Using Impala with the Amazon S3 Filesystem

Impala SQL statements work with data on S3 as follows:

• The CREATE TABLE Statement or ALTER TABLE Statement statements can specify that a table resides on the S3 filesystem by encoding an s3a:// prefix for the LOCATION property. ALTER TABLE can also set the LOCATION property for an individual partition, so that some data in a table resides on S3 and other data in the same table resides on HDFS.

• Once a table or partition is designated as residing on S3, the SELECT Statement statement transparently accesses the data files from the appropriate storage layer.

• If the S3 table is an internal table, the DROP TABLE Statement statement removes the corresponding data files from S3 when the table is dropped.

• The TRUNCATE TABLE Statement (CDH 5.5 or higher only) statement always removes the corresponding data files from S3 when the table is truncated.

• The LOAD DATA Statement can move data files residing in HDFS into an S3 table.

• The INSERT Statement statement, or the CREATE TABLE AS SELECT form of the CREATE TABLE statement, can copy data from an HDFS table or another S3 table into an S3 table. The S3_SKIP_INSERT_STAGING Query Option (CDH 5.8 or higher only) query option chooses whether or not to use a fast code path for these write operations to S3, with the tradeoff of potential inconsistency in the case of a failure during the statement.

For usage information about Impala SQL statements with S3 tables, see Creating Impala Databases, Tables, and Partitions for Data Stored on S3 and Using Impala DML Statements for S3 Data.

To allow Impala to access data in S3, specify values for the following configuration settings in your core-site.xml file:

<property>
<name>fs.s3a.access.key</name>
<value>your_access_key</value>
</property>
<property>
<name>fs.s3a.secret.key</name>
<value>your_secret_key</value>
</property>

As of CDH 5.8, these settings do not have corresponding controls in the Cloudera Manager user interface. Specify them in the HDFS Client Advanced Configuration Snippet (Safety Valve) for core-site.xml field. After specifying the credentials, restart both the Impala and Hive services. (Restarting Hive is required because Impala queries, CREATE TABLE statements, and so on go through the Hive metastore.)

If your ETL pipeline involves moving data into S3 and then querying through Impala, you can either use Impala DML statements to create, move, or copy the data, or use the same data loading techniques as you would for non-Impala data.

Impala reads data for a table or partition from S3 based on the LOCATION attribute for the table or partition. Specify the S3 details in the LOCATION clause of a CREATE TABLE or ALTER TABLE statement. The notation for the LOCATION clause is s3a://bucket_name/path/to/file. The filesystem prefix is always s3a:// because Impala does not support the s3:// or s3n:// prefixes.

For a partitioned table, either specify a separate LOCATION clause for each new partition, or specify a base LOCATION for the table and set up a directory structure in S3 to mirror the way Impala partitioned tables are structured in HDFS. Although, strictly speaking, S3 filenames do not have directory paths, Impala treats S3 filenames with / characters the same as HDFS pathnames that include directories.

You point a nonpartitioned table or an individual partition at S3 by specifying a single directory path in S3, which could be any arbitrary directory. To replicate the structure of an entire Impala partitioned table or database in S3 requires more care, with directories and subdirectories nested and named to match the equivalent directory tree in HDFS. Consider setting up an empty staging area if necessary in HDFS, and recording the complete directory structure so that you can replicate it in S3.

For convenience when working with multiple tables with data files stored in S3, you can create a database with a LOCATION attribute pointing to an S3 path. Specify a URL of the form s3a://bucket/root/path/for/database for the LOCATION attribute of the database. Any tables created inside that database automatically create directories underneath the one specified by the database LOCATION attribute.

For example, the following session creates a partitioned table where only a single partition resides on S3. The partitions for years 2013 and 2014 are located on HDFS. The partition for year 2015 includes a LOCATION attribute with an s3a:// URL, and so refers to data residing on S3, under a specific path underneath the bucket impala-demo.

