Enabling Sentry Authorization for Impala

Authorization determines which users are allowed to access which resources, and what operations they are allowed to perform. In Impala 1.1 and higher, you use the Sentry open source project for authorization. Sentry adds a fine-grained authorization framework for Hadoop. By default (when authorization is not enabled), Impala does all read and write operations with the privileges of the impala user, which is suitable for a development/test environment but not for a secure production environment. When authorization is enabled, Impala uses the OS user ID of the user who runs impala-shell or other client program, and associates various privileges with each user.

The Sentry Privilege Model

Privileges can be granted on different objects in the schema. Any privilege that can be granted is associated with a level in the object hierarchy. If a privilege is granted on a container object in the hierarchy, the child object automatically inherits it. This is the same privilege model as Hive and other database systems such as MySQL.

The object hierarchy for Impala covers Server, URI, Database, Table, and Column. (The Table privileges apply to views as well; anywhere you specify a table name, you can specify a view name instead.) Column-level authorization is available in CDH 5.5 / Impala 2.3 and higher, as described in Column-level Authorization. Previously, you constructed views to query specific columns and assigned privileges based on the views rather than the base tables.

A restricted set of privileges determines what you can do with each object:

SELECT privilege
Lets you read data from a table or view, for example with the SELECT statement, the INSERT...SELECT syntax, or CREATE TABLE...LIKE. Also required to issue the DESCRIBE statement or the EXPLAIN statement for a query against a particular table. Only objects for which a user has this privilege are shown in the output for SHOW DATABASES and SHOW TABLES statements. The REFRESH statement and INVALIDATE METADATA statements only access metadata for tables for which the user has this privilege.
INSERT privilege
Lets you write data to a table. Applies to the INSERT and LOAD DATA statements.
ALL privilege
Lets you create or modify the object. Required to run DDL statements such as CREATE TABLE, ALTER TABLE, or DROP TABLE for a table, CREATE DATABASE or DROP DATABASE for a database, or CREATE VIEW, ALTER VIEW, or DROP VIEW for a view. Also required for the URI of the "location" parameter for the CREATE EXTERNAL TABLE and LOAD DATA statements.

Privileges can be specified for a table or view before that object actually exists. If you do not have sufficient privilege to perform an operation, the error message does not disclose if the object exists or not.

Originally, privileges were encoded in a policy file, stored in HDFS. This mode of operation is still an option, but the emphasis of privilege management is moving towards being SQL-based. Although currently Impala does not have GRANT or REVOKE statements, Impala can make use of privileges assigned through GRANT and REVOKE statements done through Hive. The mode of operation with GRANT and REVOKE statements instead of the policy file requires that a special Sentry service be enabled; this service stores, retrieves, and manipulates privilege information stored inside the metastore database.

Starting the impalad Daemon with Sentry Authorization Enabled

To run the impalad daemon with authorization enabled, you add one or more options to the IMPALA_SERVER_ARGS declaration in the /etc/default/impala configuration file:

  • The -server_name option turns on Sentry authorization for Impala. The authorization rules refer to a symbolic server name, and you specify the name to use as the argument to the -server_name option.
  • If you specify just -server_name, Impala uses the Sentry service for authorization, relying on the results of GRANT and REVOKE statements issued through Hive. (This mode of operation is available in Impala 1.4.0 and higher.) Prior to Impala 1.4.0, or if you want to continue storing privilege rules in the policy file, also specify the -authorization_policy_file option as in the following item.
  • Specifying the -authorization_policy_file option in addition to -server_name makes Impala read privilege information from a policy file, rather than from the metastore database. The argument to the -authorization_policy_file option specifies the HDFS path to the policy file that defines the privileges on different schema objects.

For example, you might adapt your /etc/default/impala configuration to contain lines like the following. To use the Sentry service rather than the policy file:

IMPALA_SERVER_ARGS=" \
-server_name=server1 \
...

Or to use the policy file, as in releases prior to Impala 1.4:

IMPALA_SERVER_ARGS=" \
-authorization_policy_file=/user/hive/warehouse/auth-policy.ini \
-server_name=server1 \
...

