Monitoring through SQL v5

EDB Postgres Distributed provides several monitoring and statistics views that are specific to its distributed nature. The standard Postgres monitoring is also useful for monitoring EDB Postgres Distributed.

Monitoring overview

A PGD group consists of multiple servers, often referred to as nodes. Monitor all of the nodes to ensure the health of the whole group.

The bdr_monitor role can execute the bdr.monitor functions to provide an assessment of PGD health using one of three levels:

  • OK Often shown as green.
  • WARNING Often shown as yellow.
  • CRITICAL Often shown as red.
  • UNKNOWN For unrecognized situations, often shown as red.

PGD also provides dynamic catalog views that show the instantaneous state of various internal metrics. It also provides metadata catalogs that store the configuration defaults and configuration changes the user requests. Some of those views and tables are accessible by bdr_monitor or bdr_read_all_stats, but some contain user or internal information that has higher security requirements.

PGD allows you to monitor each of the nodes individually or to monitor the whole group by access to a single node. If you want to monitor each node individually, connect to each node and issue monitoring requests. If you want to monitor the group from a single node, then use the views starting with bdr.group since these requests make calls to other nodes to assemble a group-level information set.

If you were granted access to the bdr.run_on_all_nodes() function by bdr_superuser, then you can make your own calls to all nodes.

Monitoring node join and removal

By default, the node management functions wait for the join or part operation to complete. You can turn waiting off using the respective wait_for_completion function argument. If waiting is turned off, then to see when a join or part operation finishes, check the node state indirectly using bdr.node_summary and bdr.event_summary.

When called, the helper function bdr.wait_for_join_completion() causes a PostgreSQL session to pause until all outstanding node join operations area complete.

This example shows the output of a SELECT query from bdr.node_summary. It indicates that two nodes are active and another one is joining.

# SELECT node_name, interface_connstr, peer_state_name,
#     node_seq_id, node_local_dbname
# FROM bdr.node_summary;
-[ RECORD 1 ]-----+-----------------------------------------
node_name         | node1
interface_connstr | host=localhost dbname=postgres port=7432
peer_state_name   | ACTIVE
node_seq_id       | 1
node_local_dbname | postgres
-[ RECORD 2 ]-----+-----------------------------------------
node_name         | node2
interface_connstr | host=localhost dbname=postgres port=7433
peer_state_name   | ACTIVE
node_seq_id       | 2
node_local_dbname | postgres
-[ RECORD 3 ]-----+-----------------------------------------
node_name         | node3
interface_connstr | host=localhost dbname=postgres port=7434
peer_state_name   | JOINING
node_seq_id       | 3
node_local_dbname | postgres

Also, the table bdr.node_catchup_info gives information on the catch-up state, which can be relevant to joining nodes or parting nodes.

When a node is parted, some nodes in the cluster might not receive all the data from that parting node. So parting a node creates a temporary slot from a node that already received that data and can forward it.

The catchup_state can be one of the following:

10 = setup
20 = start
30 = catchup
40 = done

Monitoring Replication Peers

You use two main views for monitoring of replication activity:

You can also obtain most of the information provided by bdr.node_slots by querying the standard PostgreSQL replication monitoring views pg_catalog.pg_stat_replication and pg_catalog.pg_replication_slots.

Each node has one PGD group slot that must never have a connection to it and is very rarely be marked as active. This is normal and doesn't imply something is down or disconnected. See Replication slots in Node Management.

Monitoring outgoing replication

You can use another view for monitoring of outgoing replication activity:

The bdr.node_replication_rates view gives an overall picture of the outgoing replication activity along with the catchup estimates for peer nodes, specifically.

