HIVE
data warehousing using Hadoop




Facebook Data Team
Motivation

 Structured log and dimension data
  – Well known schemas, different serialization formats (binary/text)
  – Rich data structures – nesting/maps/lists

 Query language over structured data
  – SQL helps in easier adoption by business analysts + reduced learning
    curve for everyone
  – Developers love streaming and direct access to map-reduce
  – Query Language brings together SQL and Streaming

 Data Management
  – Tables/Partitions for easy data addressability
  – Abstractions allow optimizations:
        Organize data for large joins/sampling
        Add indices/manage compression/replication transparently
What is HIVE?
 Mgmt. Web UI



                                                               Map Reduce      HDFS


                              Hive CLI
                  Browsing       Queries    DDL


                 Thrift API                Parser
                                                               Execution
                                           Planner
                                                     Hive QL

                                                                 SerDe
                                                          Thrift Jute JSON..
                MetaStore
Dealing with Structured Data

 Type system
  – Primitive types
  – Recursively build up using Composition/Maps/Lists
 Generic (De)Serialization Interface (SerDe)
  – To recursively list schema
  – To recursively access fields within a row object
 Serialization families implement interface
  – Thrift (Binary and Delimited Text), RecordIO, JSON/PADS(?)
 XPath like field expressions
  – profiles.network[@is_primary=1].id
 Inbuilt DDL
  – Define schema over delimited text files
  – Leverages Thrift DDL
Data Model
                                                     #Partitions=32
                                        Schema       Sort-key=uid
                                                     uid
                                         Library




                  Hash         clicks
               Partitioning
                               views        IP
Logical Partitioning                        userId
                                 …
                                            AdId
/hive/clicks
/hive/clicks/ds=2008-03-25     Tables    Dimensions
/hive/clicks/ds=2008-03-25/0

                       HDFS    MetaStore
MetaStore

 Stores Table/Partition properties:
  –   Table schema and SerDe library
  –   Table Location on HDFS
  –   Logical Partitioning keys and types
  –   Sort column
  –   Mapping from columns to well known Dimensions


 Thrift API
  – Current clients in Php (Web Interface), Python (CLI), Java (Query
    Engine), Perl (Tests)
 Stores all properties in text files
Hive CLI

 Implemented in Python
  – uses MetaStore Thrift API
 DDL:
  – create table/drop table/rename table
  – alter table add column etc.
 Browsing:
  – show tables
  – describe table
  – cat table
 Loading Data
  – load data inpath <path1, …> into table <tablename/partition-spec>]
    [bucketed <N> ways by <dimension>]
 Queries
  – Issue queries in Hive QL.
Hive Query Language

 Philosophy
  – SQL like constructs + Hadoop Streaming


 Query Operators in initial version
  –   Projections
  –   Equijoins and Cogroups
  –   Group by
  –   Sampling


 Output of these operators can be:
  – passed to Streaming mappers/reducers
  – can be stored in another Hive Table
  – can be output to HDFS files
Hive Query Language

 Package these capabilities into a more formal SQL like query language
 in next version
 Introduce other important constructs:
  –   Views
  –   Multi table inserts
  –   Order bys
  –   Select distincts
  –   SQL like column expressions
  –   A bunch of other builtin functions
 Still work in progress
Query Language - Examples

  Multi table inserts

  FROM ad_impressions_stg imps
   INSERT INTO ad_legals/ds=2008-03-08 select imps.* where imps.legal = 1
   INSERT INTO ad_non_legals/ds=2008-03-08 select imps.* where imps.legal = 0


  Joins

 FROM ad_impressions imps, ad_dimensions ads
  INSERT INTO ad_legals_joined select imps.*, ads.campaignid
             JOIN ON(imps.adid, ads.adid)
             WHERE imps.legal = 1
Query Language - Examples

 Group By

 FROM ad_legals_joined imps
   INSERT INTO hdfs://hadoop001:9000/user/ads/adid_uu_summary
               select imps.adid, count_distinct(imps.uid)
               group by(imps.adid)
   INSERT INTO hdfs://hadoop001:9000/user/ads/campaignid_uu_summary
               select imps.campaign_id, count_distinct(imps.uid)
               group by(imps.campaignid)
Query Language – HadoopStreaming

 APPLY ON TABLE

 CREATE OPERATOR filter_legal using ‘exec://filter_legal.py’
        (ts date, adid long, uid long)

 FROM (APPLY filter_legal ON TABLE ad_impression)
   INSERT INTO ad_legals where ts >= ‘2008-03-11’ and ts < ‘2008-03-12’


 APPLY can also be applied after JOIN as reducer script

 FROM ad_impressions imps, ad_dimensions ads
      INSERT INTO ad_legals_joined select imps.*, ads.campaignid
                  JOIN ON(imps.adid, ads.adid)
                  APPLY filter_legal BEFORE OUTPUT
Query Language – Views

 Used for expressing
  – Union alls
  – APPLY operators


 Example

 CREATE VIEW actions
 SELECT photo_views.*
 UNION ALL
 SELECT video_views.*
 UNION ALL
 SELECT profile_views.* …
Hive Usage in Facebook

 Applications:
  – Summarization
       Eg: Daily/Weekly aggregations of impression/click counts
  – Ad hoc Analysis
       Eg: how many group admins broken down by state/country
  – Data Mining (Assembling training data)
       Eg: User Engagement as a function of user attributes
 Usage statistics:
  – Total Users: ~40 (about 25% of engineering !)
  – Hive Data (compressed): 22 TB total, ~200GB incoming per day
  – Jobs over last 7 days:
        Total Jobs: 3514, Projections:821, Joins: 152, Aggregates: 800,
        Loaders: 600
     * Aggregates biased because of multi-stage map-reduce
Conclusion