[localhost:21000] > create database db_on_hdfs;
[localhost:21000] > use db_on_hdfs;
[localhost:21000] > create table mostly_on_hdfs (x int) partitioned by (year int);
[localhost:21000] > alter table mostly_on_hdfs add partition (year=2013);
[localhost:21000] > alter table mostly_on_hdfs add partition (year=2014);
[localhost:21000] > alter table mostly_on_hdfs add partition (year=2015)
                  >   location 's3a://impala-demo/dir1/dir2/dir3/t1';

The following session creates a database and two partitioned tables residing entirely on S3, one partitioned by a single column and the other partitioned by multiple columns. Because a LOCATION attribute with an s3a:// URL is specified for the database, the tables inside that database are automatically created on S3 underneath the database directory. To see the names of the associated subdirectories, including the partition key values, we use an S3 client tool to examine how the directory structure is organized on S3. For example, Impala partition directories such as month=1 do not include leading zeroes, which sometimes appear in partition directories created through Hive.

[localhost:21000] > create database db_on_s3 location 's3a://impala-demo/dir1/dir2/dir3';
[localhost:21000] > use db_on_s3;

[localhost:21000] > create table partitioned_on_s3 (x int) partitioned by (year int);
[localhost:21000] > alter table partitioned_on_s3 add partition (year=2013);
[localhost:21000] > alter table partitioned_on_s3 add partition (year=2014);
[localhost:21000] > alter table partitioned_on_s3 add partition (year=2015);

[localhost:21000] > !aws s3 ls s3://impala-demo/dir1/dir2/dir3 --recursive;
2015-03-17 13:56:34          0 dir1/dir2/dir3/
2015-03-17 16:43:28          0 dir1/dir2/dir3/partitioned_on_s3/
2015-03-17 16:43:49          0 dir1/dir2/dir3/partitioned_on_s3/year=2013/
2015-03-17 16:43:53          0 dir1/dir2/dir3/partitioned_on_s3/year=2014/
2015-03-17 16:43:58          0 dir1/dir2/dir3/partitioned_on_s3/year=2015/

[localhost:21000] > create table partitioned_multiple_keys (x int)
                  >   partitioned by (year smallint, month tinyint, day tinyint);
[localhost:21000] > alter table partitioned_multiple_keys
                  >   add partition (year=2015,month=1,day=1);
[localhost:21000] > alter table partitioned_multiple_keys
                  >   add partition (year=2015,month=1,day=31);
[localhost:21000] > alter table partitioned_multiple_keys
                  >   add partition (year=2015,month=2,day=28);

[localhost:21000] > !aws s3 ls s3://impala-demo/dir1/dir2/dir3 --recursive;
2015-03-17 13:56:34          0 dir1/dir2/dir3/
2015-03-17 16:47:13          0 dir1/dir2/dir3/partitioned_multiple_keys/
2015-03-17 16:47:44          0 dir1/dir2/dir3/partitioned_multiple_keys/year=2015/month=1/day=1/
2015-03-17 16:47:50          0 dir1/dir2/dir3/partitioned_multiple_keys/year=2015/month=1/day=31/
2015-03-17 16:47:57          0 dir1/dir2/dir3/partitioned_multiple_keys/year=2015/month=2/day=28/
2015-03-17 16:43:28          0 dir1/dir2/dir3/partitioned_on_s3/
2015-03-17 16:43:49          0 dir1/dir2/dir3/partitioned_on_s3/year=2013/
2015-03-17 16:43:53          0 dir1/dir2/dir3/partitioned_on_s3/year=2014/
2015-03-17 16:43:58          0 dir1/dir2/dir3/partitioned_on_s3/year=2015/

The CREATE DATABASE and CREATE TABLE statements create the associated directory paths if they do not already exist. You can specify multiple levels of directories, and the CREATE statement creates all appropriate levels, similar to using mkdir -p.

Use the standard S3 file upload methods to actually put the data files into the right locations. You can also put the directory paths and data files in place before creating the associated Impala databases or tables, and Impala automatically uses the data from the appropriate location after the associated databases and tables are created.

You can switch whether an existing table or partition points to data in HDFS or S3. For example, if you have an Impala table or partition pointing to data files in HDFS or S3, and you later transfer those data files to the other filesystem, use an ALTER TABLE statement to adjust the LOCATION attribute of the corresponding table or partition to reflect that change. Because Impala does not have an ALTER DATABASE statement, this location-switching technique is not practical for entire databases that have a custom LOCATION attribute.

