PGD settings v5.6

You can set PGD-specific configuration settings. Unless noted otherwise, you can set the values at any time.

Conflict handling

bdr.default_conflict_detection

Sets the default conflict detection method for newly created tables. Accepts same values as bdr.alter_table_conflict_detection().

Global sequence parameters

bdr.default_sequence_kind

Sets the default sequence kind.

The default is distributed, which means snowflakeid is used for int8 sequences (that is, bigserial) and galloc sequence for int4 (that is, serial) and int2 sequences.

DDL handling

bdr.default_replica_identity

Sets the default value for REPLICA IDENTITY on newly created tables. The REPLICA IDENTITY defines the information written to the write-ahead log to identify rows that are updated or deleted.

The accepted values are:

See the PostgreSQL documentation for more details.

PGD can't replicate UPDATE and DELETE operations on tables without a PRIMARY KEY or UNIQUE constraint. The exception is when the replica identity for the table is FULL, either by table-specific configuration or by bdr.default_replica_identity.

If bdr.default_replica_identity is default and there is a UNIQUE constraint included in the table definition, it won't be automatically picked up as REPLICA IDENTITY. You need to set the REPLICA IDENTITY explicitly using ALTER TABLE ... REPLICA IDENTITY ....

Setting the replica identity of tables to full increases the volume of WAL written and the amount of data replicated on the wire for the table.

ALTER SYSTEM SET bdr.default_replica_identity="default";

bdr.ddl_replication

Automatically replicates DDL across nodes (default is on).

This parameter can be set only by bdr_superuser or superuser roles.

Running DDL or calling PGD administration functions with bdr.ddl_replication = off can create situations where replication stops until an administrator can intervene. See DDL replication for details.

A LOG-level log message is emitted to the PostgreSQL server logs whenever bdr.ddl_replication is set to off. Additionally, a WARNING-level message is written whenever replication of captured DDL commands or PGD replication functions is skipped due to this setting.

bdr.role_replication

Automatically replicates ROLE commands across nodes (default is on). Only a superuser can set this parameter. This setting works only if bdr.ddl_replication is turned on as well.

Turning this parameter off without using external methods to ensure roles are in sync across all nodes might cause replicated DDL to interrupt replication until the administrator intervenes.

See Role manipulation statements for details.

bdr.ddl_locking

Configures the operation mode of global locking for DDL.

This parameter can be set only by bdr_superuser or superuser roles.

Possible options are:

Default is all.

A LOG-level log message is emitted to the PostgreSQL server logs whenever bdr.ddl_replication is set to off. Additionally, a WARNING message is written whenever any global locking steps are skipped due to this setting. It's normal for some statements to result in two WARNING messages: one for skipping the DML lock and one for skipping the DDL lock.

For backward compatibility, bdr.ddl_locking supports aliases. on and true are an alias for all. false is an alias for off.

See also Global locking.

bdr.truncate_locking

Sets the TRUNCATE command's locking behavior (default is off). When true, TRUNCATE obeys the bdr.ddl_locking setting.

Global locking

DDL locking is controlled by bdr.ddl_locking. Other global locking settings include the following.

bdr.global_lock_max_locks

Sets the maximum number of global locks that can be held on a node (default is 1000). Can be set only at Postgres server start.

bdr.global_lock_timeout

Sets the maximum allowed duration of any wait for a global lock (default is 10 minutes). A value of zero disables this timeout.

bdr.global_lock_statement_timeout

Sets the maximum allowed duration of any statement holding a global lock (default is 60 minutes). A value of zero disables this timeout.

bdr.global_lock_idle_timeout

Sets the maximum allowed duration of idle time in a transaction holding a global lock (default is 10 minutes). A value of zero disables this timeout.

bdr.lock_table_locking

Sets locking behavior for LOCK TABLE statement (default is off). When enabled, LOCK TABLE statement also takes a global DML lock on the cluster, blocking other locking statements.

bdr.predictive_checks

Sets the log level for predictive checks (currently used only by global locks). Can be DEBUG, LOG, WARNING (default), or ERROR. Predictive checks are early validations for expected cluster state when doing certain operations. You can use them for those operations for fail early rather than wait for timeouts. In global lock terms, PGD checks that there are enough nodes connected and withing reasonable lag limit for getting the quorum needed by the global lock.

