Durability and performance options v4

Overview

Synchronous or Eager Replication synchronizes between at least two nodes of the cluster before committing a transaction. This synchronization provides three properties of interest to applications that are related but can all be implemented individually:

  • Durability: Writing to multiple nodes increases crash resilience and allows you to recover the data after a crash and restart.
  • Visibility: With the commit confirmation to the client, the database guarantees immediate visibility of the committed transaction on some sets of nodes.
  • No conflicts after commit: The client can rely on the transaction to eventually be applied on all nodes without further conflicts or get an abort directly informing the client of an error.

BDR provides a Group Commit feature to guarantee durability and visibility by providing a variant of synchronous replication. This feature is similar to the Postgres synchronous_commit feature for physical standbys but provides more flexibility for large-scale distributed systems.

In addition to Group Commit, BDR also offers two other modes that can't currently be combined with Group Commit:

  • Commit At Most Once (CAMO). This feature solves the problem with knowing whether your transaction has committed (and replicated) in case of node or network failures COMMIT. Normally, it might be hard to know whether the COMMIT was processed in. With this feature, your application can find out what happened, even if your new database connection is to a node different from your previous connection. For more information about this feature, see Commit At Most Once.
  • Eager Replication. This is an optional feature to check for conflicts prior to the commit. Every transaction is applied and prepared on all nodes simultaneously and commits only if no replication conflicts are detected. This feature reduces performance but provides strong consistency guarantees. For more information about this feature, see Eager All-Node Replication.

Postgres provides Physical Streaming Replication (PSR), which is unidirectional but offers a synchronous variant. For backward compatibility, BDR still supports configuring synchronous replication with synchronous_commit and synchronous_standby_names. See Legacy synchronous replication, but the use of Group Commit is recommended instead in all cases.

Terms and definitions

BDR nodes can take different roles. These are implicitly assigned per transaction and are unrelated even for concurrent transactions.

  • The origin is the node that receives the transaction from the client or application. It's the node processing the transaction first, initiating replication to other BDR nodes, and responding back to the client with a confirmation or an error.

  • A partner node is a BDR node expected to confirm transactions either according to Group Commit or CAMO requirements.

  • A commit group is the group of all BDR nodes involved in the commit, that is, the origin and all of its partner nodes, which can be just a few or all peer nodes.

Comparison

Most options for synchronous replication available to BDR allow for different levels of synchronization, offering different tradeoffs between performance and protection against node or network outages.

The following table summarizes what a client can expect from a peer node replicated to after having received a COMMIT confirmation from the origin node the transaction was issued to. The Mode column takes on different meaning depending on the variant. For PSR and legacy synchronous replication with BDR, it refers to the synchronous_commit setting. For CAMO, it refers to the bdr.enable_camo setting. And for Group Commit, it refers to the confirmation requirements of the commit scope configuration.

VariantModeReceivedVisibleDurable
async BDRoff (default)nonono
PSRremote_write (2)yesnono (1)
PSRon (2)yesnoyes
PSRremote_apply (2)yesyesyes
Group Commit'ON received' nodesyesnono
Group Commit'ON replicated' nodesyesnono
Group Commit'ON durable' nodesyesnoyes
Group Commit'ON visible' nodesyesyesyes
CAMOremote_write (2)yesnono
CAMOremote_commit_async (2)yesyesno
CAMOremote_commit_flush (2)yesyesyes
Eagern/ayesyesyes
legacy (3)remote_write (2)yesnono
legacy (3)on (2)yesyesyes
legacy (3)remote_apply (2)yesyesyes

(1) Written to the OS, durable if the OS remains running and only Postgres crashes.

(2) Unless switched to local mode (if allowed) by setting synchronous_replication_availability to async', otherwise the values for the asynchronous BDR default apply.

(3) Not recommended. Consider using Group Commit instead.

Reception ensures the peer operating normally can eventually apply the transaction without requiring any further communication, even in the face of a full or partial network outage. A crash of a peer node might still require retransmission of the transaction, as this confirmation doesn't involve persistent storage. All modes considered synchronous provide this protection.

Visibility implies the transaction was applied remotely. All other clients see the results of the transaction on all nodes, providing this guarantee immediately after the commit is confirmed by the origin node. Without visibility, other clients connected might not see the results of the transaction and experience stale reads.

