Setting up Exporters

The Linux instructions below use RHEL, but any Linux-based system should work. Crunchy Data customers can obtain Linux packages through the Crunchy Customer Portal.

• Installation
  • RPM installs
  • Non-RPMs installs
• Upgrading
• Setup
  • RHEL or CentOS
• Metrics Collected
  • PostgreSQL
  • System
    

Installation

RPM installs

The following RPM packages are available to Crunchy Data customers through the Crunchy Customer Portal. After installing via these packages, continue reading at the Setup section.

Available Packages

Non-RPM installs

For non-package installations on Linux, applications can be downloaded from their respective repositories:

User and Configuration Directory Installation

You will need to create a user named ccp_monitoring which you can do with the following command:

sudo useradd -m -d /var/lib/ccp_monitoring ccp_monitoring

Configuration File Installation

All executables installed via the above releases are expected to be in the /usr/bin directory. A base node_exporter systemd file is expected to be in place already. An example one can be found here:

https://github.com/lest/prometheus-rpm/tree/master/node_exporter

A base blackbox_exporter systemd file is also expected to be in place. No examples are currently available.

The files contained in this repository are assumed to be installed in the following locations with the following names. In the instructions below, you should replace a double-hash (##) with the two-digit major version of PostgreSQL you are running (ex: 12, 13, 14, etc.).

node_exporter

The node_exporter data directory should be /var/lib/ccp_monitoring/node_exporter and owned by the ccp_monitoring user. You can set it up with:

sudo install -m 0700 -o ccp_monitoring -g ccp_monitoring -d /var/lib/ccp_monitoring/node_exporter

The following pgMonitor configuration files should be placed according to the following mapping:

postgres_exporter

The following pgMonitor configuration files should be placed according to the following mapping:

blackbox_exporter

The following pgMonitor configuration files should be placed according to the following mapping:

Upgrading

• See the CHANGELOG for full details on both major & minor version upgrades.

Setup

Setup on RHEL or CentOS

Service Configuration

The following files contain defaults that should enable the exporters to run effectively on your system for the purposes of using pgMonitor. Please take some time to review them.

If you need to modify them, see the notes in the files for more details and recommendations: - /etc/systemd/system/node_exporter.service.d/crunchy-node-exporter-service-rhel - /etc/sysconfig/node_exporter - /etc/sysconfig/postgres_exporter_pg## - /etc/sysconfig/postgres_exporter_pg##_per_db

Note that /etc/sysconfig/postgres_exporter_pg## & postgres_exporter_pg##_per_db are the default sysconfig files for monitoring the database running on the local socket at /var/run/postgresql and connect to the “postgres” database. If you’ve installed the pgMonitor setup to a different database, modify these files accordingly or make new ones. If you make new ones, ensure the service name you enable references this file (see the Enable Services section below ).

Database Configuration

General Configuration

First, make sure you have installed the PostgreSQL contrib modules. You can install them with the following command:

sudo yum install postgresql##-contrib

Where ## corresponds to your current PostgreSQL version. For PostgreSQL 13 this would be:

sudo yum install postgresql13-contrib

You will need to modify your postgresql.conf configuration file to tell PostgreSQL to load shared libraries. In the default setup, this file can be found at /var/lib/pgsql/##/data/postgresql.conf .

shared_preload_libraries = 'pg_stat_statements,auto_explain'

You will need to restart your PostgreSQL instance for the change to take effect. pgMonitor has optional metrics that can be collected via pg_stat_statements. auto_explain does not do anything to your database without further configuration. But even if neither of these extensions are initially used, they are very good to have enabled here by default for when they may be needed in the future.

The following statement only needs to be run on the “global” database, typically the “postgres” database. If you want the pg_stat_statements view to be visible in other databases, this statement must be run there as well.

CREATE EXTENSION pg_stat_statements;

Monitoring Setup

By default, there are two postgres_exporter services expected to be running. One connects to the default postgres database that most PostgreSQL instances come with and is meant for collecting global metrics that are the same on all databases in the instance (connection/replication statistics, etc). This service uses the sysconfig file postgres_exporter_pg## . Connect to this database and run the setup.sql script to install the required database objects for pgMonitor.

