Comparing Accumulo, Cassandra, and HBase
•
2 likes•889 views
Organizations looking to use a NoSQL data store based on Big Table face a challenge when deciding between alternatives. Often superficial differences are overblown or worse, subtle differences aren't discovered until it's too late. In this talk we compare and contrast Apache Accumulo against Apache Cassandra and Apache HBase, diving deep into design differences and subtleties that may hinder a project only after reaching a certain amount of usage or data storage. – Speaker – Aaron Cordova Co-founder and CTO, Koverse Aaron has built multiple, large-scale, big data systems that are used by the intelligence, defense, finance and healthcare industries. Aaron co-founded Koverse Inc. Prior to that, Aaron spent five years as a researcher for the National Security Agency (NSA) where he developed and deployed into operations dozens of advanced analytical techniques. He is the founder of Apache Accumulo, a scalable and secure data store on top of Apache Hadoop and the author of the recently released O’Reilly book, Accumulo: Application Development, Table Design, and Best Practices. — More Information — For more information see http://www.accumulosummit.com/
Recommended
Recommended
More Related Content
What's hot
What's hot (20)
Similar to Comparing Accumulo, Cassandra, and HBase
Similar to Comparing Accumulo, Cassandra, and HBase (20)
Recently uploaded
Recently uploaded (20)
Comparing Accumulo, Cassandra, and HBase
- 2. We’ll look at • Architecture • Data Model • Security • Query Support • Gotchas • Popularity
- 8. Data Model: Column Visibility • “1st class” component in data model • Security filtering is implemented in a system iterator that can’t be turned off • Column Visibilities are stored as human readable strings – no mapping that could introduce confusion
- 9. Data Model: Locality Groups • Column families can be created dynamically • Column families are put into the default locality group until assigned otherwise • Column Families can be reassigned to locality groups as needed • Provides columnar scanning optimization • Locality groups are simply sections within Rfiles • no additional overhead on HDFS NameNode • no restrictions on column family names
- 11. Security • On by default • User, group/role, access control • “Cell-level” security via Column Visibilities • Fail-safe • Security label operators include & and | or Common problem for new users is writing data they can’t even see.
- 12. Query Support • No secondary indexing capabilities built in • Several secondary indexing patterns are well supported • Support for scanning long rows • BatchScanner for fetching results • Iterators enable storing index entries with records • No official query language
- 13. Some Gotchas • Usually requires tuning beyond what distros provide • Not balancing clients and tablet servers • Having many small tables vs few large tables • Not as many free resources online like blog posts, tutorials, forums, etc. • Larger individual servers mean that server failure can result in a large amount of data needing to be replicated. Accumulo only needs to process recent write-ahead log entries, however, before everything is back online.
- 14. Popularity • 4th most popular ‘wide column store’, behind Cassandra, HBase, and Microsoft Azure Cosmos DB • About a 16th as popular as HBase • 60th most popular DB overall, higher than: • Cloudant • MemSQL • Apache Drill • Oracle NoSQL • Amazon Simple DB • LevelDB • VoltDB • Google Cloud BigTable • MapR-DB https://db-engines.com/en/system/Accumulo
- 16. Recent Improvements in 1.7 • Client Authentication with Kerberos • Data-Center Replication • User-Initiated Compaction Strategies • API Clarification • Faster Startup via Configurable Threadpool Size for Assignments • Group-Commit Threshold as a Factor of Data Size • Balancing Groups of Tablets • User-specified Durability • Hadoop Metrics2 Support • Distributed Tracing with Htrace • Per-Table Volume Chooser • Table and namespace custom properties
- 17. Recent Improvements in 1.8 • Speed up WAL roll overs • User level API for Rfile • Suspend Tablet assignment for rolling restarts • Run multiple Tablet Servers on one node • Rate limiting Major Compactions • Table Sampling
- 23. Architecture “With read replicas enabled, the HMaster distributes read-only copies of regions (replicas) to different RegionServers in the cluster. One RegionServer services the default or primary replica, which is the only replica which can service write requests. If the RegionServer servicing the primary replica is down, writes will fail.” Stale reads are also now possible and unavoidable https://www.cloudera.com/documentation/enterprise/5-4-x/topics/admin_hbase_read_replicas.html
- 30. Data Model: Value Sizes "storing 10-50MB objects in HBase would probably be too much to ask”[1] “Aim to have cells no larger than 10 MB, or 50 MB if you use MOB”[2] 1. http://hbase.apache.org/0.94/book/supported.datatypes.html 2. http://hbase.apache.org/book.html#table_schema_rules_of_thumb