The preceding examples set up a symbolic name of server1 to refer to the current instance of Impala. This symbolic name is used in the following ways:

  • In an environment managed by Cloudera Manager, the server name is specified through Impala (Service-Wide) > Category > Advanced > Sentry Service and Hive > Service-Wide > Advanced > Sentry Service. The values must be the same for both, so that Impala and Hive can share the privilege rules. Restart the Impala and Hive services after setting or changing this value.

  • In an environment not managed by Cloudera Manager, you specify this value for the sentry.hive.server property in the sentry-site.xml configuration file for Hive, as well as in the -server_name option for impalad.

    If the impalad daemon is not already running, start it as described in Starting Impala. If it is already running, restart it with the command sudo /etc/init.d/impala-server restart. Run the appropriate commands on all the nodes where impalad normally runs.

  • If you use the mode of operation using the policy file, the rules in the [roles] section of the policy file refer to this same server1 name. For example, the following rule sets up a role report_generator that lets users with that role query any table in a database named reporting_db on a node where the impalad daemon was started up with the -server_name=server1 option:

    [roles]
    report_generator = server=server1->db=reporting_db->table=*->action=SELECT
    

When impalad is started with one or both of the -server_name=server1 and -authorization_policy_file options, Impala authorization is enabled. If Impala detects any errors or inconsistencies in the authorization settings or the policy file, the daemon refuses to start.

Using Impala with the Sentry Service (CDH 5.1 or higher only)

When you use the Sentry service rather than the policy file, you set up privileges through GRANT and REVOKE statement in either Impala or Hive, then both components use those same privileges automatically. (Impala added the GRANT and REVOKE statements in Impala 2.0.0 / CDH 5.2.0.)

Hive already had GRANT and REVOKE statements prior to CDH 5.1, but those statements were not production-ready. CDH 5.1 is the first release where those statements use the Sentry framework and are considered GA level. If you used the Hive GRANT and REVOKE statements prior to CDH 5.1, you must set up these privileges with the CDH 5.1 versions of GRANT and REVOKE to take advantage of Sentry authorization.

For information about using the updated Hive GRANT and REVOKE statements, see Sentry service topic in the CDH 5 Security Guide.

Using Impala with the Sentry Policy File

The policy file is a file that you put in a designated location in HDFS, and is read during the startup of the impalad daemon when you specify both the -server_name and -authorization_policy_file startup options. It controls which objects (databases, tables, and HDFS directory paths) can be accessed by the user who connects to impalad, and what operations that user can perform on the objects.

The location of the policy file is listed in the auth-site.xml configuration file. To minimize overhead, the security information from this file is cached by each impalad daemon and refreshed automatically, with a default interval of 5 minutes. After making a substantial change to security policies, restart all Impala daemons to pick up the changes immediately.

Policy File Location and Format

The policy file uses the familiar .ini format, divided into the major sections [groups] and [roles]. There is also an optional [databases] section, which allows you to specify a specific policy file for a particular database, as explained in Using Multiple Policy Files for Different Databases. Another optional section, [users], allows you to override the OS-level mapping of users to groups; that is an advanced technique primarily for testing and debugging, and is beyond the scope of this document.

In the [groups] section, you define various categories of users and select which roles are associated with each category. The group and usernames correspond to Linux groups and users on the server where the impalad daemon runs.

The group and usernames in the [groups] section correspond to Linux groups and users on the server where the impalad daemon runs. When you access Impala through the impalad interpreter, for purposes of authorization, the user is the logged-in Linux user and the groups are the Linux groups that user is a member of. When you access Impala through the ODBC or JDBC interfaces, the user and password specified through the connection string are used as login credentials for the Linux server, and authorization is based on that username and the associated Linux group membership.