The second postgres_exporter service is used to collect per-database metrics and uses the sysconfig file postgres_exporter_pg##_per_db . By default it is set to also connect to the postgres database, but you can add as many additional connection strings to this service for each individual database that you want metrics for. Per-db metrics include things like table/index statistics and bloat. See the section below for monitorig multiple databases for how to do this.

Note that your pg_hba.conf will have to be configured to allow the ccp_monitoring system user to connect as the ccp_monitoring role to any database in the instance. As of version 4.0 of pg_monitor, the postgres_exporter service is set by default to connect via local socket, so passwordless local peer authentication is the expected default. If password-based authentication is required, we recommend using SCRAM authentication, which is supported as of version 0.7.x of postgres_exporter. See our blog post for more information on SCRAM - https://info.crunchydata.com/blog/how-to-upgrade-postgresql-passwords-to-scram

postgres_exporter only takes a single yaml file as an argument for custom queries, so this requires concatenating the relevant files together. The sysconfig files for the service help with this concatenation task and define the variable QUERY_FILE_LIST . Set this variable to a space delimited list of the full path names to all files that contain queries you want to be in the single file that postgres_exporter uses.

For example, to use just the common queries for PostgreSQL 12 modify the relevant sysconfig file as follows:

QUERY_FILE_LIST="/etc/postgres_exporter/12/queries_global.yml /etc/postgres_exporter/12/queries_general.yml"

As an another example, to include queries for PostgreSQL 13 as well as pgBackRest, modify the relevant sysconfig file as follows:

QUERY_FILE_LIST="/etc/postgres_exporter/13/queries_global.yml /etc/postgres_exporter/13/queries_general.yml /etc/postgres_exporter/13/queries_backrest.yml"

For replica servers, the setup is the same except that the setup.sql file does not need to be run since writes cannot be done there and it was already run on the primary.

Access Control: GRANT statements

The ccp_monitoring database role (created by running the “setup.sql” file above) must be allowed to connect to all databases in the cluster. Note that by default, all users are granted CONNECT on all new databases, so this step can likely be skipped. Otherwise, run the following command to generate the necessary GRANT statements:

SELECT 'GRANT CONNECT ON DATABASE "' || datname || '" TO ccp_monitoring;'
FROM pg_database
WHERE datallowconn = true;

This should generate one or more statements similar to the following:

GRANT CONNECT ON DATABASE "postgres" TO ccp_monitoring;

Run these grant statements to then allow monitoring to connect.

Materialized View Metrics

With large databases/tables and some other conditions, certain metrics can cause excessive load. For those cases, materialized views and alternative metric queries have been made available. The materialized views are refreshed on their own schedule independent of the Prometheus data scrape, so any load that may be associated with gathering the underlying data is mitigated. A configuration table, seen below, contains options for how often these materialized views should be refreshed. And a single procedure can be called to refresh all materialized views relevant to monitoring.

For every database that will be collecting materialized view metrics, you will have to run the setup_metric_views.sql file against that database. This will likely need to be run as a superuser and must be run after running the base setup file mentioned above to create the necessary monitoring user first.

psql -U postgres -d alphadb -f setup_metric_views.sql
psql -U postgres -d betadb -f setup_metric_views.sql

The /etc/postgres_exporter/##/crontab.txt file has an example entry for how to call the refresh procedure. You should modify this to run as often as you need depending on how recent you need your metric data to be. This procedure is safe to run on the primary or replicas and will safely exit if the database is in recovery mode.

Configuration table monitor.metric_views :

You are also free to use this materialized view system for your own custom metrics as well. Simply make a materialized view, add its name to the configuration table and ensure the user running the refresh has permissions to do so for your view(s).

Bloat setup

Run the script on the specific database(s) you will be monitoring for bloat in the cluster. See the note below, or in crontab.txt, concerning a superuser requirement for using this script.

psql -d postgres -c "CREATE EXTENSION pgstattuple;"
/usr/bin/pg_bloat_check.py -c "host=localhost dbname=postgres user=postgres" --create_stats_table
psql -d postgres -c "GRANT SELECT,INSERT,UPDATE,DELETE,TRUNCATE ON bloat_indexes, bloat_stats, bloat_tables TO ccp_monitoring;"

The /etc/postgres_exporter/##/crontab.txt file has an example bloat check crontab entry. Modify this example to schedule bloat checking weekly during your ‘off-peak’ hours; alternatively, scheduling it monthly is usually good enough for most databases as long as the results are acted upon quickly.