# SELECT * FROM bdr.node_replication_rates;
-[ RECORD 1 ]----+-----------
peer_node_id     | 112898766
target_name      | node1
sent_lsn         | 0/28AF99C8
replay_lsn       | 0/28AF99C8
replay_lag       | 00:00:00
replay_lag_bytes | 0
replay_lag_size  | 0 bytes
apply_rate       | 822
catchup_interval | 00:00:00
-[ RECORD 2 ]----+-----------
peer_node_id     | 312494765
target_name      | node3
sent_lsn         | 0/28AF99C8
replay_lsn       | 0/28AF99C8
replay_lag       | 00:00:00
replay_lag_bytes | 0
replay_lag_size  | 0 bytes
apply_rate       | 853
catchup_interval | 00:00:00

The apply_rate refers to the rate in bytes per second. It's the rate at which the peer is consuming data from the local node. The replay_lag when a node reconnects to the cluster is immediately set to zero. This information will be fixed in a future release. As a workaround, we recommend using the catchup_interval column that refers to the time required for the peer node to catch up to the local node data. The other fields are also available from the bdr.node_slots view.

Administrators can query bdr.node_slots for outgoing replication from the local node. It shows information about replication status of all other nodes in the group that are known to the current node as well as any additional replication slots created by PGD on the current node.

# SELECT node_group_name, target_dbname, target_name, slot_name, active_pid,
#     catalog_xmin, client_addr, sent_lsn, replay_lsn, replay_lag,
#     replay_lag_bytes, replay_lag_size
# FROM bdr.node_slots;
-[ RECORD 1 ]---+----------------------------
node_group_name | bdrgroup
target_dbname   | postgres
target_name     | node3
slot_name       | bdr_postgres_bdrgroup_node3
active_pid      | 15089
catalog_xmin    | 691
client_addr     | 127.0.0.1
sent_lsn        | 0/23F7B70
replay_lsn      | 0/23F7B70
replay_lag      | [NULL]
replay_lag_bytes| 120
replay_lag_size | 120 bytes
-[ RECORD 2 ]---+----------------------------
node_group_name | bdrgroup
target_dbname   | postgres
target_name     | node2
slot_name       | bdr_postgres_bdrgroup_node2
active_pid      | 15031
catalog_xmin    | 691
client_addr     | 127.0.0.1
sent_lsn        | 0/23F7B70
replay_lsn      | 0/23F7B70
replay_lag      | [NULL]
replay_lag_bytes| 84211
replay_lag_size | 82 kB

Because PGD is a mesh network, to get the full view of lag in the cluster, you must execute this query on all nodes participating.

replay_lag_bytes reports the difference in WAL positions between the local server's current WAL write position and replay_lsn, the last position confirmed replayed by the peer node. replay_lag_size is a human-readable form of the same. It's important to understand that WAL usually contains a lot of writes that aren't replicated but still count in replay_lag_bytes, including, for example:

  • VACUUM activity
  • Index changes
  • Writes associated with other databases on the same node
  • Writes for tables that are not part of a replication set

So the lag in bytes reported here isn't the amount of data that must be replicated on the wire to bring the peer node up to date, only the amount of server-side WAL that must be processed.

Similarly, replay_lag isn't a measure of how long the peer node takes to catch up or how long it takes to replay from its current position to the write position at the time bdr.node_slots was queried. It measures the delay between when the peer confirmed the most recent commit and the current wall-clock time. We suggest that you monitor replay_lag_bytes and replay_lag_size or catchup_interval in bdr.node_replication_rates, as this column is set to zero immediately after the node reconnects.

The lag in both bytes and time doesn't advance while logical replication is streaming a transaction. It changes only when a commit is replicated. So the lag tends to "sawtooth," rising as a transaction is streamed and then falling again as the peer node commits it, flushes it, and sends confirmation. The reported LSN positions "stair-step" instead of advancing smoothly, for similar reasons.

When replication is disconnected (active = 'f'), the active_pid column is NULL, as is client_addr and the other fields that make sense only with an active connection. The state field is 'disconnected'. The _lsn fields are the same as the confirmed_flush_lsn, since that's the last position that the client is known for certain to have replayed to and saved. The _lag fields show the elapsed time between the most recent confirmed flush on the client and the current time. The _lag_size and _lag_bytes fields report the distance between confirmed_flush_lsn and the local server's current WAL insert position.