Node management

bdr.replay_progress_frequency

Sets the interval for sending replication position info to the rest of the cluster (default is 1 minute).

bdr.standby_slot_names

Sets the slots required to receive and confirm replication changes before any other ones. This setting is useful primarily when using physical standbys for failover or when using subscribe-only nodes.

Generic replication

bdr.writers_per_subscription

Sets the default number of writers per subscription. (In PGD, you can also change this with bdr.alter_node_group_config for a group.)

bdr.max_writers_per_subscription

Maximum number of writers per subscription (sets upper limit for the bdr.writers_per_subscription setting).

bdr.xact_replication

Replicates current transaction (default is on).

Turning this off makes the whole transaction local only, which means the transaction isn't visible to logical decoding by PGD and all other downstream targets of logical decoding. Data isn't transferred to any other node, including logical standby nodes.

This parameter can be set only by the bdr_superuser or superuser roles.

This parameter can be set only inside the current transaction using the SET LOCAL command unless bdr.permit_unsafe_commands = on.

bdr.permit_unsafe_commands

Overrides safety check on commands that are deemed unsafe for general use.

Requires bdr_superuser or PostgreSQL superuser.

bdr.batch_inserts

Number of consecutive inserts to one table in a single transaction that turns on batch processing of inserts for that table.

This setting allows replication of large data loads as COPY internally, rather than as a set of inserts. It's also how the initial data during node join is copied.

bdr.maximum_clock_skew

Specifies the maximum difference between the incoming transaction commit timestamp and the current time on the subscriber before triggering bdr.maximum_clock_skew_action.

It checks if the timestamp of the currently replayed transaction is in the future compared to the current time on the subscriber. If it is, and the difference is larger than bdr.maximum_clock_skew, it performs the action specified by the bdr.maximum_clock_skew_action setting.

The default is -1, which means ignore clock skew (the check is turned off). It's valid to set 0 as when the clocks on all servers are synchronized. The fact that the transaction is being replayed means it was committed in the past.

bdr.maximum_clock_skew_action

Specifies the action to take if a clock skew higher than bdr.maximum_clock_skew is detected.

There are two possible values for this setting:

bdr.accept_connections

Enables or disables connections to PGD (default is on).

Requires bdr_superuser or PostgreSQL superuser.

bdr.standby_slot_names

This setting is typically used in failover configurations to ensure that the failover-candidates streaming physical replicas for this PGD node have received and flushed all changes before they ever become visible to subscribers. That guarantees that a commit can't vanish on failover to a standby for the provider.

Replication slots whose names are listed in the comma-separated bdr.standby_slot_names list are treated specially by the WAL sender on a PGD node.

PGD's logical replication WAL senders ensure that all local changes are sent and flushed to the replication slots in bdr.standby_slot_names before the node sends those changes to any other PGD replication clients. Effectively, it provides a synchronous replication barrier between the named list of slots and all other replication clients.

Any replication slot can be listed in bdr.standby_slot_names. Both logical and physical slots work, but it's generally used for physical slots.

Without this safeguard, two anomalies are possible where a commit can be received by a subscriber and then vanish from the provider on failover because the failover candidate hadn't received it yet:

• For 1+ subscribers, the subscriber might have applied the change but the new provider might execute new transactions that conflict with the received change, as it never happened as far as the provider is concerned.

• For 2+ subscribers, at the time of failover, not all subscribers have applied the change. The subscribers now have inconsistent and irreconcilable states because the subscribers that didn't receive the commit have no way to get it.

Setting bdr.standby_slot_names by design causes other subscribers not listed in there to lag behind the provider if the required number of listed nodes are not keeping up. Monitoring is thus essential.