GRANT pg_read_all_data TO ccp_monitoring;

Blackbox Exporter

The configuration file for the blackbox_exporter provided by pgMonitor (/etc/blackbox_exporter/crunchy-blackbox.yml ) provides a probe for monitoring any IPv4 TCP port status. The actual target and port being monitored are controlled via the Prometheus target configuration system. Please see the pgMonitor Prometheus documentation for further details. If any additional Blackbox probes are desired, please see the upstream documentation.

PGBouncer

In order to monitor pgbouncer with pgMonitor, the pgbouncer_fdw maintained by CrunchyData is required. Please see its repository for full installation instructions. A package for this is available for Crunchy customers.

https://github.com/CrunchyData/pgbouncer_fdw

Once that is working, you should be able to add the queries_pgbouncer.yml file to the QUERY_FILE_LIST for the exporter that is monitoring the database where the FDW was installed.

Enable Services

sudo systemctl enable node_exporter
sudo systemctl start node_exporter
sudo systemctl status node_exporter

If you’ve installed the blackbox exporter:

sudo systemctl enable blackbox_exporter
sudo systemctl start blackbox_exporter
sudo systemctl status blackbox_exporter

To most easily allow the use of multiple postgres exporters, running multiple major versions of PostgreSQL, and to avoid maintaining many similar service files, a systemd template service file is used. The name of the sysconfig EnvironmentFile to be used by the service is passed as the value after the “@” and before “.service” in the service name. The default exporter’s sysconfig file is named “postgres_exporter_pg##” and tied to the major version of postgres that it was installed for. A similar EnvironmentFile exists for the per-db service. Be sure to replace the ## in the below commands first!

sudo systemctl enable crunchy-postgres-exporter@postgres_exporter_pg##
sudo systemctl start crunchy-postgres-exporter@postgres_exporter_pg##
sudo systemctl status crunchy-postgres-exporter@postgres_exporter_pg##

sudo systemctl enable crunchy-postgres-exporter@postgres_exporter_pg##_per_db
sudo systemctl start crunchy-postgres-exporter@postgres_exporter_pg##_per_db
sudo systemctl status crunchy-postgres-exporter@postgres_exporter_pg##_per_db

Monitoring multiple databases and/or running multiple postgres exporters (RHEL / CentOS)

Certain metrics are not cluster-wide, so multiple exporters must be run to avoid duplication when monitoring multiple databases in a single PostgreSQL instance. To collect these per-database metrics, an additional exporter service is required and pgMonitor provides this using the following query file: (queries_per_db.yml ). In Prometheus, you can then define the global and per-db exporter targets for a single job. This will place all the metrics that are collected for a single database instance together.

pgMonitor provides and recommends an example sysconfig file for this per-db exporter: sysconfig.postgres_exporter_pg##_per_db . If you’d like to create additional exporter services for different query files, just copy the existing ones and modify the relevant lines, mainly the port, database name, and query file. The below example shows connecting to 3 databases in the same instance to collect their per-db metrics: postgres, mydb1, and mydb2.

OPT="--web.listen-address=0.0.0.0:9188 --extend.query-path=/etc/postgres_exporter/14/queries_per_db.yml"
DATA_SOURCE_NAME="postgresql:///postgres?host=/var/run/postgresql/&user=ccp_monitoring&sslmode=disable,postgresql:///mydb1?host=/var/run/postgresql/&user=ccp_monitoring&sslmode=disable,postgresql:///mydb2?host=/var/run/postgresql/&user=ccp_monitoring&sslmode=disable"

As was done with the exporter service that is collecting the global metrics, also modify the QUERY_LIST_FILE in the new sysconfig file to only collect per-db metrics

QUERY_FILE_LIST="/etc/postgres_exporter/14/queries_per_db.yml"

Since a systemd template is used for the postgres_exporter services, all you need to do is pass the sysconfig file name as part of the new service name.

sudo systemctl enable crunchy-postgres-exporter@postgres_exporter_pg14_per_db
sudo systemctl start cruncy-postgres-exporter@postgres_exporter_pg14_per_db
sudo systemctl status crunchy-postgres-exporter@postgres_exporter_pg14_per_db

Lastly, update the Prometheus auto.d target file to include the new exporter in the same job you already had running for this system

Metrics Collected

The metrics collected by our exporters are outlined below.