Note

It's normal for restart_lsn to be behind the other lsn columns. This doesn't indicate a problem with replication or a peer node lagging. The restart_lsn is the position that PostgreSQL's internal logical decoding must be reading WAL at if interrupted. It generally reflects the position of the oldest transaction that's not yet replicated and flushed. A very old restart_lsn can make replication slow to restart after disconnection and force retention of more WAL than is desirable, but it's otherwise harmless. If you're concerned, look for very long-running transactions and forgotten prepared transactions.

Monitoring incoming replication

You can monitor incoming replication (also called subscriptions) by querying the bdr.subscription_summary view. This query shows the list of known subscriptions to other nodes in the EDB Postgres Distributed cluster and the state of the replication worker:

# SELECT node_group_name, origin_name, sub_enabled, sub_slot_name,
#     subscription_status
# FROM bdr.subscription_summary;
-[ RECORD 1 ]-------+----------------------------
node_group_name     | bdrgroup
origin_name         | node2
sub_enabled         | t
sub_slot_name       | bdr_postgres_bdrgroup_node1
subscription_status | replicating
-[ RECORD 2 ]-------+----------------------------
node_group_name     | bdrgroup
origin_name         | node3
sub_enabled         | t
sub_slot_name       | bdr_postgres_bdrgroup_node1
subscription_status | replicating

Monitoring WAL senders using LCR

If the decoding worker is enabled, you can monitor information about the current logical change record (LCR) file for each WAL sender using the function bdr.wal_sender_stats(). For example:

postgres=# SELECT * FROM bdr.wal_sender_stats();
   pid   | is_using_lcr |       decoder_slot_name       |              lcr_file_name
---------+--------------+-------------------------------+------------------------------------------
 2059904 | f            |                               |
 2059909 | t            | bdr_postgres_bdrgroup_decoder | 0000000000000000000000140000000000000000
 2059916 | t            | bdr_postgres_bdrgroup_decoder | 0000000000000000000000140000000000000000
(3 rows)

If is_using_lcr is FALSE, decoder_slot_name/lcr_file_name is NULL. This is the case if the decoding worker isn't enabled or the WAL sender is serving a logical standby.

Also, you can monitor information about the decoding worker using the function bdr.get_decoding_worker_stat(). For example:

postgres=# SELECT * FROM bdr.get_decoding_worker_stat();
   pid   | decoded_upto_lsn | waiting | waiting_for_lsn
---------+------------------+---------+-----------------
 1153091 | 0/1E5EEE8        | t       | 0/1E5EF00
(1 row)

Monitoring PGD replication workers

All PGD workers show up in the system view bdr.stat_activity, which has the same columns and information content as pg_stat_activity. So this view offers these insights into the state of a PGD system:

  • The wait_event column has enhanced information, if the reason for waiting is related to PGD.
  • The query column is blank in PGD workers, except when a writer process is executing DDL.

The bdr.workers view shows PGD worker-specific details that aren't available from bdr.stat_activity.

The view bdr.event_summary shows the last error (if any) reported by any worker that has a problem continuing the work. This information is persistent, so it's important to note the time of the error and not just its existence. Most errors are transient, and PGD workers will retry the failed operation.

Monitoring PGD writers

Another system view, bdr.writers, monitors writer activities. This view shows only the current status of writer workers. It includes:

  • sub_name to identify the subscription that the writer belongs to
  • pid of the writer process
  • streaming_allowed to know if the writer supports applying in-progress streaming transactions
  • is_streaming to know if the writer is currently applying a streaming transaction
  • commit_queue_position to check the position of the writer in the commit queue

PGD honors commit ordering by following the same commit order as happened on the origin. In case of parallel writers, multiple writers might apply different transactions at the same time. The commit_queue_position shows the order in which they will commit. Value 0 means that the writer is the first one to commit. Value -1 means that the commit position isn't yet known, which can happen for a streaming transaction or when the writer isn't currently applying any transaction.

Monitoring global locks

The global lock, which is currently used only for DDL replication, is a heavyweight lock that exists across the whole PGD group.

There are currently two types of global locks:

  • DDL lock, used for serializing all DDL operations on permanent (not temporary) objects (that is, tables) in the database
  • DML relation lock, used for locking out writes to relations during DDL operations that change the relation definition

You can create either or both entry types for the same transaction, depending on the type of DDL operation and the value of the bdr.ddl_locking setting.