In the [roles] section, you a set of roles. For each role, you specify precisely the set of privileges is available. That is, which objects users with that role can access, and what operations they can perform on those objects. This is the lowest-level category of security information; the other sections in the policy file map the privileges to higher-level divisions of groups and users. In the [groups] section, you specify which roles are associated with which groups. The group and usernames correspond to Linux groups and users on the server where the impalad daemon runs. The privileges are specified using patterns like:
server=server_name->db=database_name->table=table_name->action=SELECT
server=server_name->db=database_name->table=table_name->action=CREATE
server=server_name->db=database_name->table=table_name->action=ALL
For the server_name value, substitute the same symbolic name you specify with the impalad -server_name option. You can use * wildcard characters at each level of the privilege specification to allow access to all such objects. For example:
server=impala-host.example.com->db=default->table=t1->action=SELECT
server=impala-host.example.com->db=*->table=*->action=CREATE
server=impala-host.example.com->db=*->table=audit_log->action=SELECT
server=impala-host.example.com->db=default->table=t1->action=*

When authorization is enabled, Impala uses the policy file as a whitelist, representing every privilege available to any user on any object. That is, only operations specified for the appropriate combination of object, role, group, and user are allowed; all other operations are not allowed. If a group or role is defined multiple times in the policy file, the last definition takes precedence.

To understand the notion of whitelisting, set up a minimal policy file that does not provide any privileges for any object. When you connect to an Impala node where this policy file is in effect, you get no results for SHOW DATABASES, and an error when you issue any SHOW TABLES, USE database_name, DESCRIBE table_name, SELECT, and or other statements that expect to access databases or tables, even if the corresponding databases and tables exist.

The contents of the policy file are cached, to avoid a performance penalty for each query. The policy file is re-checked by each impalad node every 5 minutes. When you make a non-time-sensitive change such as adding new privileges or new users, you can let the change take effect automatically a few minutes later. If you remove or reduce privileges, and want the change to take effect immediately, restart the impalad daemon on all nodes, again specifying the -server_name and -authorization_policy_file options so that the rules from the updated policy file are applied.

Examples of Policy File Rules for Security Scenarios

The following examples show rules that might go in the policy file to deal with various authorization-related scenarios. For illustration purposes, this section shows several very small policy files with only a few rules each. In your environment, typically you would define many roles to cover all the scenarios involving your own databases, tables, and applications, and a smaller number of groups, whose members are given the privileges from one or more roles.

A User with No Privileges

If a user has no privileges at all, that user cannot access any schema objects in the system. The error messages do not disclose the names or existence of objects that the user is not authorized to read.

This is the experience you want a user to have if they somehow log into a system where they are not an authorized Impala user. In a real deployment with a filled-in policy file, a user might have no privileges because they are not a member of any of the relevant groups mentioned in the policy file.

Examples of Privileges for Administrative Users

When an administrative user has broad access to tables or databases, the associated rules in the [roles] section typically use wildcards and/or inheritance. For example, in the following sample policy file, db=* refers to all databases and db=*->table=* refers to all tables in all databases.

Omitting the rightmost portion of a rule means that the privileges apply to all the objects that could be specified there. For example, in the following sample policy file, the all_databases role has all privileges for all tables in all databases, while the one_database role has all privileges for all tables in one specific database. The all_databases role does not grant privileges on URIs, so a group with that role could not issue a CREATE TABLE statement with a LOCATION clause. The entire_server role has all privileges on both databases and URIs within the server.

[groups]
supergroup = all_databases

[roles]
read_all_tables = server=server1->db=*->table=*->action=SELECT
all_tables = server=server1->db=*->table=*
all_databases = server=server1->db=*
one_database = server=server1->db=test_db
entire_server = server=server1

A User with Privileges for Specific Databases and Tables

If a user has privileges for specific tables in specific databases, the user can access those things but nothing else. They can see the tables and their parent databases in the output of SHOW TABLES and SHOW DATABASES, USE the appropriate databases, and perform the relevant actions (SELECT and/or INSERT) based on the table privileges. To actually create a table requires the ALL privilege at the database level, so you might define separate roles for the user that sets up a schema and other users or applications that perform day-to-day operations on the tables.

The following sample policy file shows some of the syntax that is appropriate as the policy file grows, such as the # comment syntax, \ continuation syntax, and comma separation for roles assigned to groups or privileges assigned to roles.

[groups]
cloudera = training_sysadmin, instructor
visitor = student

[roles]
training_sysadmin = server=server1->db=training, \
server=server1->db=instructor_private, \
server=server1->db=lesson_development
instructor = server=server1->db=training->table=*->action=*, \
server=server1->db=instructor_private->table=*->action=*, \
server=server1->db=lesson_development->table=lesson*
# This particular course is all about queries, so the students can SELECT but not INSERT or CREATE/DROP.
student = server=server1->db=training->table=lesson_*->action=SELECT

Privileges for Working with External Data Files

When data is being inserted through the LOAD DATA statement, or is referenced from an HDFS location outside the normal Impala database directories, the user also needs appropriate permissions on the URIs corresponding to those HDFS locations.