Durability relates to the peer node's storage and provides protection against loss of data after a crash and recovery of the peer node. This can either relate to the reception of the data (as with physical streaming replication) or to visibility (as with Group Commit, CAMO, and Eager). The former eliminates the need for retransmissions after a crash, while the latter ensures visibility is maintained across restarts.

Internal timing of operations

For a better understanding of how the different modes work, it's helpful to realize PSR and BDR apply transactions differently.

With physical streaming replication, the order of operations is:

  • Origin flushes a commit record to WAL, making the transaction visible locally.
  • Peer node receives changes and issues a write.
  • Peer flushes the received changes to disk.
  • Peer applies changes, making the transaction visible locally.

With BDR, the order of operations is different:

  • Origin flushes a commit record to WAL, making the transaction visible locally.
  • Peer node receives changes into its apply queue in memory.
  • Peer applies changes, making the transaction visible locally.
  • Peer persists the transaction by flushing to disk.

For Group Commit, CAMO, and Eager, the origin node waits for a certain number of confirmations prior to making the transaction visible locally. The order of operations is:

  • Origin flushes a prepare or precommit record to WAL.
  • Peer node receives changes into its apply queue in memory.
  • Peer applies changes, making the transaction visible locally.
  • Peer persists the transaction by flushing to disk.
  • Origin commits and makes the transaction visible locally.

The following table summarizes the differences.

VariantOrder of apply vs persistReplication before or after commit
PSRpersist firstafter WAL flush of commit record
BDRapply firstafter WAL flush of commit record
Group Commitapply firstbefore COMMIT on origin
CAMOapply firstbefore COMMIT on origin
Eagerapply firstbefore COMMIT on origin

Configuration

The following table provides an overview of the configuration settings that are required to be set to a nondefault value (req) or optional (opt) but affecting a specific variant.

setting (GUC)Group CommitCAMOEagerPSR (1)
synchronous_standby_namesn/an/an/areq
synchronous_commitn/an/an/aopt
synchronous_replication_availabilityn/aoptn/aopt
bdr.enable_camon/areqn/an/a
bdr.commit_scopereqn/aoptn/a
bdr.global_commit_timeoutoptoptoptn/a

(1) values in this column apply also to synchronous_commit and synchronous_standby_names being used in combination with BDR.

Planned shutdown and restarts

When using Group Commit with receive confirmations or CAMO in combination with remote_write, take care with planned shutdown or restart. By default, the apply queue is consumed prior to shutting down. However, in the immediate shutdown mode, the queue is discarded at shutdown, leading to the stopped node "forgetting" transactions in the queue. A concurrent failure of the origin node can lead to loss of data, as if both nodes failed.

To ensure the apply queue gets flushed to disk, use either smart or fast shutdown for maintenance tasks. This approach maintains the required synchronization level and prevents loss of data.

Legacy synchronous replication using BDR

Note

We don't recommend this approach. Consider using Group Commit instead.

Usage

To enable synchronous replication using BDR, you need to add the application name of the relevant BDR peer nodes to synchronous_standby_names. The use of FIRST x or ANY x offers a some flexibility if this doesn't conflict with the requirements of non-BDR standby nodes.

Once you've added it, you can configure the level of synchronization per transaction using synchronous_commit, which defaults to on. This setting means that adding to synchronous_standby_names already enables synchronous replication. Setting synchronous_commit to local or off turns off synchronous replication.

Due to BDR applying the transaction before persisting it, the values on and remote_apply are equivalent (for logical replication).

Migration to Group Commit

The Group Commit feature of BDR is configured independent of synchronous_commit and synchronous_standby_names. Instead, the bdr.commit_scope GUC allows you to select the scope per transaction. And instead of synchronous_standby_names configured on each node individually, Group Commit uses globally synchronized Commit Scopes.

Note

While the grammar for synchronous_standby_names and Commit Scopes looks similar, the former doesn't account for the origin node, but the latter does. Therefore, for example, synchronous_standby_names = 'ANY 1 (..)' is equivalent to a Commit Scope of ANY 2 (...). This choice makes reasoning about majority easier and reflects that the origin node also contributes to the durability and visibility of the transaction.