Another use case where bdr.standby_slot_names is useful is when using a subscriber-only node, to ensure that it doesn't move ahead of any of the regular PGD nodes. This can best be achieved by listing the logical slots of all regular PGD peer nodes in combination with setting bdr.standby_slots_min_confirmed to at least 1.

bdr.standby_slots_min_confirmed

Controls how many of the bdr.standby_slot_names have to confirm before sending data to PGD subscribers.

bdr.writer_input_queue_size

Specifies the size of the shared memory queue used by the receiver to send data to the writer process. If the writer process is stalled or making slow progress, then the queue might get filled up, stalling the receiver process too. So it's important to provide enough shared memory for this queue. The default is 1 MB, and the maximum allowed size is 1 GB. While any storage size specifier can be used to set the GUC, the default is KB.

bdr.writer_output_queue_size

Specifies the size of the shared memory queue used by the receiver to receive data from the writer process. Since the writer isn't expected to send a large amount of data, a relatively smaller sized queue is enough. The default is 32 KB, and the maximum allowed size is 1 MB. While any storage size specifier can be used to set the GUC, the default is KB.

bdr.min_worker_backoff_delay

Allows for rate limiting of PGD background worker launches by preventing a given worker from being relaunched more often than every bdr.min_worker_backoff_delay milliseconds. On repeated errors, the backoff increases exponentially with added jitter up to a maximum of bdr.max_worker_backoff_delay.

Time-unit suffixes are supported.

The default for bdr.min_worker_backoff_delay is 1 second. For bdr.max_worker_backoff_delay, it's 1 minute.

If the backoff delay setting is changed and the PostgreSQL configuration is reloaded, then all current backoffs wait for reset. Additionally, the bdr.worker_task_reset_backoff_all() function is provided to allow the administrator to force all backoff intervals to immediately expire.

A tracking table in shared memory is maintained to remember the last launch time of each type of worker. This tracking table isn't persistent. It's cleared by PostgreSQL restarts, including soft restarts during crash recovery after an unclean backend exit.

You can use the view bdr.worker_tasks to inspect this state so the administrator can see any backoff rate limiting currently in effect.

For rate-limiting purposes, workers are classified by task. This key consists of the worker role, database OID, subscription ID, subscription writer ID, extension library name and function name, extension-supplied worker name, and the remote relation ID for sync writers. NULL is used where a given classifier doesn't apply, for example, when manager workers don't have a subscription ID and receivers don't have a writer ID.

CRDTs

bdr.crdt_raw_value

Sets the output format of CRDT data types.

The default output (when this setting is off) is to return only the current value of the base CRDT type, for example, a bigint for crdt_pncounter. When set to on, the returned value represents the full representation of the CRDT value, which can, for example, include the state from multiple nodes.

Commit scope

bdr.commit_scope

Sets the current (or default) commit scope (default is an empty string).

Commit At Most Once

bdr.camo_local_mode_delay

The commit delay that applies in CAMO's asynchronous mode to emulate the overhead that normally occurs with the CAMO partner having to confirm transactions (default is 5 ms). Set to 0 to disable this feature.

bdr.camo_enable_client_warnings

Emits warnings if an activity is carried out in the database for which CAMO properties can't be guaranteed (default is enabled). Well-informed users can choose to disable this setting to reduce the amount of warnings going into their logs.

Transaction streaming

bdr.default_streaming_mode

Controls transaction streaming by the subscriber node. Possible values are: off, writer, file, and auto. Defaults to auto. If set to off, the subscriber doesn't request transaction streaming. If set to one of the other values, the subscriber requests transaction streaming and the publisher provides it if it supports them and if configured at group level. For more details, see Transaction streaming.

