Querying and Inserting Data

The Hive query operations are documented in Select, and the insert operations are documented in Inserting data into Hive Tables from queries and Writing data into the filesystem from queries.

Simple Query 简单查询

For all the active users, one can use the query of the following form:

Note that unlike SQL, we always insert the results into a table. We will illustrate later how the user can inspect these results and even dump them to a local file. You can also run the following query on Hive CLI:

This will be internally rewritten to some temporary file and displayed to the Hive client side.

Partition Based Query 分区查询

What partitions to use in a query is determined automatically by the system on the basis of where clause conditions on partition columns. For example, in order to get all the page_views in the month of 03/2008 referred from domain xyz.com, one could write the following query:

Note that page_views.date is used here because the table (above) was defined with PARTITIONED BY(date DATETIME, country STRING) ; if you name your partition something different, don't expect .date to do what you think!

Joins　　表连接

In order to get a demographic breakdown (by gender) of page_view of 2008-03-03 one would need to join the page_view table and the user table on the userid column. This can be accomplished with a join as shown in the following query:

In order to do outer joins the user can qualify the join with LEFT OUTER, RIGHT OUTER or FULL OUTER keywords in order to indicate the kind of outer join (left preserved, right preserved or both sides preserved). For example, in order to do a full outer join in the query above, the corresponding syntax would look like the following query:

In order check the existence of a key in another table, the user can use LEFT SEMI JOIN as illustrated by the following example.

In order to join more than one tables, the user can use the following syntax:

Note that Hive only supports equi-joins. Also it is best to put the largest table on the rightmost side of the join to get the best performance.

目前Hive只支持等连接，并且最好将达标放到表连接的右边。

Aggregations　　聚合操作

In order to count the number of distinct users by gender one could write the following query:

Multiple aggregations can be done at the same time, however, no two aggregations can have different DISTINCT columns .e.g while the following is possible

DISTINCT关键在不允许在两个聚合函数里，针对不同的列。

however, the following query is not allowed

下面这种写法是不允许的：

Multi Table/File Inserts　　多表插入和文件插入

The output of the aggregations or simple selects can be further sent into multiple tables or even to hadoop dfs files (which can then be manipulated using hdfs utilities). e.g. if along with the gender breakdown, one needed to find the breakdown of unique page views by age, one could accomplish that with the following query:

The first insert clause sends the results of the first group by to a Hive table while the second one sends the results to a hadoop dfs files.

Dynamic-Partition Insert 动态分区插入

In the previous examples, the user has to know which partition to insert into and only one partition can be inserted in one insert statement. If you want to load into multiple partitions, you have to use multi-insert statement as illustrated below.

In order to load data into all country partitions in a particular day, you have to add an insert statement for each country in the input data. This is very inconvenient since you have to have the priori knowledge of the list of countries exist in the input data and create the partitions beforehand. If the list changed for another day, you have to modify your insert DML as well as the partition creation DDLs. It is also inefficient since each insert statement may be turned into a MapReduce Job.

Dynamic-partition insert (or multi-partition insert) is designed to solve this problem by dynamically determining which partitions should be created and populated while scanning the input table. This is a newly added feature that is only available from version 0.6.0. In the dynamic partition insert, the input column values are evaluated to determine which partition this row should be inserted into. If that partition has not been created, it will create that partition automatically. Using this feature you need only one insert statement to create and populate all necessary partitions. In addition, since there is only one insert statement, there is only one corresponding MapReduce job. This significantly improves performance and reduce the Hadoop cluster workload comparing to the multiple insert case.

Below is an example of loading data to all country partitions using one insert statement:

There are several syntactic differences from the multi-insert statement:

• country appears in the PARTITION specification, but with no value associated. In this case, country is a dynamic partition column. On the other hand, ds has a value associated with it, which means it is a static partition column. If a column is dynamic partition column, its value will be coming from the input column. Currently we only allow dynamic partition columns to be the last column(s) in the partition clause because the partition column order indicates its hierarchical order (meaning dt is the root partition, and country is the child partition). 目前只支持最后一个分区列为动态分区列。You cannot specify a partition clause with (dt, country='US') because that means you need to update all partitions with any date and its country sub-partition is 'US'.
• An additional pvs.country column is added in the select statement. This is the corresponding input column for the dynamic partition column. Note that you do not need to add an input column for the static partition column because its value is already known in the PARTITION clause. Note that the dynamic partition values are selected by ordering, not name, and taken as the last columns from the select clause.

