Hive HBase Integration
Introduction
This page documents the Hive/HBase integration support originally introduced in HIVE-705. This feature allows Hive QL statements to access HBase tables for both read (SELECT) and write (INSERT). It is even possible to combine access to HBase tables with native Hive tables via joins and unions.
A presentation is available from the HBase HUG10 Meetup
This feature is a work in progress, and suggestions for its improvement are very welcome.
Storage Handlers
Before proceeding, please read StorageHandlers for an overview of the generic storage handler framework on which HBase integration depends.
Usage
The storage handler is built as an independent module, hive-hbase-handler-x.y.z.jar, which must be available on the Hive client auxpath, along with HBase and Zookeeper jars. It also requires the correct configuration property to be set in order to connect to the right HBase master. See the HBase documentation for how to set up an HBase cluster.
Here's an example using CLI from a source build environment, targeting a single-node HBase server. (Note that the jar locations and names have changed in Hive 0.7, so for earlier releases, some changes are needed.)
$HIVE_SRC/build/dist/bin/hive --auxpath $HIVE_SRC/build/dist/lib/hive-hbase-handler-0.7.0.jar,$HIVE_SRC/build/dist/lib/hbase-0.89.0-SNAPSHOT.jar,$HIVE_SRC/build/dist/lib/zookeeper-3.3.1.jar -hiveconf hbase.master=hbase.yoyodyne.com:60000
Here's an example which instead targets a distributed HBase cluster where a quorum of 3 zookeepers is used to elect the HBase master:
$HIVE_SRC/build/dist/bin/hive --auxpath $HIVE_SRC/build/dist/lib/hive-hbase-handler-0.7.0.jar,$HIVE_SRC/build/dist/lib/hbase-0.89.0-SNAPSHOT.jar,$HIVE_SRC/build/dist/lib/zookeeper-3.3.1.jar -hiveconf hbase.zookeeper.quorum=zk1.yoyodyne.com,zk2.yoyodyne.com,zk3.yoyodyne.com
The handler requires Hadoop 0.20 or higher, and has only been tested with dependency versions hadoop-0.20.x, hbase-0.89.0 and zookeeper-3.3.1. If you are not using hbase-0.89.0, you will need to rebuild the handler with the HBase jar matching your version, and change the --auxpath above accordingly. Failure to use matching versions will lead to misleading connection failures such as MasterNotRunningException since the HBase RPC protocol changes often.
In order to create a new HBase table which is to be managed by Hive, use the STORED BY clause on CREATE TABLE:
CREATE TABLE hbase_table_1(key int, value string)
STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler'
WITH SERDEPROPERTIES ("hbase.columns.mapping" = ":key,cf1:val")
TBLPROPERTIES ("hbase.table.name" = "xyz");
The hbase.columns.mapping property is required and will be explained in the next section. The hbase.table.name property is optional; it controls the name of the table as known by HBase, and allows the Hive table to have a different name. In this example, the table is known as hbase_table_1 within Hive, and as xyz within HBase. If not specified, then the Hive and HBase table names will be identical.
After executing the command above, you should be able to see the new (empty) table in the HBase shell:
$ hbase shell
HBase Shell; enter 'help<RETURN>' for list of supported commands.
Version: 0.20.3, r902334, Mon Jan 25 13:13:08 PST 2010
hbase(main):001:0> list
xyz
1 row(s) in 0.0530 seconds
hbase(main):002:0> describe "xyz"
DESCRIPTION ENABLED
{NAME => 'xyz', FAMILIES => [{NAME => 'cf1', COMPRESSION => 'NONE', VE true
RSIONS => '3', TTL => '2147483647', BLOCKSIZE => '65536', IN_MEMORY =>
'false', BLOCKCACHE => 'true'}]}
1 row(s) in 0.0220 seconds
hbase(main):003:0> scan "xyz"
ROW COLUMN+CELL
0 row(s) in 0.0060 seconds
Notice that even though a column name "val" is specified in the mapping, only the column family name "cf1" appears in the DESCRIBE output in the HBase shell. This is because in HBase, only column families (not columns) are known in the table-level metadata; column names within a column family are only present at the per-row level.
Here's how to move data from Hive into the HBase table (see GettingStarted for how to create the example table pokes in Hive first):
INSERT OVERWRITE TABLE hbase_table_1 SELECT * FROM pokes WHERE foo=98;
Use HBase shell to verify that the data actually got loaded:
hbase(main):009:0> scan "xyz" ROW COLUMN+CELL 98 column=cf1:val, timestamp=1267737987733, value=val_98 1 row(s) in 0.0110 seconds
And then query it back via Hive:
hive> select * from hbase_table_1; Total MapReduce jobs = 1 Launching Job 1 out of 1 ... OK 98 val_98 Time taken: 4.582 seconds
Inserting large amounts of data may be slow due to WAL overhead; if you would like to disable this, make sure you have HIVE-1383, and then issue this command before the INSERT:
set hive.hbase.wal.enabled=false;
Warning: disabling WAL may lead to data loss if an HBase failure occurs, so only use this if you have some other recovery strategy available.