- 32. Security • To use cell level security • Ensure HBase is configured to use v 3 Hfile storage • VisibilityController must be added to the list of co- processors • Setup Hadoop Group Mapping mechanism • By default, visibility labels are lost on replication • ! (not) included as an operator, making it more important to ensure that clients can’t drop user authorization tokens to avoid elevation of privilege
- 34. Gotchas Covered above https://www.cloudera.com/documentation/enterprise/5-6-x/topics/search.html
- 35. Popularity • 2nd most popular ‘wide columnar store’ • About half as popular as Cassandra • Still growing in popularity, but growth has slowed https://db-engines.com/en/system/HBase
- 38. Architecture • Tries to combine parts of BigTable and Amazon’s Dynamo • Designed to span data centers, allows users to choose between CP and AP • Every node is the same, no masters, no zookeeper, storage is coupled with service • Each server still uses a memtable, sstable files on disk, compaction, sorting, etc • Use the ‘gossip’ peer-to-peer protocol
- 40. Data Model • Same as BigTable, but interaction with the data model directly has been completely eclipsed by a more table-like abstraction as part of CQL • Table schemas have to be declared up front • Now features ‘partition keys’, ‘clustering columns’, and other, regular columns
- 42. Data Model Videos by User user_id K uploaded_timestamp C video_id C email first_name last_name title description
- 43. Data Model user_id uploaded timestamp video_id email first_name last_name title description Partition key clustering column clustering column column column column column column Row Col Value hash(user_id)::uploaded_timestamp::video_id email … hash(user_id)::uploaded_timestamp::video_id first_name … hash(user_id)::uploaded_timestamp::video_id last_name … hash(user_id)::uploaded_timestamp::video_id title … hash(user_id)::uploaded_timestamp::video_id description …
- 44. Data Model • All dynamism appears to be absent from the data model • All data has to be modeled up front • Columns are even typed
- 45. Security • Cassandra supports a security model similar to that of relational databases, which supports controlling access to key- spaces, tables, and rows.[1] • Row level access relies on exact string matches.[2] • Column level permissions are not yet implemented.[3] 1 http://docs.datastax.com/en/dse/5.1/dse-admin/datastax_enterprise/security/secPermissions.html 2 http://docs.datastax.com/en/dse/5.1/dse-admin/datastax_enterprise/security/secRlac.html 3 https://issues.apache.org/jira/browse/CASSANDRA-12859
- 46. Query Support • Cassandra features a SQL like language called CQL • Used to specify table schemas • Used for querying • Also features built-in secondary indexing
- 48. Query Support • hash(user_id) implies no range scans over partition keys • only lookups on partition keys or clustering columns are efficient without secondary indexing Row Col Value hash(user_id)::uploaded_timestamp::video_id email … hash(user_id)::uploaded_timestamp::video_id first_name … hash(user_id)::uploaded_timestamp::video_id last_name … hash(user_id)::uploaded_timestamp::video_id title … hash(user_id)::uploaded_timestamp::video_id description …
- 49. Query Support: Caveats • If not including partition keys in your query, have to filter (not allowed by default) • Possible to order partition keys to allow range scans, but not on by default • Without all the clustering columns in a query, it will be rejected unless filtering is on • Can do some range scans on clustering columns
- 50. Query Support: Caveats • Queries on secondary index restricted to equality or ‘contains’ • Non-indexed columns can be part of a query with additional filtering • When querying secondary index, Cassandra must query all partitions but avoids naively querying all at once
- 55. Other Gotchas • Adding a replacement node can take a long time (days) because data must replicate – more smaller servers alleviates this • Hinted handoffs can be problematic • Large number of tables can be problematic • People who run 9000 nodes (Netflix) spread them across 100s of cluster – averaging 100 nodes per cluster
- 56. Popularity • Most popular ‘wide columnar store’ • 8th most popular database overall • Popularity has plateaued somewhat https://db-engines.com/en/system/Cassandra
- 58. So which to use?
- 59. 0 2 4 6 8 10 12 14 16 18 ad built on cloud enterprise gaming marketing mobile social web Relative Usage by Declaration Cassandra Hbase https://en.wikipedia.org/wiki/Apache_Cassandra https://hbase.apache.org/poweredbyhbase.html
- 61. Ideas for Accumulo • Security is still good, but might not be a strong enough differentiator • Grow community support • Offer a secondary indexing with high level query language?
- 63. Questions?