Using Hive to interact with HBase, Part 1
This is the first of two posts examining the use of Hive for interaction with HBase tables. The second post is here.
One of the things I’m frequently asked about is how to use HBase from Apache Hive. Not just how to do it, but what works, how well it works, and how to make good use of it. I’ve done a bit of research in this area, so hopefully this will be useful to someone besides myself. This is a topic that we did not get to cover in HBase in Action, perhaps these notes will become the basis for the 2nd edition 😉 These notes are applicable to Hive 0.11.x used in conjunction with HBase 0.94.x. They should be largely applicable to 0.12.x + 0.96.x, though I haven’t tested everything yet.
The hive project includes an optional library for interacting with HBase. This is where the bridge layer between the two systems is implemented. The primary interface you use when accessing HBase from Hive queries is called the
BaseStorageHandler. You can also interact with HBase tables directly via Input and Output formats, but the handler is simpler and works for most uses.
HBase tables from Hive
HBaseStorageHandler to register HBase tables with the Hive metastore. You can optionally specify the HBase table as
EXTERNAL, in which case Hive will not create to drop that table directly – you’ll have to use the HBase shell to do so.
CREATE [EXTERNAL] TABLE foo(...) STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler' TBLPROPERTIES ('hbase.table.name' = 'bar');
The above statement registers the HBase table named
bar in the Hive metastore, accessible from Hive by the name
Under the hood,
HBaseStorageHandler is delegating interaction with the HBase table to
HiveHBaseTableOutputFormat. You can register your HBase table in Hive using those classes directly if you desire. The above statement is roughly equivalent to:
CREATE TABLE foo(...) STORED AS INPUTFORMAT 'org.apache.hadoop.hive.hbase.HiveHBaseTableInputFormat' OUTPUTFORMAT 'org.apache.hadoop.hive.hbase.HiveHBaseTableOutputFormat' TBLPROPERTIES ('hbase.table.name' = 'bar');
Also provided is the
HiveHFileOutputFormat which means it should be possible to generate HFiles for bulkloading from Hive as well. In practice, I haven’t gotten this to work end-to-end (see HIVE-4627).
Registering the table is only the first step. As part of that registration, you also need to specify a column mapping. This is how you link Hive column names to the HBase table’s rowkey and columns. Do so using the
hbase.columns.mapping SerDe property.
CREATE TABLE foo(rowkey STRING, a STRING, b STRING) STORED BY 'org.apache.hadoop.hive.hbase.HBaseStorageHandler' WITH SERDEPROPERTIES ('hbase.columns.mapping' = ':key,f:c1,f:c2') TBLPROPERTIES ('hbase.table.name' = 'bar'); ...
The values provided in the mapping property correspond one-for-one with column names of the hive table. HBase column names are fully qualified by column family, and you use the special token
:key to represent the rowkey. The above
example makes rows from the HBase table
bar available via the Hive table
rowkey maps to the HBase’s table’s rowkey,
c1 in the
f column family, and
c2, also in the
You can also associate Hive’s
MAP data structures to HBase column families. In this case, only the
STRING Hive type is used. The other Hive type currently supported is
BINARY. See the wiki page for more examples.
Interacting with data
With the column mappings defined, you can now access HBase data just like you would any other Hive data. Only simple query predicates are currently supported.
SELECT * FROM foo WHERE ...;
FROM source_hive_table INSERT INTO TABLE my_hbase_table SELECT source_hive_table.* WHERE ...;
Be advised that there is a regression in Hive 0.12.0 which breaks this feature, see HIVE-5515.
There’s still a little finesse required to get everything wired up properly at runtime. The HBase interaction module is completely optional, so you have to make sure it and it’s HBase dependencies are available on Hive’s classpath.
$ export HADOOP_CLASSPATH=... $ hive -e "CREATE TABLE ... STORED BY 'org.apache...HBaseStorageHandler'"
The installation environment could do a better job of handling this for users, but for the time being you must manage it yourself. Ideally the
hive bin script can detect the presence of HBase and automatically make the necessary
CLASSPATH adjustments. This enhancement appears to be tracked in HIVE-2055. The last mile is provided by the distribution itself, ensuring the environment variables are set for
hive. This functionality is provided by BIGTOP-955.
You also need to make sure the necessary jars are shipped out to the MapReduce jobs when you execute your Hive statements. Hive provides a mechanism for shipping additional job dependencies via the auxjars feature.
$ export HIVE_AUX_JARS_PATH=... $ hive -e "SELECT * FROM ..."
