The gpperfmon Database
The gpperfmon Database
The gpperfmon database is a dedicated database where data collection agents on Greenplum segment hosts save query and system statistics. The optional Greenplum Command Center management tool depends upon the gpperfmon database for query history.
The gpperfmon database is created using the gpperfmon_install command-line utility. The utility creates the database and the gpmon database role and enables the data collection agents on the master and segment hosts. See the gpperfmon_install reference in the Greenplum Database Utility Guide for information about using the utility and configuring the data collection agents.
The gpperfmon database consists of three sets of tables that capture query and system status information at different stages.
- _now tables store current system metrics such as active queries.
- _tail tables are used to stage data before it is saved to the _history tables. The _tail tables are for internal use only and not to be queried by users.
- _history tables store historical metrics.
The data for _now and _tail tables are stored as text files on the master host file system, and are accessed in the gpperfmon database via external tables. The history tables are regular heap database tables in the gpperfmon database. History is saved only for queries that run for a minimum number of seconds, 20 by default. You can set this threshold to another value by setting the min_query_time parameter in the $MASTER_DATA_DIRECTORY/gpperfmon/conf/gpperfmon.conf configuration file. Setting the value to 0 saves history for all queries.
The history tables are partitioned by month. See History Table Partition Retention for information about removing old partitions.
The database contains the following categories of tables:
- The database_* tables store query workload information for a Greenplum Database instance.
- The diskspace_* tables store diskspace metrics.
- The log_alert_* tables store error and warning messages from pg_log.
- The queries_* tables store high-level query status information.
- The segment_* tables store memory allocation statistics for the Greenplum Database segment instances.
- The socket_stats_* tables store statistical metrics about socket usage for a Greenplum Database instance. Note: These tables are in place for future use and are not currently populated.
- The system_* tables store system utilization metrics.
The gpperfmon database also contains the following views:
- The dynamic_memory_info view shows an aggregate of all the segments per host and the amount of dynamic memory used per host.
- The memory_info view shows per-host memory information from the system_history and segment_history tables.
History Table Partition Retention
The history tables in the gpperfmon database are partitioned by month. Partitions are automatically added in two month increments as needed.
The partition_age parameter in the $MASTER_DATA_DIRECTORY/gpperfmon/conf/gpperfmon.conf file can be set to the maximum number of monthly partitions to keep. Partitions older than the specified value are removed automatically when new partitions are added.
The default value for partition_age is 0, which means that administrators must manually remove unneeded partitions.
Alert Log Processing and Log Rotation
When the gp_gperfmon_enable server configuration parameter is set to true, the Greenplum Database syslogger writes alert messages to a .csv file in the $MASTER_DATA_DIRECTORY/gpperfmon/logs directory.
The level of messages written to the log can be set to none, warning, error, fatal, or panic by setting the gpperfmon_log_alert_level server configuration parameter in postgresql.conf. The default message level is warning.
The directory where the log is written can be changed by setting the log_location configuration variable in the $MASTER_DATA_DIRECTORY/gpperfmon/conf/gpperfmon.conf configuration file.
The syslogger rotates the alert log every 24 hours or when the current log file reaches or exceeds 1MB.
A rotated log file can exceed 1MB if a single error message contains a large SQL statement or a large stack trace. Also, the syslogger processes error messages in chunks, with a separate chunk for each logging process. The size of a chunk is OS-dependent; on Red Hat Enterprise Linux, for example, it is 4096 bytes. If many Greenplum Database sessions generate error messages at the same time, the log file can grow significantly before its size is checked and log rotation is triggered.
gpperfmon Data Collection Process
When Greenplum Database starts up with gpperfmon support enabled, it forks a gpmmon agent process. gpmmon then starts a gpsmon agent process on the master host and every segment host in the Greenplum Database cluster. The Greenplum Database postmaster process monitors the gpmmon process and restarts it if needed, and the gpmmon process monitors and restarts gpsmon processes as needed.
The gpmmon process runs in a loop and at configurable intervals retrieves data accumulated by the gpsmon processes, adds it to the data files for the _now and _tail external database tables, and then into the _history regular heap database tables.
Two configuration parameters in the $MASTER_DATA_DIRECTORY/gpperfmon/conf/gpperfmon.conf configuration file control how often gpmmon activities are triggered:
- The quantum parameter is how frequently, in seconds, gpmmon requests data from the gpsmon agents on the segment hosts and adds retrieved data to the _now and _tail external table data files. Valid values for the quantum parameter are 10, 15, 20, 30, and 60. The default is 15.
- The harvest_interval parameter is how frequently, in seconds, data in the _tail tables is moved to the _history tables. The harvest_interval must be at least 30. The default is 120.
See the gpperfmon_install management utility reference in the Greenplum Database Utility Guide for the complete list of gpperfmon configuration parameters.
The following steps describe the flow of data from Greenplum Database into the gpperfmon database when gpperfmon support is enabled.
- While executing queries, the Greenplum Database query dispatcher and query executor processes send out query status messages in UDP datagrams. The gp_gpperfmon_send_interval server configuration variable determines how frequently the database sends these messages. The default is every second.
- The gpsmon process on each host receives the UDP packets, consolidates and summarizes the data they contain, and adds additional host metrics, such as CPU and memory usage.
- The gpsmon processes continue to accumulate data until they receive a dump command from gpmmon.
- The gpsmon processes respond to a dump command by sending their accumulated status data and log alerts to a listening gpmmon event handler thread.
- The gpmmon event handler saves the metrics to .txt files in the $MASTER_DATA_DIRECTORY/gpperfmon/data directory on the master host.
At each quantum interval (15 seconds by default), gpmmon performs the following steps:
- Sends a dump command to the gpsmon processes.
- Gathers and converts the .txt files saved in the
$MASTER_DATA_DIRECTORY/gpperfmon/data directory into .dat
external data files for the _now and _tail external
tables in the gpperfmon database.
For example, disk space metrics are added to the diskspace_now.dat and _diskspace_tail.dat delimited text files. These text files are accessed via the diskspace_now and _diskspace_tail tables in the gpperfmon database.
- Renames the _tail file to a _stage file.
- Creates a new _tail file.
- Appends data from the _stage file into the _tail file.
- Runs a SQL command to insert the data from the _tail external table
into the corresponding _history table.
For example, the contents of the _database_tail external table is inserted into the database_history regular (heap) table.
- Deletes the _tail file after its contents have been loaded into the database table.
- Gathers all of the gpdb-alert-*.csv files in the $MASTER_DATA_DIRECTORY/gpperfmon/logs directory (except the most recent, which the syslogger has open and is writing to) into a single file, alert_log_stage.
- Loads the alert_log_stage file into the log_alert_history table in the gpperfmon database.
- Truncates the alert_log_stage file.
The following topics describe the contents of the tables in the gpperfmon database.