If you want to give Hive access to an existing HBase table, use CREATE EXTERNAL TABLE:
CREATE EXTERNAL TABLE hbase_table_2(key int, value string)
STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler'
WITH SERDEPROPERTIES ("hbase.columns.mapping" = "cf1:val")
TBLPROPERTIES("hbase.table.name" = "some_existing_table");
Again, hbase.columns.mapping is required (and will be validated against the existing HBase table's column families), whereas hbase.table.name is optional.
Column Mapping
The column mapping support currently available is somewhat cumbersome and restrictive:
- for each Hive column, the table creator must specify a corresponding entry in the comma-delimited
hbase.columns.mappingstring (so for a Hive table with n columns, the string should have n entries); whitespace should not be used in between entries since these will be interperted as part of the column name, which is almost certainly not what you want - a mapping entry must be either
:keyor of the formcolumn-family-name:column-name - there must be exactly one
:keymapping (we don't support compound keys yet) - (note that before HIVE-1228,
:keywas not supported, and the first Hive column implicitly mapped to the key; as of HIVE-1228, it is now strongly recommended that you always specify the key explictly; we will drop support for implicit key mapping in the future) - if no column-name is given, then the Hive column will map to all columns in the corresponding HBase column family, and the Hive MAP datatype must be used to allow access to these (possibly sparse) columns
- there is currently no way to access the HBase timestamp attribute, and queries always access data with the latest timestamp.
- since HBase does not associate datatype information with columns, the serde converts everything to string representation before storing it in HBase; there is currently no way to plug in a custom serde per column
- it is not necessary to reference every HBase column family, but those that are not mapped will be inaccessible via the Hive table; it's possible to map multiple Hive tables to the same HBase table
The next few sections provide detailed examples of the kinds of column mappings currently possible.
Multiple Columns and Families
Here's an example with three Hive columns and two HBase column families, with two of the Hive columns (value1 and value2) corresponding to one of the column families (a, with HBase column names b and c), and the other Hive column corresponding to a single column (e) in its own column family (d).
CREATE TABLE hbase_table_1(key int, value1 string, value2 int, value3 int) STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler' WITH SERDEPROPERTIES ( "hbase.columns.mapping" = ":key,a:b,a:c,d:e" ); INSERT OVERWRITE TABLE hbase_table_1 SELECT foo, bar, foo+1, foo+2 FROM pokes WHERE foo=98 OR foo=100;
Here's how this looks in HBase:
hbase(main):014:0> describe "hbase_table_1"
DESCRIPTION ENABLED
{NAME => 'hbase_table_1', FAMILIES => [{NAME => 'a', COMPRESSION => 'N true
ONE', VERSIONS => '3', TTL => '2147483647', BLOCKSIZE => '65536', IN_M
EMORY => 'false', BLOCKCACHE => 'true'}, {NAME => 'd', COMPRESSION =>
'NONE', VERSIONS => '3', TTL => '2147483647', BLOCKSIZE => '65536', IN
_MEMORY => 'false', BLOCKCACHE => 'true'}]}
1 row(s) in 0.0170 seconds
hbase(main):015:0> scan "hbase_table_1"
ROW COLUMN+CELL
100 column=a:b, timestamp=1267740457648, value=val_100
100 column=a:c, timestamp=1267740457648, value=101
100 column=d:e, timestamp=1267740457648, value=102
98 column=a:b, timestamp=1267740457648, value=val_98
98 column=a:c, timestamp=1267740457648, value=99
98 column=d:e, timestamp=1267740457648, value=100
2 row(s) in 0.0240 seconds
And when queried back into Hive:
hive> select * from hbase_table_1; Total MapReduce jobs = 1 Launching Job 1 out of 1 ... OK 100 val_100 101 102 98 val_98 99 100 Time taken: 4.054 seconds
Hive MAP to HBase Column Family
Here's how a Hive MAP datatype can be used to access an entire column family. Each row can have a different set of columns, where the column names correspond to the map keys and the column values correspond to the map values.
CREATE TABLE hbase_table_1(value map<string,int>, row_key int) STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler' WITH SERDEPROPERTIES ( "hbase.columns.mapping" = "cf:,:key" ); INSERT OVERWRITE TABLE hbase_table_1 SELECT map(bar, foo), foo FROM pokes WHERE foo=98 OR foo=100;
(This example also demonstrates using a Hive column other than the first as the HBase row key.)