PostgreSQL

PostgreSQL metrics are collected by the postgres_exporter. pgMonitor uses custom queries for its PG metrics. The default metrics that postgres_exporter comes with are all disabled except for the pg_up metric.

General Metrics

pg_up - Database is up and connectable by metric collector. This is the only metrics that comes with postgres_exporter that is currently used

Common Metrics

Metrics contained in the queries_global.yml file. These metrics are common to all versions of PostgreSQL and are recommended as a minimum default for the global exporter.

• ccp_archive_command_status_seconds_since_last_fail - Seconds since the last archive_command run failed. If zero, the archive_command is succeeding without error.

• ccp_database_size_bytes - Total size of each database in PostgreSQL instance

• ccp_is_in_recovery_status - Current value of the pg_is_in_recovery() function expressed as 1 for true (instance is a replica) and 2 for false (instance is a primary)

• ccp_connection_stats_active - Count of active connections

• ccp_connection_stats_idle - Count of idle connections

• ccp_connection_stats_idle_in_txn - Count of idle in transaction connections

• ccp_connection_stats_max_blocked_query_time - Runtime of longest running query that has been blocked by a heavyweight lock

• ccp_connection_stats_max_connections - Current value of max_connections for reference

• ccp_connection_stats_max_idle_in_txn_time - Runtime of longest idle in transaction (IIT) session.

• ccp_connection_stats_max_query_time - Runtime of longest general query (inclusive of IIT).

• ccp_connection_stats_max_blocked_query_time - Runtime of the longest running query that has been blocked by a heavyweight lock

• ccp_locks_count - Count of active lock types per database

• ccp_pg_hba_checksum_status - Value of checksum monitioring status for pg_catalog.pg_hba_file_rules (pg_hba.conf). 0 = valid config. 1 = settings changed. Settings history is available for review in the table monitor.pg_hba_checksum. To reset current config to valid after alert, run monitor.pg_hba_checksum_set_valid(). Note this will clear the history table.

• ccp_pg_settings_checksum_status - Value of checksum monitioring status for pg_catalog.pg_settings (postgresql.conf). 0 = valid config. 1 = settings changed. Settings history is available for review in the table monitor.pg_settings_checksum. To reset current config to valid after alert, run monitor.pg_settings_checksum_set_valid(). Note this will clear the history table.

• ccp_postmaster_uptime_seconds - Time interval in seconds since PostgreSQL database was last restarted

• ccp_postgresql_version_current - Version of PostgreSQL that this exporter is monitoring. Value is the 6 digit integer returned by the server_version_num PostgreSQL configuration variable to allow easy monitoring for version changes.

• ccp_replication_lag_replay_time - Time since a replica received and replayed a WAL file; only shown on replica instances. Note that this is not the main way to determine if a replica is behind its primary. This metric only monitors the time since the replica replayed the WAL vs when it was received. It also does not monitor when a WAL replay replica completely stops receiving WAL (see received_time metric). It is a secondary metric for monitoring WAL replay on the replica itself. This metric always returns zero on a primary.

• ccp_replication_lag_received_time - Similar to ccp_replication_lag_replay_time, however this value always increases between replay of WAL files. Effective for monitoring that a WAL replay replica has actually received WAL files. Note this will cause false positives when used as an alert for replica lag if the primary receives little to no writes (which means there is no WAL to send). This metric always returns zero on a primary.

• ccp_replication_lag_size_bytes - Only provides values on instances that have attached replicas (primary, cascading replica). Tracks byte lag of every streaming replica connected to this database instance. This is the main way that replication lag is monitored. Note that if you have WAL replay only replicas, this will not be reflected here.