Semantics of the dynamic partition insert statement:

• When there are already non-empty partitions exists for the dynamic partition columns, (e.g., country='CA' exists under some ds root partition), it will be overwritten if the dynamic partition insert saw the same value (say 'CA') in the input data. This is in line with the 'insert overwrite' semantics. However, if the partition value 'CA' does not appear in the input data, the existing partition will not be overwritten.如果有对应的分区数据，则会被覆盖。没有则不会被覆盖。
• Since a Hive partition corresponds to a directory in HDFS, the partition value has to conform to the HDFS path format (URI in Java). Any character having a special meaning in URI (e.g., '%', ':', '/', '#') will be escaped with '%' followed by 2 bytes of its ASCII value.
• If the input column is a type different than STRING, its value will be first converted to STRING to be used to construct the HDFS path.
• If the input column value is NULL or empty string, the row will be put into a special partition, whose name is controlled by the hive parameter hive.exec.default.partition.name. The default value is HIVE_DEFAULT_PARTITION{}. Basically this partition will contain all "bad" rows whose value are not valid partition names. The caveat of this approach is that the bad value will be lost and is replaced by HIVE_DEFAULT_PARTITION{} if you select them Hive. JIRA HIVE-1309 is a solution to let user specify "bad file" to retain the input partition column values as well. (空值使用默认的partition名字)
• Dynamic partition insert could potentially resource hog in that it could generate a large number of partitions in a short time. To get yourself buckled, we define three parameters:（因为动态分区插入过程中，会在短时间内产生大量的partition,可以通过一下参数进行控制：）
• hive.exec.max.dynamic.partitions.pernode (default value being 100) is the maximum dynamic partitions that can be created by each mapper or reducer. If one mapper or reducer created more than that the threshold, a fatal error will be raised from the mapper/reducer (through counter) and the whole job will be killed. （单个Mapper或者Redcuer最大的动态Partition数量）
• hive.exec.max.dynamic.partitions (default value being 1000) is the total number of dynamic partitions could be created by one DML. If each mapper/reducer did not exceed the limit but the total number of dynamic partitions does, then an exception is raised at the end of the job before the intermediate data are moved to the final destination.（在最后一个JOB，执行中间数据迁移到最终目的地的时候，如果一个DML语句中所有的Mapper/Reducer产生的动态分区之后大于设定值，则会抛出异常。）
• hive.exec.max.created.files (default value being 100000) is the maximum total number of files created by all mappers and reducers. This is implemented by updating a Hadoop counter by each mapper/reducer whenever a new file is created. If the total number is exceeding hive.exec.max.created.files, a fatal error will be thrown and the job will be killed.（该参数定义为：所有的Mapper和Reducer产生的文件数量不能超过此参数设定值，否则当前JOB会被KILL）

• Another situation we want to protect against dynamic partition insert is that the user may accidentally specify all partitions to be dynamic partitions without specifying one static partition, while the original intention is to just overwrite the sub-partitions of one root partition. We define another parameter hive.exec.dynamic.partition.mode=strict to prevent the all-dynamic partition case. In the strict mode, you have to specify at least one static partition. The default mode is strict. In addition, we have a parameter hive.exec.dynamic.partition=true/false to control whether to allow dynamic partition at all. The default value is false.（为了防止用户不小心将所有的分区设置成动态分区，没有静态分区，通过hive.exec.dynamic.partition.mode=strict参数来解决此问题，当设置成strict时候，用户至少指定一个静态分区。）
• In Hive 0.6, dynamic partition insert does not work with hive.merge.mapfiles=true or hive.merge.mapredfiles=true, so it internally turns off the merge parameters. Merging files in dynamic partition inserts are supported in Hive 0.7 (see JIRA HIVE-1307 for details).

Troubleshooting and best practices: 问题解决及最佳实践

• As stated above, there are too many dynamic partitions created by a particular mapper/reducer, a fatal error could be raised and the job will be killed. The error message looks something like:

  hive> set hive.exec.dynamic.partition.mode=nonstrict;     hive> FROM page_view_stg pvs           INSERT OVERWRITE TABLE page_view PARTITION(dt, country)                  SELECT pvs.viewTime, pvs.userid, pvs.page_url, pvs.referrer_url, null, null, pvs.ip,                         from_unixtimestamp(pvs.viewTime, 'yyyy-MM-dd') ds, pvs.country; ... 2010-05-07 11:10:19,816 Stage-1 map = 0%,  reduce = 0% [Fatal Error] Operator FS_28 (id=41): fatal error. Killing the job. Ended Job = job_201005052204_28178 with errors ...