Lag Control

bdr.lag_control_max_commit_delay

Maximum acceptable post-commit delay that can be tolerated, in fractional milliseconds.

bdr.lag_control_max_lag_size

Maximum acceptable lag size that can be tolerated, in kilobytes.

bdr.lag_control_max_lag_time

Maximum acceptable lag time that can be tolerated, in milliseconds.

bdr.lag_control_min_conforming_nodes

Minimum number of nodes required to stay below acceptable lag measures.

bdr.lag_control_commit_delay_adjust

Commit delay micro adjustment measured as a fraction of the maximum commit delay time. At a default value of 0.01%, it takes 100 net increments to reach the maximum commit delay.

bdr.lag_control_sample_interval

Minimum time between lag samples and commit delay micro adjustments, in milliseconds.

bdr.lag_control_commit_delay_start

The lag threshold at which commit delay increments start to be applied, expressed as a fraction of acceptable lag measures. At a default value of 1.0%, commit delay increments don't begin until acceptable lag measures are breached.

By setting a smaller fraction, it might be possible to prevent a breach by "bending the lag curve" earlier so that it's asymptotic with the acceptable lag measure.

Timestamp-based snapshots

bdr.timestamp_snapshot_keep

Time to keep valid snapshots for the timestamp-based snapshot use (default is 0, meaning don't keep past snapshots).

Monitoring and logging

bdr.debug_level

Defines the log level that PGD uses to write its debug messages. The default value is debug2. If you want to see detailed PGD debug output, set bdr.debug_level = 'log'.

bdr.trace_level

Similar to bdr.debug_level, defines the log level to use for PGD trace messages. Enabling tracing on all nodes of an EDB Postgres Distributed cluster might help EDB Support to diagnose issues. You can set this parameter only at Postgres server start.

bdr.track_subscription_apply

Tracks apply statistics for each subscription with bdr.stat_subscription (default is on).

bdr.track_relation_apply

Tracks apply statistics for each relation with bdr.stat_relation (default is off).

bdr.track_apply_lock_timing

Tracks lock timing when tracking statistics for relations with bdr.stat_relation (default is off).

Decoding worker

bdr.enable_wal_decoder

Enables logical change record (LCR) sending on a single node with a decoding worker (default is false). When set to true, a decoding worker process starts, and WAL senders send the LCRs it produces. If set back to false, any WAL senders using LCR are restarted and use the WAL directly.

bdr.receive_lcr

When subscribing to another node, this setting enables the node to request the use of logical change records (LCRs) for the subscription (default is false). When this setting is true on a downstream node, the node requests that upstream nodes use LCRs when sending to it. If you set bdr.enable_wal_decoder to true on a node, also set this setting to true.

bdr.lcr_cleanup_interval

Logical change record (LCR) file cleanup interval (default is 3 minutes). When the decoding worker is enabled, the decoding worker stores LCR files as a buffer. These files are periodically cleaned, and this setting controls the interval between any two consecutive cleanups. Setting it to zero disables cleanup.

Connectivity settings

The following are a set of connectivity settings affecting all cross-node libpq connections. The defaults are set to fairly conservative values and cover most production needs. All variables have SIGHUP context, meaning changes are applied upon reload.

bdr.global_connection_timeout

Maximum time to wait while connecting, in seconds (default is 15 seconds). Write as a decimal integer, for example, 10. Zero, negative, or not specified means wait indefinitely. The minimum allowed timeout is 2 seconds, therefore a value of 1 is interpreted as 2.

bdr.global_keepalives

Controls whether TCP keepalives are used (default is 1, meaning on). If you don't want keepalives, you can change this to 0, meaning off. This parameter is ignored for connections made by a Unix-domain socket.

bdr.global_keepalives_idle

Controls the number of seconds of inactivity after which TCP sends a keepalive message to the server (default is 1 second). A value of zero uses the system default. This parameter is ignored for connections made by a Unix-domain socket or if keepalives are disabled. It's supported only on systems where TCP_KEEPIDLE or an equivalent socket option is available. On other systems, it has no effect.