Global locks held on the local node are visible in the bdr.global_locks view. This view shows the type of the lock. For relation locks, it shows the relation that's being locked, the PID holding the lock (if local), and whether the lock was globally granted. In case of global advisory locks, lock_type column shows GLOBAL_LOCK_ADVISORY, and relation column shows the advisory keys on which the lock is acquired.

This example shows the output of bdr.global_locks while running an ALTER TABLE statement with bdr.ddl_locking = 'all':

# SELECT lock_type, relation, pid FROM bdr.global_locks;
-[ RECORD 1 ]--------------
lock_type | GLOBAL_LOCK_DDL
relation  | [NULL]
pid       | 15534
-[ RECORD 2 ]--------------
lock_type | GLOBAL_LOCK_DML
relation  | someschema.sometable
pid       | 15534

See Catalogs for details on all fields, including lock timing information.

Monitoring conflicts

Replication conflicts can arise when multiple nodes make changes that affect the same rows in ways that can interact with each other. Monitor the PGD system to identify conflicts and, where possible, make application changes to eliminate the conflicts or make them less frequent.

By default, all conflicts are logged to bdr.conflict_history. Since this log contains full details of conflicting data, the rows are protected by row-level security to ensure they're visible only by owners of replicated tables. Owners should expect conflicts and analyze them to see which, if any, might be considered as problems to resolve.

For monitoring purposes, use bdr.conflict_history_summary, which doesn't contain user data. This example shows a query to count the number of conflicts seen in the current day using an efficient query plan:

SELECT count(*)
FROM bdr.conflict_history_summary
WHERE local_time > date_trunc('day', current_timestamp)
  AND local_time < date_trunc('day', current_timestamp + '1 day');

Apply statistics

PGD collects statistics about replication apply, both for each subscription and for each table.

Two monitoring views exist: bdr.stat_subscription for subscription statistics and bdr.stat_relation for relation statistics. These views both provide:

  • Number of INSERTs/UPDATEs/DELETEs/TRUNCATEs replicated
  • Block accesses and cache hit ratio
  • Total I/O time for read/write
  • Number of in-progress transactions streamed to file
  • Number of in-progress transactions streamed to writers
  • Number of in-progress streamed transactions committed/aborted

For relations only, bdr.stat_relation also includes:

  • Total time spent processing replication for the relation
  • Total lock wait time to acquire lock (if any) for the relation (only)

For subscriptions only, bdr.stat_subscription includes:

  • Number of COMMITs/DDL replicated for the subscription
  • Number of times this subscription has connected upstream

Tracking of these statistics is controlled by the PGD GUCs bdr.track_subscription_apply and bdr.track_relation_apply, respectively.

The following shows the example output from these:

# SELECT sub_name, nconnect, ninsert, ncommit, nupdate, ndelete, ntruncate, nddl
FROM bdr.stat_subscription;
-[ RECORD 1 ]----------------------------------
sub_name  | bdr_regression_bdrgroup_node1_node2
nconnect  | 3
ninsert   | 10
ncommit   | 5
nupdate   | 0
ndelete   | 0
ntruncate | 0
nddl      | 2

In this case, the subscription connected three times to the upstream, inserted 10 rows, and performed two DDL commands inside five transactions.

You can reset the stats counters for these views to zero using the functions bdr.reset_subscription_stats and bdr.reset_relation_stats.

Standard PostgreSQL statistics views

Statistics on table and index usage are normally updated by the downstream master. This is essential for the correct function of autovacuum. If there are no local writes on the downstream master and statistics haven't been reset, these two views show corresponding results between upstream and downstream:

  • pg_stat_user_tables
  • pg_statio_user_tables
Note

We don't necessarily expect the upstream table statistics to be similar to the downstream ones. We only expect them to change by the same amounts. Consider the example of a table whose statistics show 1M inserts and 1M updates. When a new node joins the PGD group, the statistics for the same table in the new node show 1M inserts and zero updates. However, from that moment, the upstream and downstream table statistics change by the same amounts because all changes on one side are replicated to the other side.