In this sample policy file:

  • The external_table role lets us insert into and query the Impala table, external_table.sample.
  • The staging_dir role lets us specify the HDFS path /user/cloudera/external_data with the LOAD DATA statement. Remember, when Impala queries or loads data files, it operates on all the files in that directory, not just a single file, so any Impala LOCATION parameters refer to a directory rather than an individual file.
  • We included the IP address and port of the Hadoop name node in the HDFS URI of the staging_dir rule. We found those details in /etc/hadoop/conf/core-site.xml, under the fs.default.name element. That is what we use in any roles that specify URIs (that is, the locations of directories in HDFS).
  • We start this example after the table external_table.sample is already created. In the policy file for the example, we have already taken away the external_table_admin role from the cloudera group, and replaced it with the lesser-privileged external_table role.
  • We assign privileges to a subdirectory underneath /user/cloudera in HDFS, because such privileges also apply to any subdirectories underneath. If we had assigned privileges to the parent directory /user/cloudera, it would be too likely to mess up other files by specifying a wrong location by mistake.
  • The cloudera under the [groups] section refers to the cloudera group. (In the demo VM used for this example, there is a cloudera user that is a member of a cloudera group.)

Policy file:

[groups]
cloudera = external_table, staging_dir

[roles]
external_table_admin = server=server1->db=external_table
external_table = server=server1->db=external_table->table=sample->action=*
staging_dir = server=server1->uri=hdfs://127.0.0.1:8020/user/cloudera/external_data->action=*

impala-shell session:

[localhost:21000] > use external_table;
Query: use external_table
[localhost:21000] > show tables;
Query: show tables
Query finished, fetching results ...
+--------+
| name   |
+--------+
| sample |
+--------+
Returned 1 row(s) in 0.02s

[localhost:21000] > select * from sample;
Query: select * from sample
Query finished, fetching results ...
+-----+
| x   |
+-----+
| 1   |
| 5   |
| 150 |
+-----+
Returned 3 row(s) in 1.04s

[localhost:21000] > load data inpath '/user/cloudera/external_data' into table sample;
Query: load data inpath '/user/cloudera/external_data' into table sample
Query finished, fetching results ...
+----------------------------------------------------------+
| summary                                                  |
+----------------------------------------------------------+
| Loaded 1 file(s). Total files in destination location: 2 |
+----------------------------------------------------------+
Returned 1 row(s) in 0.26s
[localhost:21000] > select * from sample;
Query: select * from sample
Query finished, fetching results ...
+-------+
| x     |
+-------+
| 2     |
| 4     |
| 6     |
| 8     |
| 64738 |
| 49152 |
| 1     |
| 5     |
| 150   |
+-------+
Returned 9 row(s) in 0.22s

[localhost:21000] > load data inpath '/user/cloudera/unauthorized_data' into table sample;
Query: load data inpath '/user/cloudera/unauthorized_data' into table sample
ERROR: AuthorizationException: User 'cloudera' does not have privileges to access: hdfs://127.0.0.1:8020/user/cloudera/unauthorized_data

Separating Administrator Responsibility from Read and Write Privileges

Remember that to create a database requires full privilege on that database, while day-to-day operations on tables within that database can be performed with lower levels of privilege on specific table. Thus, you might set up separate roles for each database or application: an administrative one that could create or drop the database, and a user-level one that can access only the relevant tables.

For example, this policy file divides responsibilities between users in 3 different groups:

  • Members of the supergroup group have the training_sysadmin role and so can set up a database named training.
  • Members of the cloudera group have the instructor role and so can create, insert into, and query any tables in the training database, but cannot create or drop the database itself.
  • Members of the visitor group have the student role and so can query those tables in the training database.
[groups]
supergroup = training_sysadmin
cloudera = instructor
visitor = student

[roles]
training_sysadmin = server=server1->db=training
instructor = server=server1->db=training->table=*->action=*
student = server=server1->db=training->table=*->action=SELECT

Using Multiple Policy Files for Different Databases

For an Impala cluster with many databases being accessed by many users and applications, it might be cumbersome to update the security policy file for each privilege change or each new database, table, or view. You can allow security to be managed separately for individual databases, by setting up a separate policy file for each database:

  • Add the optional [databases] section to the main policy file.
  • Add entries in the [databases] section for each database that has its own policy file.
  • For each listed database, specify the HDFS path of the appropriate policy file.