I did discover a small bug in HDP-1.3 builds which masks user-specified values of
HIVE_AUX_JARS_PATH. With administrative rights, this is easily fixed by correcting the line in
hive-env.sh to respect an existing value. The
work-around in user scripts is to use the
SET statement to provide a value once you’ve launched the Hive CLI.
SET hive.aux.jars.path = ...
Hive should be able to detect which jars are necessary and add them itself. HBase provides the
TableMapReduceUtils#addDependencyJars methods for this purpose. It appears that this is done in hive-0.12.0, at least according to HIVE-2379.
Much has been said about proper support for predicate pushdown (HIVE-1643, HIVE-2854, HIVE-3617,
HIVE-3684) and data type awareness (HIVE-1245, HIVE-2599). These go hand-in-hand as predicate semantics are defined in terms of the types upon which they operate. More could be done to map Hive’s complex data types like Maps and Structs onto HBase column families as well (HIVE-3211). Support for HBase timestamps is a bit of a mess; they’re not made available to Hive applications with any level of granularity (HIVE-2828, HIVE-2306). The only interaction a user has is via storage handler setting for writing a custom timestamp with all operations.
From a performance perspective, there are things Hive can do today (ie, not dependent on data types) to take advantage of HBase. There’s also the possibility of an HBase-aware Hive to make use of HBase tables as intermediate storage location (HIVE-3565), facilitating map-side joins against dimension tables loaded into HBase. Hive could make use of HBase’s natural indexed structure (HIVE-3634, HIVE-3727), potentially saving huge scans. Currently, the user doesn’t have (any?) control over the scans which are executed. Configuration on a per-job, or at least per-table basis should be enabled (HIVE-1233). That would enable an HBase-savy user to provide Hive with hints regarding how it should interact with HBase. Support for simple split sampling of HBase tables (HIVE-3399) could also be easily done because HBase manages table partitions already.
Other access channels
Everything discussed thus far has required Hive to interact with online HBase RegionServers. Applications may stand to gain significant throughput and enjoy greater flexibility by interacting directly with HBase data persisted to HDFS. This also has the benefit of preventing Hive workloads from interfering with online SLA-bound HBase applications (at least, until we see HBase improvements in QOS isolation between tasks, HBASE-4441).
As mentioned earlier, there is the
HiveHFileOutputFormat. Resolving HIVE-4627 should make Hive a straight-forward way to generate HFiles for bulk loading. Once you’ve created the HFiles using Hive, there’s still the last step of running the
LoadIncrementalHFiles utility to copy and register them in the regions. For this, the
HiveStorageHandler interface will need some kind of hook to influence the query plan as it’s created, allowing it to append steps. Once in place, it should be possible to
SET a runtime flag, switching an
INSERT operation to use bulkload.
HBase recently introduced the table snapshot feature. This allows a user to create a persisted point-in-time view of a table, persisted to HDFS. HBase is able to restore a table from a snapshot to a previous state, and to create an entirely new table from an existing snapshot. Hive does not currently support reading from an HBase snapshot. For that matter, HBase doesn’t yet support MapReduce jobs over snapshots, though the feature is a work in progress (HBASE-8369).
The interface between HBase and Hive is young, but has nice potential. There’s a lot of low-hanging fruit that can be picked up to make things easier and faster. The most glaring issue barring real application development is the impedance mismatch between Hive’s typed, dense schema and HBase’s untyped, sparse schema. This is as much a cognitive problem as technical issue. Solutions here would allow a number of improvements to fall out, including much in the way of performance improvements. I’m hopeful that continuing work to add data types to HBase (HBASE-8089) can help bridge this gap.
Basic operations mostly work, at least in a rudimentary way. You can read data out of and write data back into HBase using Hive. Configuring the environment is an opaque and manual process, one which likely stymies novices from adopting the tools. There’s also the question of bulk operations – support for writing HFiles and reading HBase snapshots using Hive is entirely lacking at this point. And of course, there are bugs sprinkled throughout. The biggest recent improvement is the deprecation of HCatalog’s interface, removing the necessary upfront decision regarding which interface to use.
Hive provides a very usable SQL interface on top of HBase, one which integrates easily into many existing ETL workflows. That interface requires simplifying some of the BigTable semantics HBase provides, but the result will be to open up HBase to a much broader audience of users. The Hive interop compliments extremely well the experience provided by Phoenix. Hive has the benefit of not requiring the deployment complexities currently required by that system. Hopefully the common definition of types will allow a complimentary future.
Try it with Sandbox
Hortonworks Sandbox is a self-contained virtual machine with Apache Hadoop pre-configured alongside a set of hands-on, step-by-step Hadoop tutorials.