 Release to Open Source in 3-4 months
 People:
  –   Suresh Anthony (suresh@facebook.com)
  –   Jeff Hammerbacher (jeffh@)
  –   Joydeep Sarma (jssarma@)
  –   Ashish Thusoo (athusoo@)
  –   Pete Wyckoff (pwyckoff@)

Facebook hadoop-summit

  • 1.
    HIVE data warehousing usingHadoop Facebook Data Team
  • 2.
    Motivation Structured logand dimension data – Well known schemas, different serialization formats (binary/text) – Rich data structures – nesting/maps/lists Query language over structured data – SQL helps in easier adoption by business analysts + reduced learning curve for everyone – Developers love streaming and direct access to map-reduce – Query Language brings together SQL and Streaming Data Management – Tables/Partitions for easy data addressability – Abstractions allow optimizations: Organize data for large joins/sampling Add indices/manage compression/replication transparently
  • 3.
    What is HIVE? Mgmt. Web UI Map Reduce HDFS Hive CLI Browsing Queries DDL Thrift API Parser Execution Planner Hive QL SerDe Thrift Jute JSON.. MetaStore
  • 4.
    Dealing with StructuredData Type system – Primitive types – Recursively build up using Composition/Maps/Lists Generic (De)Serialization Interface (SerDe) – To recursively list schema – To recursively access fields within a row object Serialization families implement interface – Thrift (Binary and Delimited Text), RecordIO, JSON/PADS(?) XPath like field expressions – profiles.network[@is_primary=1].id Inbuilt DDL – Define schema over delimited text files – Leverages Thrift DDL
  • 5.
    Data Model #Partitions=32 Schema Sort-key=uid uid Library Hash clicks Partitioning views IP Logical Partitioning userId … AdId /hive/clicks /hive/clicks/ds=2008-03-25 Tables Dimensions /hive/clicks/ds=2008-03-25/0 HDFS MetaStore
  • 6.
    MetaStore Stores Table/Partitionproperties: – Table schema and SerDe library – Table Location on HDFS – Logical Partitioning keys and types – Sort column – Mapping from columns to well known Dimensions Thrift API – Current clients in Php (Web Interface), Python (CLI), Java (Query Engine), Perl (Tests) Stores all properties in text files
  • 7.
    Hive CLI Implementedin Python – uses MetaStore Thrift API DDL: – create table/drop table/rename table – alter table add column etc. Browsing: – show tables – describe table – cat table Loading Data – load data inpath <path1, …> into table <tablename/partition-spec>] [bucketed <N> ways by <dimension>] Queries – Issue queries in Hive QL.
  • 8.
    Hive Query Language Philosophy – SQL like constructs + Hadoop Streaming Query Operators in initial version – Projections – Equijoins and Cogroups – Group by – Sampling Output of these operators can be: – passed to Streaming mappers/reducers – can be stored in another Hive Table – can be output to HDFS files
  • 9.
    Hive Query Language Package these capabilities into a more formal SQL like query language in next version Introduce other important constructs: – Views – Multi table inserts – Order bys – Select distincts – SQL like column expressions – A bunch of other builtin functions Still work in progress
  • 10.
    Query Language -Examples Multi table inserts FROM ad_impressions_stg imps INSERT INTO ad_legals/ds=2008-03-08 select imps.* where imps.legal = 1 INSERT INTO ad_non_legals/ds=2008-03-08 select imps.* where imps.legal = 0 Joins FROM ad_impressions imps, ad_dimensions ads INSERT INTO ad_legals_joined select imps.*, ads.campaignid JOIN ON(imps.adid, ads.adid) WHERE imps.legal = 1
  • 11.
    Query Language -Examples Group By FROM ad_legals_joined imps INSERT INTO hdfs://hadoop001:9000/user/ads/adid_uu_summary select imps.adid, count_distinct(imps.uid) group by(imps.adid) INSERT INTO hdfs://hadoop001:9000/user/ads/campaignid_uu_summary select imps.campaign_id, count_distinct(imps.uid) group by(imps.campaignid)
  • 12.
    Query Language –HadoopStreaming APPLY ON TABLE CREATE OPERATOR filter_legal using ‘exec://filter_legal.py’ (ts date, adid long, uid long) FROM (APPLY filter_legal ON TABLE ad_impression) INSERT INTO ad_legals where ts >= ‘2008-03-11’ and ts < ‘2008-03-12’ APPLY can also be applied after JOIN as reducer script FROM ad_impressions imps, ad_dimensions ads INSERT INTO ad_legals_joined select imps.*, ads.campaignid JOIN ON(imps.adid, ads.adid) APPLY filter_legal BEFORE OUTPUT
  • 13.
    Query Language –Views Used for expressing – Union alls – APPLY operators Example CREATE VIEW actions SELECT photo_views.* UNION ALL SELECT video_views.* UNION ALL SELECT profile_views.* …
  • 14.
    Hive Usage inFacebook Applications: – Summarization Eg: Daily/Weekly aggregations of impression/click counts – Ad hoc Analysis Eg: how many group admins broken down by state/country – Data Mining (Assembling training data) Eg: User Engagement as a function of user attributes Usage statistics: – Total Users: ~40 (about 25% of engineering !) – Hive Data (compressed): 22 TB total, ~200GB incoming per day – Jobs over last 7 days: Total Jobs: 3514, Projections:821, Joins: 152, Aggregates: 800, Loaders: 600 * Aggregates biased because of multi-stage map-reduce
  • 15.
    Conclusion Release toOpen Source in 3-4 months People: – Suresh Anthony (suresh@facebook.com) – Jeff Hammerbacher (jeffh@) – Joydeep Sarma (jssarma@) – Ashish Thusoo (athusoo@) – Pete Wyckoff (pwyckoff@)