For example:

[databases]
# Defines the location of the per-DB policy files for the 'customers' and 'sales' databases.
customers = hdfs://ha-nn-uri/etc/access/customers.ini
sales = hdfs://ha-nn-uri/etc/access/sales.ini

To enable URIs in per-DB policy files, add the following string in the Cloudera Manager field Impala Service Environment Advanced Configuration Snippet (Safety Valve):

JAVA_TOOL_OPTIONS="-Dsentry.allow.uri.db.policyfile=true"

Setting Up Schema Objects for a Secure Impala Deployment

Remember that in your role definitions, you specify privileges at the level of individual databases and tables, or all databases or all tables within a database. To simplify the structure of these rules, plan ahead of time how to name your schema objects so that data with different authorization requirements is divided into separate databases.

If you are adding security on top of an existing Impala deployment, remember that you can rename tables or even move them between databases using the ALTER TABLE statement. In Impala, creating new databases is a relatively inexpensive operation, basically just creating a new directory in HDFS.

You can also plan the security scheme and set up the policy file before the actual schema objects named in the policy file exist. Because the authorization capability is based on whitelisting, a user can only create a new database or table if the required privilege is already in the policy file: either by listing the exact name of the object being created, or a * wildcard to match all the applicable objects within the appropriate container.

Privilege Model and Object Hierarchy

Privileges can be granted on different objects in the schema. Any privilege that can be granted is associated with a level in the object hierarchy. If a privilege is granted on a container object in the hierarchy, the child object automatically inherits it. This is the same privilege model as Hive and other database systems such as MySQL.

The kinds of objects in the schema hierarchy are:

Server
URI
Database
  Table

The server name is specified by the -server_name option when impalad starts. Specify the same name for all impalad nodes in the cluster.

URIs represent the HDFS paths you specify as part of statements such as CREATE EXTERNAL TABLE and LOAD DATA. Typically, you specify what look like UNIX paths, but these locations can also be prefixed with hdfs:// to make clear that they are really URIs. To set privileges for a URI, specify the name of a directory, and the privilege applies to all the files in that directory and any directories underneath it.

In CDH 5.5 / Impala 2.3 and higher, you can specify privileges for individual columns, as described in Column-level Authorization. Formerly, to specify read privileges at this level, you created a view that queried specific columns and/or partitions from a base table, and gave SELECT privilege on the view but not the underlying table.

URIs must start with either hdfs:// or file://. If a URI starts with anything else, it will cause an exception and the policy file will be invalid. When defining URIs for HDFS, you must also specify the NameNode. For example:
data_read = server=server1->uri=file:///path/to/dir, \
server=server1->uri=hdfs://namenode:port/path/to/dir
Valid privilege types and objects they apply to
Privilege Object
INSERT DB, TABLE
SELECT DB, TABLE, VIEW, COLUMN
ALL SERVER, TABLE, DB, URI
Operation Scope Privileges Required URI
CREATE DATABASE SERVER ALL  
DROP DATABASE DATABASE ALL  
CREATE TABLE DATABASE ALL  
DROP TABLE TABLE ALL  
CREATE VIEW

-This operation is allowed if you have column-level SELECT access to the columns being used.

DATABASE; SELECT on TABLE; ALL  
ALTER VIEW

-This operation is allowed if you have column-level SELECT access to the columns being used.