Here's how this looks in HBase (with different column names in different rows):
hbase(main):012:0> scan "hbase_table_1" ROW COLUMN+CELL 100 column=cf:val_100, timestamp=1267739509194, value=100 98 column=cf:val_98, timestamp=1267739509194, value=98 2 row(s) in 0.0080 seconds
And when queried back into Hive:
hive> select * from hbase_table_1;
Total MapReduce jobs = 1
Launching Job 1 out of 1
...
OK
{"val_100":100} 100
{"val_98":98} 98
Time taken: 3.808 seconds
Note that the key of the MAP must have datatype string, since it is used for naming the HBase column, so the following table definition will fail:
CREATE TABLE hbase_table_1(key int, value map<int,int>) STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler' WITH SERDEPROPERTIES ( "hbase.columns.mapping" = ":key,cf:" ); FAILED: Error in metadata: java.lang.RuntimeException: MetaException(message:org.apache.hadoop.hive.serde2.SerDeException org.apache.hadoop.hive.hbase.HBaseSerDe: hbase column family 'cf:' should be mapped to map<string,?> but is mapped to map<int,int>)
Illegal: Hive Primitive to HBase Column Family
Table definitions such as the following are illegal because a
Hive column mapped to an entire column family must have MAP type:
CREATE TABLE hbase_table_1(key int, value string) STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler' WITH SERDEPROPERTIES ( "hbase.columns.mapping" = ":key,cf:" ); FAILED: Error in metadata: java.lang.RuntimeException: MetaException(message:org.apache.hadoop.hive.serde2.SerDeException org.apache.hadoop.hive.hbase.HBaseSerDe: hbase column family 'cf:' should be mapped to map<string,?> but is mapped to string)
Key Uniqueness
One subtle difference between HBase tables and Hive tables is that HBase tables have a unique key, whereas Hive tables do not. When multiple rows with the same key are inserted into HBase, only one of them is stored (the choice is arbitrary, so do not rely on HBase to pick the right one). This is in contrast to Hive, which is happy to store multiple rows with the same key and different values.
For example, the pokes table contains rows with duplicate keys. If it is copied into another Hive table, the duplicates are preserved:
CREATE TABLE pokes2(foo INT, bar STRING); INSERT OVERWRITE TABLE pokes2 SELECT * FROM pokes; -- this will return 3 SELECT COUNT(1) FROM POKES WHERE foo=498; -- this will also return 3 SELECT COUNT(1) FROM pokes2 WHERE foo=498;
But in HBase, the duplicates are silently eliminated:
CREATE TABLE pokes3(foo INT, bar STRING) STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler' WITH SERDEPROPERTIES ( "hbase.columns.mapping" = ":key,cf:bar" ); INSERT OVERWRITE TABLE pokes3 SELECT * FROM pokes; -- this will return 1 instead of 3 SELECT COUNT(1) FROM pokes3 WHERE foo=498;
Overwrite
Another difference to note between HBase tables and other Hive tables is that when INSERT OVERWRITE is used, existing rows are not deleted from the table. However, existing rows are overwritten if they have keys which match new rows.
Potential Followups
There are a number of areas where Hive/HBase integration could definitely use more love:
- more flexible column mapping (HIVE-806, HIVE-1245)
- default column mapping in cases where no mapping spec is given
- filter pushdown and indexing (see FilterPushdownDev and IndexDev)
- expose timestamp attribute, possibly also with support for treating it as a partition key
- allow per-table hbase.master configuration
- run profiler and minimize any per-row overhead in column mapping
- user defined routines for lookups and data loads via HBase client API (HIVE-758 and HIVE-791)
- logging is very noisy, with a lot of spurious exceptions; investigate these and either fix their cause or squelch them
Build
Code for the storage handler is located under hive/trunk/hbase-handler.
HBase and Zookeeper dependencies are fetched via ivy.
Tests
Class-level unit tests are provided under hbase-handler/src/test/org/apache/hadoop/hive/hbase.
Positive QL tests are under hbase-handler/src/test/queries. These use a HBase+Zookeeper mini-cluster for hosting the fixture tables in-process, so no free-standing HBase installation is needed in order to run them. To avoid failures due to port conflicts, don't try to run these tests on the same machine where a real HBase master or zookeeper is running.
The QL tests can be executed via ant like this:
ant test -Dtestcase=TestHBaseCliDriver -Dqfile=hbase_queries.q
An Eclipse launch template remains to be defined.
Links
- For information on how to bulk load data from Hive into HBase, see HBaseBulkLoad.
- For another project which adds SQL-like query language support on top of HBase, see HBQL (unrelated to Hive).
Acknowledgements
- Primary credit for this feature goes to Samuel Guo, who did most of the development work in the early drafts of the patch