• ccp_replication_slots_active - Active state of given replication slot. 1 = true. 0 = false.

• ccp_replication_slots_retained_bytes - The amount of WAL (in bytes) being retained for given slot.

• ccp_sequence_exhaustion_count - Checks for any sequences that may be close to exhaustion (by default greater than 75% usage). Note this checks the sequences themselves, not the values contained in the columns that use said sequences. Function monitor.sequence_status() can provide more details if run directly on database instance.

• ccp_settings_pending_restart_count - Number of settings from pg_settings catalog in a pending_restart state. This value is from the similarly named column found in pg_catalog.pg_settings.

• ccp_wal_activity_total_size_bytes - Current size in bytes of the WAL directory

• ccp_wal_activity_last_5_min_size_bytes - Current size in bytes of the last 5 minutes of WAL generation. Includes recycled WALs.

The meaning of the following ccp_transaction_wraparound metrics, and how to manage when they are triggered, is covered more extensively in this blog post: https://info.crunchydata.com/blog/managing-transaction-id-wraparound-in-postgresql

• ccp_transaction_wraparound_percent_towards_emergency_autovac - Recommended thresholds set to 75%/95% when first evaluating vacuum settings on new systems. Once those have been reviewed and at least one instance-wide vacuum has been run, recommend thresholds of 110%/125%. Reaching 100% is not a cause for immediate concern, but alerting above 100% for extended periods of time means that autovacuum is not able to keep up with current transaction rate and needs further tuning.

• ccp_transaction_wraparound_percent_towards_wraparound - Recommend thresholds set to 50%/75%. If any of these thresholds is tripped, current vacuum settings must be evaluated and tuned ASAP. If critical threshold is reached, it is vitally important that vacuum be run on tables with old transaction IDs to avoid the cluster being forced to shut down for extended offline maintenance.

The following ccp_stat_bgwriter metrics are statistics collected from the pg_stat_bgwriter view for monitoring performance. These metrics cover important performance information about flushing data out to disk. Please see the documentation for further details on these metrics.

• ccp_stat_bgwriter_buffers_alloc

• ccp_stat_bgwriter_buffers_backend

• ccp_stat_bgwriter_buffers_backend_fsync

• ccp_stat_bgwriter_buffers_checkpoint

• ccp_stat_bgwriter_buffers_clean

The following ccp_stat_database_* metrics are statistics collected from the pg_stat_database view.

• ccp_stat_database_blks_hit

• ccp_stat_database_blks_read

• ccp_stat_database_conflicts

• ccp_stat_database_deadlocks

• ccp_stat_database_tup_deleted

• ccp_stat_database_tup_fetched

• ccp_stat_database_tup_inserted

• ccp_stat_database_tup_returned

• ccp_stat_database_tup_updated

• ccp_stat_database_xact_commit

• ccp_stat_database_xact_rollback

PostgreSQL Version Specific Metrics

The following metrics either require special considerations when monitoring specific versions of PostgreSQL, or are only available for specific versions. These metrics are found in the queries_pg##.yml files, where ## is the major version of PG. Unless otherwise noted, the below metrics are available for all versions of PG. These metrics are recommend as a minimum default for the global exporter.

• ccp_data_checksum_failure_count - PostgreSQL 12 and later only. Total number of checksum failures on this database.

• ccp_data_checksum_failure_time_since_last_failure_seconds - PostgreSQL 12 and later only. Time interval in seconds since the last checksum failure was encountered.

Backup Metrics

Backup monitoring only covers pgBackRest at this time. These metrics are found in the queries_backrest.yml file. These metrics only need to be collected once per database instance so should be collected by the global postgres_exporter.

• ccp_backrest_last_full_backup_time_since_completion_seconds - Time since completion of last pgBackRest FULL backup

• ccp_backrest_last_diff_backup_time_since_completion_seconds - Time since completion of last pgBackRest DIFFERENTIAL backup. Note that FULL backup counts as a successful DIFFERENTIAL for the given stanza.

• ccp_backrest_last_incr_backup_time_since_completion_seconds - Time since completion of last pgBackRest INCREMENTAL backup. Note that both FULL and DIFFERENTIAL backups count as a successful INCREMENTAL for the given stanza.

