Hypertable Doug Judd Zvents Inc Background hypertable org

Hypertable Doug Judd Zvents, Inc.

Background hypertable. org

Web 2. 0 = Data Explosion Web 1. 0 Web 2. 0

Traditional Tools Don’t Scale Well § Designed for a single machine § Typical scaling solutions § ad-hoc § manual/static resource allocation

The Google Stack § Google File System (GFS) § Map-reduce § Bigtable hypertable. org

Architectural Overview hypertable. org

What is Hypertable? § A open source high performance, scalable database, modelled after Google's Bigtable § Not relational § Does not support transactions hypertable. org

Hypertable Improvements Over Traditional RDBMS § Scalable § High random insert, update, and delete rate

Data Model § Sparse, two-dimensional table with cell versions § Cells are identified by a 4 -part key § § Row Column Family Column Qualifier Timestamp hypertable. org

Table: Visual Representation hypertable. org

Table: Actual Representation hypertable. org

Anatomy of a Key § § Row key is terminated Column Family is represented with 1 byte Column qualifier is terminated Timestamp is stored big-endian ones-compliment hypertable. org

Concurrency § Bigtable uses copy-on-write § Hypertable uses a form of MVCC (multi-version concurrency control) § Deletes are carried out by inserting “delete” records

Cell. Store § Sequence of 65 K blocks of compressed key/value pairs

System Overview

Range Server § Manages ranges of table data § Caches updates in memory (Cell. Cache) § Periodically spills (compacts) cached updates to disk (Cell. Store) hypertable. org

Client API class Client { void create_table(const String &name, const String &schema); Table *open_table(const String &name); String get_schema(const String &name); void get_tables(vector<String> &tables); }; void drop_table(const String &name, bool if_exists);

Client API (cont. ) class Table { Table. Mutator *create_mutator(); Table. Scanner *create_scanner(Scan. Spec &scan_spec); }; class Table. Mutator { void set(Key. Spec &key, const void *value, int value_len); void set_delete(Key. Spec &key); void flush(); }; class Table. Scanner { bool next(Cell. T &cell); }; hypertable. org

Language Bindings § Currently C++ only § Thrift Broker

Write Ahead Commit Log § Persists all modifications (inserts and deletes) § Written into underlying DFS

Range Meta-Operation Log § Facilitates Range meta operation § Loads § Splits § Moves § Part of Master and Range. Server § Ensures Range state and location consistency

Compression § Cell Stores store compressed blocks of key/value pairs § Commit Log stores compressed blocks of updates § Supported Compression Schemes § § § zlib (--best and --fast) lzo quicklz bmz none hypertable. org

Caching § Block Cache § Caches Cell. Store blocks § Blocks are cached uncompressed § Query Cache § Caches query results § TBD hypertable. org

Bloom Filter § Negative Cache § Probabilistic data structure § Indicates if key is not present

Scaling (part I) hypertable. org

Scaling (part II) hypertable. org

Scaling (part III) hypertable. org

Access Groups § Provides control of physical data layout -hybrid row/column oriented § Improves performance by minimizing I/O CREATE TABLE crawldb { Title MAX_VERSIONS=3, Content MAX_VERSIONS=3, Page. Rank MAX_VERSIONS=10, Click. Rank MAX_VERSIONS=10, ACCESS GROUP default (Title, Content), ACCESS GROUP ranking (Page. Rank, Click. Rank) }; hypertable. org

Filesystem Broker Architecture § Hypertable can run on top of any distributed filesystem (e. g. Hadoop, KFS, etc. ) hypertable. org

Keys To Performance § C++ § Asynchronous communication

C++ vs. Java § Hypertable is CPU intensive § Manages large in-memory key/value map § Alternate compression codecs (e. g. BMZ) § Hypertable is memory intensive § Java uses 2 -3 times the amount of memory to manage large in-memory map (e. g. Tree. Map) § Poor processor cache performance hypertable. org

Performance Test (AOL Query Logs) § 75, 274, 825 inserted cells § 8 node cluster § § § § § 1 1. 8 GHz Dual-core Opteron 4 GB RAM 3 x 7200 RPM SATA drives Average row key: 7 bytes Average value: 15 bytes Replication factor: 3 4 simultaneous insert clients 500 K random inserts/s 680 K scanned cells/s hypertable. org

Performance Test II § Simulated AOL query log data § 1 TB data § 9 node cluster § § § § 1 2. 33 GHz quad-core Intel 16 GB RAM 3 x 7200 RPM SATA drives Average row key: 9 bytes Average value: 18 bytes Replication factor: 3 4 simultaneous insert clients Over 1 M random inserts/s (sustained) hypertable. org

Weaknesses § Range data managed by a single range server § Though no data loss, can cause periods of unavailability § Can be mitigated with client-side cache or memcached hypertable. org

Project Status § Currently in “alpha” § Just released version 0. 9. 0. 7 § Will release “beta” version end of August § Waiting on Hadoop JIRA 1700 hypertable. org

License § GPL 2. 0 § Why not Apache?

Questions? § www. hypertable. org