Just as with tables located on HDFS storage, you can designate S3-based tables as either internal (managed by Impala) or external, by using the syntax CREATE TABLE or CREATE EXTERNAL TABLE respectively. When you drop an internal table, the files associated with the table are removed, even if they are on S3 storage. When you drop an external table, the files associated with the table are left alone, and are still available for access by other tools or components. See Overview of Impala Tables for details.

If the data on S3 is intended to be long-lived and accessed by other tools in addition to Impala, create any associated S3 tables with the CREATE EXTERNAL TABLE syntax, so that the files are not deleted from S3 when the table is dropped.

If the data on S3 is only needed for querying by Impala and can be safely discarded once the Impala workflow is complete, create the associated S3 tables using the CREATE TABLE syntax, so that dropping the table also deletes the corresponding data files on S3.

For example, this session creates a table in S3 with the same column layout as a table in HDFS, then examines the S3 table and queries some data from it. The table in S3 works the same as a table in HDFS as far as the expected file format of the data, table and column statistics, and other table properties. The only indication that it is not an HDFS table is the s3a:// URL in the LOCATION property. Many data files can reside in the S3 directory, and their combined contents form the table data. Because the data in this example is uploaded after the table is created, a REFRESH statement prompts Impala to update its cached information about the data files.

[localhost:21000] > create table usa_cities_s3 like usa_cities location 's3a://impala-demo/usa_cities';
[localhost:21000] > desc usa_cities_s3;
+-------+----------+---------+
| name  | type     | comment |
+-------+----------+---------+
| id    | smallint |         |
| city  | string   |         |
| state | string   |         |
+-------+----------+---------+

-- Now from a web browser, upload the same data file(s) to S3 as in the HDFS table,
-- under the relevant bucket and path. If you already have the data in S3, you would
-- point the table LOCATION at an existing path.

[localhost:21000] > refresh usa_cities_s3;
[localhost:21000] > select count(*) from usa_cities_s3;
+----------+
| count(*) |
+----------+
| 289      |
+----------+
[localhost:21000] > select distinct state from sample_data_s3 limit 5;
+----------------------+
| state                |
+----------------------+
| Louisiana            |
| Minnesota            |
| Georgia              |
| Alaska               |
| Ohio                 |
+----------------------+
[localhost:21000] > desc formatted usa_cities_s3;
+------------------------------+------------------------------+---------+
| name                         | type                         | comment |
+------------------------------+------------------------------+---------+
| # col_name                   | data_type                    | comment |
|                              | NULL                         | NULL    |
| id                           | smallint                     | NULL    |
| city                         | string                       | NULL    |
| state                        | string                       | NULL    |
|                              | NULL                         | NULL    |
| # Detailed Table Information | NULL                         | NULL    |
| Database:                    | s3_testing                   | NULL    |
| Owner:                       | jrussell                     | NULL    |
| CreateTime:                  | Mon Mar 16 11:36:25 PDT 2015 | NULL    |
| LastAccessTime:              | UNKNOWN                      | NULL    |
| Protect Mode:                | None                         | NULL    |
| Retention:                   | 0                            | NULL    |
| Location:                    | s3a://impala-demo/usa_cities | NULL    |
| Table Type:                  | MANAGED_TABLE                | NULL    |
...
+------------------------------+------------------------------+---------+

In this case, we have already uploaded a Parquet file with a million rows of data to the sample_data directory underneath the impala-demo bucket on S3. This session creates a table with matching column settings pointing to the corresponding location in S3, then queries the table. Because the data is already in place on S3 when the table is created, no REFRESH statement is required.