  The problem of this that one mapper will take a random set of rows and it is very likely that the number of distinct (dt, country) pairs will exceed the limit of hive.exec.max.dynamic.partitions.pernode. One way around it is to group the rows by the dynamic partition columns in the mapper and distribute them to the reducers where the dynamic partitions will be created. In this case the number of distinct dynamic partitions will be significantly reduced. The above example query could be rewritten to:

  hive> set hive.exec.dynamic.partition.mode=nonstrict; hive> FROM page_view_stg pvs       INSERT OVERWRITE TABLE page_view PARTITION(dt, country)              SELECT pvs.viewTime, pvs.userid, pvs.page_url, pvs.referrer_url, null, null, pvs.ip,                     from_unixtimestamp(pvs.viewTime, 'yyyy-MM-dd') ds, pvs.country              DISTRIBUTE BY ds, country;

  This query will generate a MapReduce job rather than Map-only job. The SELECT-clause will be converted to a plan to the mappers and the output will be distributed to the reducers based on the value of (ds, country) pairs. The INSERT-clause will be converted to the plan in the reducer which writes to the dynamic partitions.

Additional documentation:

• Design Document for Dynamic Partitions
  • Original design doc
  • HIVE-936
• Hive DML: Dynamic Partition Inserts
• HCatalog Dynamic Partitioning
  • Usage with Pig
  • Usage from MapReduce

Inserting into Local Files　　添加本地文件

In certain situations you would want to write the output into a local file so that you could load it into an excel spreadsheet. This can be accomplished with the following command:

Sampling 抽样插入

The sampling clause allows the users to write queries for samples of the data instead of the whole table. Currently the sampling is done on the columns that are specified in the CLUSTERED BY clause of the CREATE TABLE statement. In the following example we choose 3rd bucket out of the 32 buckets of the pv_gender_sum table:

In general the TABLESAMPLE syntax looks like:

y has to be a multiple or divisor of the number of buckets in that table as specified at the table creation time. The buckets chosen are determined if bucket_number module y is equal to x. So in the above example the following tablesample clause

would pick out the 3rd and 19th buckets. The buckets are numbered starting from 0.

On the other hand the tablesample clause

would pick out half of the 3rd bucket.

Union All

The language also supports union all, e.g. if we suppose there are two different tables that track which user has published a video and which user has published a comment, the following query joins the results of a union all with the user table to create a single annotated stream for all the video publishing and comment publishing events:

Array Operations 数组操作

Array columns in tables can be as follows:

Assuming that pv.friends is of the type ARRAY<INT> (i.e. it is an array of integers), the user can get a specific element in the array by its index as shown in the following command:

The select expression gets the third item in the pv.friends array.

The user can also get the length of the array using the size function as shown below:

Map (Associative Arrays) Operations

Maps provide collections similar to associative arrays. Such structures can only be created programmatically currently. We will be extending this soon. For the purpose of the current example assume that pv.properties is of the type map<String, String> i.e. it is an associative array from strings to string. Accordingly, the following query:

can be used to select the 'page_type' property from the page_views table.

Similar to arrays, the size function can also be used to get the number of elements in a map as shown in the following query:

Custom Map/Reduce Scripts  [??]

Users can also plug in their own custom mappers and reducers in the data stream by using features natively supported in the Hive language. e.g. in order to run a custom mapper script - map_script - and a custom reducer script - reduce_script - the user can issue the following command which uses the TRANSFORM clause to embed the mapper and the reducer scripts.

Note that columns will be transformed to string and delimited by TAB before feeding to the user script, and the standard output of the user script will be treated as TAB-separated string columns. User scripts can output debug information to standard error which will be shown on the task detail page on hadoop.

Sample map script (weekday_mapper.py )

Of course, both MAP and REDUCE are "syntactic sugar" for the more general select transform. The inner query could also have been written as such:

Schema-less map/reduce: If there is no "AS" clause after "USING map_script", Hive assumes the output of the script contains 2 parts: key which is before the first tab, and value which is the rest after the first tab. Note that this is different from specifying "AS key, value" because in that case value will only contains the portion between the first tab and the second tab if there are multiple tabs.

In this way, we allow users to migrate old map/reduce scripts without knowing the schema of the map output. User still needs to know the reduce output schema because that has to match what is in the table that we are inserting to.

Distribute By and Sort By: Instead of specifying "cluster by", the user can specify "distribute by" and "sort by", so the partition columns and sort columns can be different. The usual case is that the partition columns are a prefix of sort columns, but that is not required.

Co-Groups

Amongst the user community using map/reduce, cogroup is a fairly common operation wherein the data from multiple tables are sent to a custom reducer such that the rows are grouped by the values of certain columns on the tables. With the UNION ALL operator and the CLUSTER BY specification, this can be achieved in the Hive query language in the following way. Suppose we wanted to cogroup the rows from the actions_video and action_comments table on the uid column and send them to the 'reduce_script' custom reducer, the following syntax can be used by the user: