) • Hash example: PRIMARY KEY (tweet_id) DISTRIBUTE BY HASH(tweet_id) INTO 100 BUCKETS
- 14. 14© Cloudera, Inc. All rights reserved. Each tablet is fault tolerant via Raft consensus • A single Tablet Server can host many tablets • Metadata is stored on just another tablet, but only Master Server processes host that tablet • Raft consensus: • Strong consistency • Leader election on failure • Replication factor 3 or 5 is typical
- 15. 15© Cloudera, Inc. All rights reserved. Consistency model • Consistency and replication enforced by Raft consensus (similar to Paxos) • Replication by operation not data • Single-row transactions now • Multi-row transactions later • Geo-distributed replicas will be possible under strict time synchronization • Techniques drawn from Google Spanner and others
- 16. 16© Cloudera, Inc. All rights reserved. Kudu interfaces • NoSQL-style APIs • Insert(), Update(), Delete(), Scan() • Java, C++, limited Python • Integrations with MapReduce, Spark, and Impala • No direct access to underlying Kudu tablet files • Beta does not have authentication, authorization, encryption
- 17. 17© Cloudera, Inc. All rights reserved. Impala integration • Opens up Kudu to JDBC/ODBC clients • Intuitive way to get data into Kudu • INSERT INTO kudu_table SELECT * FROM src_table; • Additional commands • UPDATE • DELETE • Efficient INSERT VALUES • Runs on the Kudu C++ client
- 18. 18© Cloudera, Inc. All rights reserved. Performance characteristics Very CPU efficient • Written in modern C++, uses specialized CPU instructions (SIMD), JIT compilation with LLVM Latency dependent on storage hardware capabilities • Expect sub-millisecond response on SSDs and upcoming technologies No garbage collection allows very large memory footprint with no pauses Bloom filters reduce the need for many disk accesses
- 19. 19© Cloudera, Inc. All rights reserved. Operating Kudu • Easiest through Cloudera Manager integration • Separate parcel for now • Kudu is always compacting • No minor vs. major compaction • No compaction latency spikes • Web UI is full of metrics and logs
- 20. 20© Cloudera, Inc. All rights reserved. Cluster layout • One or multiple masters for failover • Only one in current beta • Low CPU and memory impact • One tablet server per worker node • Can share disks with HDFS • One SSD per worker node just for Kudu WAL can speed up writes • No dependencies on other Hadoop ecosystem components • But interfacing components like Impala or Spark do
- 21. 21© Cloudera, Inc. All rights reserved. Kudu: Cluster metadata management • Replicated master • Acts as a tablet directory • Acts as a catalog (which tables exist, etc) • Acts as a load balancer (tracks TS liveness, re-replicates under-replicated tablets) • Caches all metadata in RAM for high performance • Under heavy load, 99.99% response times still in microseconds (650us) • Client configured with master addresses • Client asks master for tablet locations as needed and caches them locally
- 22. 22© Cloudera, Inc. All rights reserved. Real-time analytics in Hadoop today Merging in new data = storage complexity Downsides: ● Multiple storage layers ● Latest data is hidden ● Files are messy ● Complex to do updates without breaking running queriesNew Partition Most Recent Partition Historic Data HBase Parquet File Have we accumulated enough data? Reorganize HBase file into Parquet • Wait for running operations to complete • Define new Impala partition referencing the newly written Parquet file Incoming Data (Messaging System) Reporting Request HDFS + Impala
- 23. 23© Cloudera, Inc. All rights reserved. Real-time analytics in Hadoop with Kudu Improvements: ● One system to operate ● No schedules or background processes ● Handle late arrivals or data corrections with ease ● New data available immediately for analytics or operations Historical and Real-time Data Incoming Data (Messaging System) Reporting Request Kudu + Impala
- 24. 24© Cloudera, Inc. All rights reserved. Kudu for data warehousing • Near real time data visibility • BI tools can display events that happened seconds earlier • Excellent for star schemas • Fast scans of deep fact tables • Efficient wide fact tables • Simplified updates of slowly changing dimensions
- 25. 25© Cloudera, Inc. All rights reserved. Near real time data warehousing on Kudu Files RDBMS Streams K A F K A K U D U IMPALA HUE BI tools User FLUME SPARK STREAMING Simple Complex
- 26. 26© Cloudera, Inc. All rights reserved. Resources Join the community http://getkudu.io kudu-user@googlegroups.com Download the beta cloudera.com/downloads Read the whitepaper getkudu.io/kudu.pdf
- 27. 27© Cloudera, Inc. All rights reserved. Demo • Load a standard HDFS Parquet table • Create a Kudu table via Impala • INSERT/UPDATE/DELETE statement with Impala • Spark insert into Kudu table sample job • Reading from Parquet files as input • Map task to insert each row into Kudu
- 28. 28© Cloudera, Inc. All rights reserved. Creating a Kudu table - Impala Kudu Storage handler Table name in Impala does NOT match table name in Kudu. Kudu is its own storage layer. Kudu Master hostname and port A primary key is mandatory
- 29. 29© Cloudera, Inc. All rights reserved. Spark (Scala) code – Java APIDataFrame Row Kudu Master hostname and port Kudu table name Create a client, session and table object Extract values from the row, strong types Create an insert object and row Set the values by type, column name and column valule Perform the actual insert Cleanup
- 30. 30© Cloudera, Inc. All rights reserved. Kudu with Spark DataFrames Tie in Kudu with Spark Map of the Kudu table name, and Kudu master Table name to refer to within Spark SparkSQL statement
- 31. 31© Cloudera, Inc. All rights reserved. Kudu code examples and docs https://github.com/cloudera/kudu-examples http://www.cloudera.com/documentation/betas/kudu/0-7- 0/topics/kudu_development.html http://getkudu.io/docs/developing.html
- 32. 32© Cloudera, Inc. All rights reserved. Thank you mladen@cloudera.com
Editor's Notes
- #29: Copy and paste text of the above slide: CREATE TABLE kudu_table ( cityid STRING , event_dtm STRING , temperature DOUBLE ) TBLPROPERTIES ( 'storage_handler' = 'com.cloudera.kudu.hive.KuduStorageHandler' , 'kudu.table_name' = 'table_in_kudu' , 'kudu.master_addresses' = 'mladen-kudu-1.vpc.cloudera.com:7051' , 'kudu.key_columns' = 'cityid,event_dtm' );
- #31: Copy and paste text of the above slide: CREATE TABLE kudu_table ( cityid STRING , event_dtm STRING , temperature DOUBLE ) TBLPROPERTIES ( 'storage_handler' = 'com.cloudera.kudu.hive.KuduStorageHandler' , 'kudu.table_name' = 'table_in_kudu' , 'kudu.master_addresses' = 'mladen-kudu-1.vpc.cloudera.com:7051' , 'kudu.key_columns' = 'cityid,event_dtm' );