VIEW/TABLE ALL  
DROP VIEW VIEW/TABLE ALL  
ALTER TABLE .. ADD COLUMNS TABLE ALL  
ALTER TABLE .. REPLACE COLUMNS TABLE ALL  
ALTER TABLE .. CHANGE column TABLE ALL  
ALTER TABLE .. RENAME TABLE ALL  
ALTER TABLE .. SET TBLPROPERTIES TABLE ALL  
ALTER TABLE .. SET FILEFORMAT TABLE ALL  
ALTER TABLE .. SET LOCATION TABLE ALL URI
ALTER TABLE .. ADD PARTITION TABLE ALL  
ALTER TABLE .. ADD PARTITION location TABLE ALL URI
ALTER TABLE .. DROP PARTITION TABLE ALL  
ALTER TABLE .. PARTITION SET FILEFORMAT TABLE ALL  
SHOW CREATE TABLE TABLE SELECT/INSERT  
SHOW PARTITIONS TABLE SELECT/INSERT  
SHOW TABLES

-Output includes all the tables for which the user has table-level privileges and all the tables for which the user has some column-level privileges.

TABLE SELECT/INSERT  
SHOW GRANT ROLE

-Output includes an additional field for any column-level privileges.

TABLE SELECT/INSERT  
DESCRIBE TABLE

-Output shows all columns if the user has table level-privileges or SELECT privilege on at least one table column

TABLE SELECT/INSERT  
LOAD DATA TABLE INSERT URI
SELECT

-You can grant the SELECT privilege on a view to give users access to specific columns of a table they do not otherwise have access to.

-See Column-level Authorization for details on allowed column-level operations.

VIEW/TABLE; COLUMN SELECT  
INSERT OVERWRITE TABLE TABLE INSERT  
CREATE TABLE .. AS SELECT

-This operation is allowed if you have column-level SELECT access to the columns being used.

DATABASE; SELECT on TABLE ALL  
USE <dbName> Any    
CREATE FUNCTION SERVER ALL  
ALTER TABLE .. SET SERDEPROPERTIES TABLE ALL  
ALTER TABLE .. PARTITION SET SERDEPROPERTIES TABLE ALL  
EXPLAIN SELECT TABLE; COLUMN SELECT  
EXPLAIN INSERT TABLE; COLUMN INSERT  
INVALIDATE METADATA SERVER ALL  
INVALIDATE METADATA <table name> TABLE SELECT/INSERT  
REFRESH <table name> or REFRESH <table name> PARTITION (<partition_spec>) TABLE SELECT/INSERT  
DROP FUNCTION SERVER ALL  
COMPUTE STATS TABLE ALL  

Debugging Failed Sentry Authorization Requests

Sentry logs all facts that lead up to authorization decisions at the debug level. If you do not understand why Sentry is denying access, the best way to debug is to temporarily turn on debug logging:
  • In Cloudera Manager, add log4j.logger.org.apache.sentry=DEBUG to the logging settings for your service through the corresponding Logging Safety Valve field for the Impala, Hive Server 2, or Solr Server services.
  • On systems not managed by Cloudera Manager, add log4j.logger.org.apache.sentry=DEBUG to the log4j.properties file on each host in the cluster, in the appropriate configuration directory for each service.
Specifically, look for exceptions and messages such as:
FilePermission server..., RequestPermission server...., result [true|false]
which indicate each evaluation Sentry makes. The FilePermission is from the policy file, while RequestPermission is the privilege required for the query. A RequestPermission will iterate over all appropriate FilePermission settings until a match is found. If no matching privilege is found, Sentry returns false indicating "Access Denied" .

Managing Sentry for Impala through Cloudera Manager

To enable the Sentry service for Impala and Hive, set Hive/Impala > Service-Wide > Sentry Service parameter to the Sentry service. Then restart Impala and Hive. Simply adding Sentry service as a dependency and restarting enables Impala and Hive to use the Sentry service.

To set the server name to use when granting server level privileges, set the Hive > Service-Wide > Advanced > Server Name for Sentry Authorization parameter. When using Sentry with the Hive Metastore, you can specify the list of users that are allowed to bypass Sentry Authorization in Hive Metastore using Hive > Service-Wide > Security > Bypass Sentry Authorization Users. These are usually service users that already ensure all activity has been authorized.

The DEFAULT Database in a Secure Deployment

Because of the extra emphasis on granular access controls in a secure deployment, you should move any important or sensitive information out of the DEFAULT database into a named database whose privileges are specified in the policy file. Sometimes you might need to give privileges on the DEFAULT database for administrative reasons; for example, as a place you can reliably specify with a USE statement when preparing to drop a database.