[localhost:21000] > create table sample_data_s3
                  > (id int, id bigint, val int, zerofill string,
                  > name string, assertion boolean, city string, state string)
                  > stored as parquet location 's3a://impala-demo/sample_data';
[localhost:21000] > select count(*) from sample_data_s3;;
+----------+
| count(*) |
+----------+
| 1000000  |
+----------+
[localhost:21000] > select count(*) howmany, assertion from sample_data_s3 group by assertion;
+---------+-----------+
| howmany | assertion |
+---------+-----------+
| 667149  | true      |
| 332851  | false     |
+---------+-----------+

Once the appropriate LOCATION attributes are set up at the table or partition level, you query data stored in S3 exactly the same as data stored on HDFS or in HBase:

• Queries against S3 data support all the same file formats as for HDFS data.
• Tables can be unpartitioned or partitioned. For partitioned tables, either manually construct paths in S3 corresponding to the HDFS directories representing partition key values, or use ALTER TABLE ... ADD PARTITION to set up the appropriate paths in S3.
• HDFS and HBase tables can be joined to S3 tables, or S3 tables can be joined with each other.
• Authorization using the Sentry framework to control access to databases, tables, or columns works the same whether the data is in HDFS or in S3.
• The catalogd daemon caches metadata for both HDFS and S3 tables. Use REFRESH and INVALIDATE METADATA for S3 tables in the same situations where you would issue those statements for HDFS tables.
• Queries against S3 tables are subject to the same kinds of admission control and resource management as HDFS tables.
• Metadata about S3 tables is stored in the same metastore database as for HDFS tables.
• You can set up views referring to S3 tables, the same as for HDFS tables.
• The COMPUTE STATS, SHOW TABLE STATS, and SHOW COLUMN STATS statements work for S3 tables also.

Impala requires that the default filesystem for the cluster be HDFS. You cannot use S3 as the only filesystem in the cluster.

Prior to CDH 5.8 / Impala 2.6, Impala could not perform DML operations (INSERT, LOAD DATA, or CREATE TABLE AS SELECT) where the destination is a table or partition located on an S3 filesystem. This restriction is lifted in CDH 5.8 / Impala 2.6 and higher.

Impala does not support the old s3:// block-based and s3n:// filesystem schemes, only s3a://.

Although S3 is often used to store JSON-formatted data, the current Impala support for S3 does not include directly querying JSON data. For Impala queries, use data files in one of the file formats listed in How Impala Works with Hadoop File Formats. If you have data in JSON format, you can prepare a flattened version of that data for querying by Impala as part of your ETL cycle.

You cannot use the ALTER TABLE ... SET CACHED statement for tables or partitions that are located in S3.

The following guidelines represent best practices derived from testing and field experience with Impala on S3:

• Any reference to an S3 location must be fully qualified. (This rule applies when S3 is not designated as the default filesystem.)

• Set the safety valve fs.s3a.connection.maximum to 1500 for impalad.

• Set safety valve fs.s3a.block.size to 134217728 (128 MB in bytes) if most Parquet files queried by Impala were written by Hive or ParquetMR jobs. Set the block size to 268435456 (256 MB in bytes) if most Parquet files queried by Impala were written by Impala.

• DROP TABLE .. PURGE is much faster than the default DROP TABLE. The same applies to ALTER TABLE ... DROP PARTITION PURGE versus the default DROP PARTITION operation. However, due to the eventually consistent nature of S3, the files for that table or partition could remain for some unbounded time when using PURGE. The default DROP TABLE/PARTITION is slow because Impala copies the files to the HDFS trash folder, and Impala waits until all the data is moved. DROP TABLE/PARTITION .. PURGE is a fast delete operation, and the Impala statement finishes quickly even though the change might not have propagated fully throughout S3.

• INSERT statements are faster than INSERT OVERWRITE for S3. The query option S3_SKIP_INSERT_STAGING, which is set to true by default, skips the staging step for regular INSERT (but not INSERT OVERWRITE). This makes the operation much faster, but consistency is not guaranteed: if a node fails during execution, the table could end up with inconsistent data. Set this option to false if stronger consistency is required, however this setting will make the INSERT operations slower.

• Too many files in a table can make metadata loading and updating slow on S3. If too many requests are made to S3, S3 has a back-off mechanism and responds slower than usual. You might have many small files because of:

  • Too many partitions due to over-granular partitioning. Prefer partitions with many megabytes of data, so that even a query against a single partition can be parallelized effectively.

  • Many small INSERT queries. Prefer bulk INSERTs so that more data is written to fewer files.