Since indexes are used to apply changes, the identifying indexes on the downstream side might appear more heavily used with workloads that perform UPDATE and DELETE than non-identifying indexes are.

The built-in index monitoring views are:

  • pg_stat_user_indexes
  • pg_statio_user_indexes

All these views are discussed in detail in the PostgreSQL documentation on the statistics views.

Monitoring PGD versions

PGD allows running different Postgres versions as well as different BDR extension versions across the nodes in the same cluster. This capability is useful for upgrading.

The view bdr.group_versions_details uses the function bdr.run_on_all_nodes() to retrieve Postgres and BDR extension versions from all nodes at the same time. For example:

bdrdb=# SELECT node_name, postgres_version, bdr_version
        FROM bdr.group_versions_details;
 node_name | postgres_version | bdr_version
-----------+------------------+-------------
 node1     | 15.2.0           | 5.0.0
 node2     | 15.2.0           | 5.0.0

The recommended setup is to try to have all nodes running the same (and latest) versions as soon as possible. We recommend that the cluster doesn't run different versions of the BDR extension for too long.

For monitoring purposes, we recommend the following alert levels:

  • status=UNKNOWN, message=This node is not part of any PGD group
  • status=OK, message=All nodes are running same PGD versions
  • status=WARNING, message=There is at least 1 node that is not accessible
  • status=WARNING, message=There are node(s) running different PGD versions when compared to other nodes

The described behavior is implemented in the function bdr.monitor_group_versions(), which uses PGD version information returned from the view bdr.group_version_details to provide a cluster-wide version check. For example:

bdrdb=# SELECT * FROM bdr.monitor_group_versions();
 status |                message
--------+-----------------------------------------
 OK     | All nodes are running same BDR versions

Monitoring Raft consensus

Raft consensus must be working cluster-wide at all times. The impact of running a EDB Postgres Distributed cluster without Raft consensus working might be as follows:

  • The replication of PGD data changes might still work correctly.
  • Global DDL/DML locks doesn't work.
  • Galloc sequences eventually run out of chunks.
  • Eager Replication doesn't work.
  • Cluster maintenance operations (join node, part node, promote standby) are still allowed, but they might not finish (hanging instead).
  • Node statuses might not be correctly synced among the PGD nodes.
  • PGD group replication slot doesn't advance LSN and thus keeps WAL files on disk.

The view bdr.group_raft_details uses the functions bdr.run_on_all_nodes() and bdr.get_raft_status() to retrieve Raft consensus status from all nodes at the same time. For example:

bdrdb=# SELECT node_id, node_name, state, leader_id
FROM bdr.group_raft_details;
  node_id   | node_name | node_group_name |     state     | leader_id
------------+-----------+-----------------+---------------+------------
 1148549230 | node1     | top_group       | RAFT_LEADER   | 1148549230
 3367056606 | node2     | top_group       | RAFT_FOLLOWER | 1148549230

Raft consensus is working correctly if all of these conditions are met:

  • A valid state (RAFT_LEADER or RAFT_FOLLOWER) is defined on all nodes.
  • Only one of the nodes is the RAFT_LEADER.
  • The leader_id is the same on all rows and must match the node_id of the row where state = RAFT_LEADER.

From time to time, Raft consensus starts a new election to define a new RAFT_LEADER. During an election, there might be an intermediary situation where there's no RAFT_LEADER, and some of the nodes consider themselves as RAFT_CANDIDATE. The whole election can't take longer than bdr.raft_election_timeout (by default it's set to 6 seconds). If the query above returns an in-election situation, then wait for bdr.raft_election_timeout, and run the query again. If after bdr.raft_election_timeout has passed and some the listed conditions are still not met, then Raft consensus isn't working.

Raft consensus might not be working correctly on only a single node. For example, one of the nodes doesn't recognize the current leader and considers itself as a RAFT_CANDIDATE. In this case, it's important to make sure that:

  • All PGD nodes are accessible to each other through both regular and replication connections (check file pg_hba.conf).
  • PGD versions are the same on all nodes.
  • bdr.raft_election_timeout is the same on all nodes.

