Limiting the number and length of client connections
You can manage how many active sessions a user can open to the server, and the duration of those sessions. Doing so helps prevent overuse of available resources, and can improve overall throughput.
You can define connection limits at two levels:
Set the MAXCONNECTIONS property on individual users. This property specifies how many sessions a user can open concurrently on individual nodes, or across the database cluster. For example, the following ALTER USER statement allows user Joe up to 10 concurrent sessions:
=> ALTER USER Joe MAXCONNECTIONS 10 ON DATABASE;
Set the configuration parameter MaxClientSessions on the database or individual nodes. This parameter specifies the maximum number of client sessions that can run on nodes in the database cluster, by default set to 50. An extra five sessions are always reserved to dbadmin users. This enables them to log in when the total number of client sessions equals MaxClientSessions.
Total client connections to a given node cannot exceed the limits set in MaxClientSessions.
Changes to a client's MAXCONNECTIONS property have no effect on current sessions; these changes apply only to new sessions. For example, if you change user's connection mode from DATABASE to NODE, current node connections are unaffected. This change applies only to new sessions, which are reserved on the invoking node.
Managing TCP keepalive settings
Vertica uses kernel TCP keepalive parameters to detect unresponsive clients and determine when the connection should be closed.Vertica also supports a set of equivalent KeepAlive parameters that can override TCP keepalive parameter settings. By default, all Vertica KeepAlive parameters are set to 0, which signifies to use TCP keepalive settings. To override TCP keepalive settings, set the equivalent parameters at the database level with ALTER DATABASE, or for the current session with ALTER SESSION.
|TCP keepalive Parameter||Vertica Parameter||Description|
|tcp_keepalive_time||KeepAliveIdleTime||Length (in seconds) of the idle period before the first TCP keepalive probe is sent to ensure that the client is still connected.|
|tcp_keepalive_probes||KeepAliveProbeCount||Number of consecutive keepalive probes that must go unacknowledged by the client before the client connection is considered lost and closed|
|tcp_keepalive_intvl||KeepAliveProbeInterval||Time interval (in seconds) between keepalive probes.|
The following examples show how to use Vertica KeepAlive parameters to override TCP keepalive parameters as follows:
After 600 seconds (ten minutes), the first keepalive probe is sent to the client.
Consecutive keepalive probes are sent every 30 seconds.
If the client fails to respond to 10 keepalive probes, the connection is considered lost and is closed.
To make this the default policy for client connections, use ALTER DATABASE:
=> ALTER DATABASE DEFAULT SET KeepAliveIdleTime = 600; => ALTER DATABASE DEFAULT SET KeepAliveProbeInterval = 30; => ALTER DATABASE DEFAULT SET KeepAliveProbeCount = 10;
To override database-level policies for the current session, use ALTER SESSION:
=> ALTER SESSION SET KeepAliveIdleTime = 400; => ALTER SESSION SET KeepAliveProbeInterval = 72; => ALTER SESSION SET KeepAliveProbeCount = 60;
Query system table CONFIGURATION_PARAMETERS to verify database and session settings of the three Vertica KeepAlive parameters:
=> SELECT parameter_name, database_value, current_value FROM configuration_parameters WHERE parameter_name ILIKE 'KeepAlive%'; parameter_name | database_value | current_value ------------------------+----------------+--------------- KeepAliveProbeCount | 10 | 60 KeepAliveIdleTime | 600 | 400 KeepAliveProbeInterval | 30 | 72 (3 rows)
Limiting idle session length
If a client continues to respond to TCP keepalive probes, but is not running any queries, the client's session is considered idle. Idle sessions eventually time out. The maximum time that sessions are allowed to idle can be set at three levels, in descending order of precedence:
- As dbadmin, set the IDLESESSIONTIMEOUT property for individual users. This property overrides all other session timeout settings.
- Users can limit the idle time of the current session with SET SESSION IDLESESSIONTIMEOUT. Non-superusers can only set their session idle time to a value equal to or lower than their own IDLESESSIONTIMEOUT setting. If no session idle time is explicitly set for a user, the session idle time for that user is inherited from the node or database settings.
- As dbadmin, set configuration parameter DEFAULTIDLESESSIONTIMEOUT. on the database or on individual nodes. This You can limit the default database cluster or individual nodes, with configuration parameter DEFAULTIDLESESSIONTIMEOUT. This parameter sets the default timeout value for all non-superusers.
All settings apply to sessions that are continuously idle—that is, sessions where no queries are running. If a client is slow or unresponsive during query execution, that time does not apply to timeouts. For example, the time that is required for a streaming batch insert is not counted towards timeout. The server identifies a session as idle starting from the moment it starts to wait for any type of message from that session.
Viewing session settings
The following sections demonstrate how you can query the database for details about the session and connection limits.
Session length limits
Use SHOW DATABASE to view the session length limit for the database:
=> SHOW DATABASE DEFAULT DEFAULTIDLESESSIONTIMEOUT; name | setting ---------------------------+--------- DefaultIdleSessionTimeout | 2 day (1 row)
Use SHOW to view the length limit for the current session:
=> SHOW IDLESESSIONTIMEOUT; name | setting --------------------+--------- idlesessiontimeout | 1 (1 row)
Use SHOW DATABASE to view the connection limits for the database:
=> SHOW DATABASE DEFAULT MaxClientSessions; name | setting -------------------+--------- MaxClientSessions | 50 (1 row)
Query USERS to view the connection limits for users:
=> SELECT user_name, max_connections, connection_limit_mode FROM users WHERE user_name != 'dbadmin'; user_name | max_connections | connection_limit_mode -----------+-----------------+----------------------- SuzyX | 3 | database Joe | 10 | database (2 rows)
Closing user sessions
To manually close a user session, use CLOSE_USER_SESSIONS:
=> SELECT CLOSE_USER_SESSIONS ('Joe'); close_user_sessions ------------------------------------------------------------------------------ Close all sessions for user Joe sent. Check v_monitor.sessions for progress. (1 row)
A user executes a query, and for some reason the query takes an unusually long time to finish (for example, because of server traffic or query complexity). In this case, the user might think the query failed, and opens another session to run the same query. Now, two sessions run the same query, using an extra connection.
To prevent this situation, you can limit how many sessions individual users can run, by modifying their MAXCONNECTIONS user property. This can help minimize the chances of running redundant queries. It also helps prevent users from consuming all available connections, as set by the database For example, the following setting on user
SuzyQ limits her to two database sessions at any time:
=> CREATE USER SuzyQ MAXCONNECTIONS 2 ON DATABASE;
Limiting Another issue setting client connections prevents is when a user connects to the server many times. Too many user connections exhausts the number of allowable connections set by database configuration parameter MaxClientSessions.
NoteNo user can have a MAXCONNECTIONS limit greater than the MaxClientSessions setting.
Cluster changes and connections
Behavior changes can occur with client connection limits when the following changes occur to a cluster:
You add or remove a node.
A node goes down or comes back up.
Changes in node availability between connection requests have little impact on connection limits.
In terms of honoring connection limits, no significant impact exists when nodes go down or come up in between connection requests. No special actions are needed to handle this. However, if a node goes down, its active session exits and other nodes in the cluster also drop their sessions. This frees up connections. The query may hang in which case the blocked sessions are reasonable and as expected.