• ccp_backrest_last_info_runtime_backup_runtime_seconds - Last successful runtime of each backup type (full/diff/incr).

• ccp_backrest_last_info_repo_backup_size_bytes - Actual size of only this individual backup in the pgbackrest repository

• ccp_backrest_last_info_repo_total_size_bytes - Total size of this backup in the pgbackrest repository, including all required previous backups and WAL

• ccp_backrest_last_info_backup_error - Count of errors tracked for this backup. Note this does not track incomplete backups, only errors encountered during the backup (checksum errors, file truncation, invalid headers, etc)

Per-Database Metrics

These are metrics that are only available on a per-database level. These metrics are found in the queries_per_db.yml file. These metrics are optional and recommended for the non-global, per-db postgres_exporter. They can be included in the global exporter as well if the global database needs per-database metrics monitored. Please note that depending on the number of objects in your database, collecting these metrics can greatly increase the storage requirements for Prometheus since all of these metrics are being collected for each individual object.

• ccp_table_size_size_bytes - Table size inclusive of all indexes in that table

The following ccp_stat_user_tables_* metrics are statistics collected from the pg_stat_user_tables. Please see the PG documentation for descriptions of these metrics.

• ccp_stat_user_tables_analyze_count

• ccp_stat_user_tables_autoanalyze_count

• ccp_stat_user_tables_autovacuum_count

• ccp_stat_user_tables_n_tup_del

• ccp_stat_user_tables_n_tup_ins

• ccp_stat_user_tables_n_tup_upd

• ccp_stat_user_tables_vacuum_count

Bloat Metrics

Bloat metrics are only available if the pg_bloat_check script has been setup to run. See instructions above. These metrics are found in the queries_bloat.yml file. These metrics are per-database so, should be used by the per-db postgres_exporter.

• ccp_bloat_check_size_bytes - Size of object in bytes

• ccp_bloat_check_total_wasted_space_bytes - Total wasted space in bytes of given object

pgBouncer Metrics

The following metric prefixes correspond to the SHOW command views found in the pgBouncer documentation. Each column found in the SHOW view is a separate metric under the respective prefix. Ex: ccp_pgbouncer_pools_client_active corresponds to the SHOW POOLS view’s client_active column. These metrics are found in the queries_bouncer.yml file. These metrics only need to be collected once per database instance so should be collected by the global postgres_exporter.

• ccp_pgbouncer_pools - SHOW POOLS

• ccp_pgbouncer_databases - SHOW DATABASES

• ccp_pgbouncer_clients - SHOW CLIENTS

• ccp_pgbouncer_servers - SHOW SERVERS

• ccp_pgbouncer_lists - SHOW LISTS

pg_stat_statements Metrics

Collecting all per-query metrics into Prometheus could greatly increase storage requirements and heavily impact performance without sufficient resources. Therefore the metrics below give simplified numeric metrics on overall statistics and Top N queries. N can be set with the PG_STAT_STATEMENTS_LIMIT variable in the exporter sysconfig file (defaults to 20). Note that the statistics for individual queries can only be reset on PG12+. Prior to that, pg_stat_statements must have all statistics reset to redo the top N queries.

• ccp_pg_stat_statements_top_max_time_ms - Maximum time spent in the statement in milliseconds per database/user/query for the top N queries

• ccp_pg_stat_statements_top_mean_time_ms - Average query runtime in milliseconds per database/user/query for the top N queries

• ccp_pg_stat_statements_top_total_time_ms - Total time spent in the statement in milliseconds per database/user/query for the top N queries

• ccp_pg_stat_statements_total_calls_count - Total number of queries run per user/database

• ccp_pg_stat_statements_total_mean_time_ms - Mean runtime of all queries per user/database

• ccp_pg_stat_statements_total_row_count - Total rows returned from all queries per user/database

• ccp_pg_stat_statements_total_time_ms - Total runtime of all queries per user/database

System

*NIX Operating System metrics (Linux, BSD, etc) are collected using the node_exporter provided by the Prometheus team. pgMonitor only collects the default metrics provided by node_exporter, but many additional metrics are available if needed.