In some cases, especially if nodes are geographically distant from each other or network latency is high, the default value of bdr.raft_election_timeout (6 seconds) might not be enough. If Raft consensus is still not working even after making sure everything is correct, consider increasing bdr.raft_election_timeout to 30 seconds on all nodes. For PGD 3.6.11 and later, setting bdr.raft_election_timeout requires only a server reload.

Given how Raft consensus affects cluster operational tasks, and also as Raft consensus is directly responsible for advancing the group slot, monitoring alert levels are defined as follows:

  • status=UNKNOWN, message=This node is not part of any PGD group
  • status=OK, message=Raft Consensus is working correctly
  • status=WARNING, message=There is at least 1 node that is not accessible
  • status=WARNING, message=There are node(s) as RAFT_CANDIDATE, an election might be in progress
  • status=WARNING, message=There is no RAFT_LEADER, an election might be in progress
  • status=CRITICAL, message=There is a single node in Raft Consensus
  • status=CRITICAL, message=There are node(s) as RAFT_CANDIDATE while a RAFT_LEADER is defined
  • status=CRITICAL, message=There are node(s) following a leader different than the node set as RAFT_LEADER

The described behavior is implemented in the function bdr.monitor_group_raft(), which uses Raft consensus status information returned from the view bdr.group_raft_details to provide a cluster-wide Raft check. For example:

bdrdb=# SELECT * FROM bdr.monitor_group_raft();
node_group_name | status |               message
----------------|--------+-------------------------------------
myroup          | OK     | Raft Consensus is working correctly

Monitoring replication slots

Each PGD node keeps:

  • One replication slot per active PGD peer
  • One group replication slot

For example:

bdrdb=# SELECT slot_name, database, active, confirmed_flush_lsn
FROM pg_replication_slots ORDER BY slot_name;
        slot_name         | database | active | confirmed_flush_lsn
--------------------------+----------+--------+---------------------
 bdr_bdrdb_bdrgroup       | bdrdb    | f      | 0/3110A08
 bdr_bdrdb_bdrgroup_node2 | bdrdb    | t      | 0/31F4670
 bdr_bdrdb_bdrgroup_node3 | bdrdb    | t      | 0/31F4670
 bdr_bdrdb_bdrgroup_node4 | bdrdb    | t      | 0/31F4670

Peer slot names follow the convention bdr_<DATABASE>_<GROUP>_<PEER>, while the PGD group slot name follows the convention bdr_<DATABASE>_<GROUP>. You can access the group slot using the function bdr.local_group_slot_name().

Peer replication slots must be active on all nodes at all times. If a peer replication slot isn't active, then it might mean either:

  • The corresponding peer is shut down or not accessible.
  • PGD replication is broken.

Grep the log file for ERROR or FATAL, and also check bdr.event_summary on all nodes. The root cause might be, for example, an incompatible DDL was executed with DDL replication disabled on one of the nodes.

The PGD group replication slot is, however, inactive most of the time. PGD maintains this slot and advances its LSN when all other peers already consumed the corresponding transactions. Consequently, it's not necessary to monitor the status of the group slot.

The function bdr.monitor_local_replslots() provides a summary of whether all PGD node replication slots are working as expected. For example:

bdrdb=# SELECT * FROM bdr.monitor_local_replslots();
 status |                    message
--------+-------------------------------------------------
 OK     | All BDR replication slots are working correctly

One of the following status summaries is returned:

  • UNKNOWN: This node is not part of any BDR group
  • OK: All BDR replication slots are working correctly
  • OK: This node is part of a subscriber-only group
  • CRITICAL: There is at least 1 BDR replication slot which is inactive
  • CRITICAL: There is at least 1 BDR replication slot which is missing

Monitoring transaction COMMITs

By default, PGD transactions are committed only to the local node. In that case, a transaction's COMMIT is processed quickly.

PGD's Commit Scopes feature offers a range of synchronous transaction commit scopes that allow you to balance durability, consistency, and performance for your particular queries. You can monitor these transactions by examining the bdr.stat_activity catalog. The processes report different wait_event states as a transaction is committed. This monitoring only covers transactions in progress and doesn't provide historical timing information.