bdr.global_keepalives_interval

Controls the number of seconds after which to retransmit a TCP keepalive message that isn't acknowledged by the server (default is 2 seconds). A value of zero uses the system default. This parameter is ignored for connections made by a Unix-domain socket or if keepalives are disabled. It's supported only on systems where TCP_KEEPINTVL or an equivalent socket option is available. On other systems, it has no effect.

bdr.global_keepalives_count

Controls the number of TCP keepalives that can be lost before the client's connection to the server is considered dead (default is 3). A value of zero uses the system default. This parameter is ignored for connections made by a Unix-domain socket or if keepalives are disabled. It's supported only on systems where TCP_KEEPCNT or an equivalent socket option is available. On other systems, it has no effect.

bdr.global_tcp_user_timeout

Controls the number of milliseconds that transmitted data can remain unacknowledged before a connection is forcibly closed (default is 5000, that is, 5 seconds). A value of zero uses the system default. This parameter is ignored for connections made by a Unix-domain socket. It's supported only on systems where TCP_USER_TIMEOUT is available. On other systems, it has no effect.

Topology settings

bdr.force_full_mesh

Forces the full mesh topology (default is on). When set to off, PGD will attempt to use the optimized topology for subscriber-only groups. This setting is only effective when the requirements for the optimized topology are met. See Optimizing subscriber-only groups for more information.

Internal settings - Raft timeouts

bdr.raft_global_election_timeout

To account for network failures, the Raft consensus protocol implements timeouts for elections and requests. This value is used when a request is being sent to the global (top-level) group. The default is 6 seconds (6s).

bdr.raft_group_election_timeout

To account for network failures, the Raft consensus protocol implements timeouts for elections and requests. This value is used when a request is being sent to the sub-group. The default is 3 seconds (3s).

bdr.raft_response_timeout

For responses, the settings of bdr.raft_global_election_timeout and bdr.raft_group_election_timeout are used as appropriate. You can override this behavior by setting this variable. The setting of bdr.raft_response_timeout must be less than either of the election timeout values. Set this variable to -1 to disable the override. The default is -1.

Internal settings - Other Raft values

bdr.raft_keep_min_entries

The minimum number of entries to keep in the Raft log when doing log compaction (default is 1000; PGD 5.3 and earlier: 100). The value of 0 disables log compaction. You can set this parameter only at Postgres server start.

bdr.raft_log_min_apply_duration

To move the state machine forward, Raft appends entries to its internal log. During normal operation, appending takes only a few milliseconds. This poses an upper threshold on the duration of that append action, above which an INFO message is logged. This can indicate a problem. Default is 3000 ms.

bdr.raft_log_min_message_duration

When to log a consensus request. Measures roundtrip time of a PGD consensus request and logs an INFO message if the time exceeds this parameter (default is 5000 ms).

bdr.raft_group_max_connections

The maximum number of connections across all PGD groups for a Postgres server (default is 100 connections). These connections carry PGD consensus requests between the groups' nodes. You can set this parameter only at Postgres server start.

Internal settings - Other values

bdr.backwards_compatibility

Specifies the version to be backward compatible to, in the same numerical format as used by bdr.bdr_version_num, for example, 30618. (Default is the current PGD version.) Enables exact behavior of a former PGD version, even if this has generally unwanted effects. Since this changes from release to release, we advise against explicit use in the configuration file unless the value is different from the current version.

bdr.track_replication_estimates

Tracks replication estimates in terms of apply rates and catchup intervals for peer nodes. Protocols like CAMO can use this information to estimate the readiness of a peer node. This parameter is enabled by default.

bdr.lag_tracker_apply_rate_weight

PGD monitors how far behind peer nodes are in terms of applying WAL from the local node and calculate a moving average of the apply rates for the lag tracking. This parameter specifies how much contribution newer calculated values have in this moving average calculation. Default is 0.1.

bdr.enable_auto_sync_reconcile

When enabled, nodes perform automatic synchronization of data from a node that is furthest ahead with respect to the down node. Default (from 5.5.1) is off.