PGD overview v5
EDB Postgres Distributed (PGD) provides multi-master replication and data distribution with advanced conflict management, data-loss protection, and throughput up to 5X faster than native logical replication. It also enables distributed Postgres clusters with high availability up to five 9s.
PGD provides loosely coupled, multimaster logical replication using a mesh topology. This means that you can write to any server and the changes are sent directly, row-by-row, to all the other servers that are part of the same PGD group.
By default, PGD uses asynchronous replication, applying changes on the peer nodes only after the local commit. Multiple synchronous replication options are also available.
Basic architecture
Multiple groups
A PGD node is a member of at least one node group. In the most basic architecture, there's a single node group for the whole PGD cluster.
Multiple masters
Each node (database) participating in a PGD group both receives changes from other members and can be written to directly by the user.
This is distinct from hot or warm standby, where only one master server accepts writes and all the other nodes are standbys that replicate either from the master or from another standby.
You don't have to write to all the masters all of the time. A frequent configuration directs writes mostly to just one master called the write leader.
Asynchronous, by default
Changes made on one PGD node aren't replicated to other nodes until they're committed locally. As a result, the data isn't exactly the same on all nodes at any given time. Some nodes have data that hasn't yet arrived at other nodes. PostgreSQL's block-based replication solutions default to asynchronous replication as well. In PGD, there are multiple masters and, as a result, multiple data streams. So data on different nodes might differ even when synchronous_commit and synchronous_standby_names are used.
Mesh topology
PGD is structured around a mesh network where every node connects to every other node, and all nodes exchange data directly with each other. There's no forwarding of data in PGD except in special circumstances, such as adding and removing nodes. Data can arrive from outside the EDB Postgres Distributed cluster or be sent onward using native PostgreSQL logical replication.
Logical replication
Logical replication is a method of replicating data rows and their changes based on their replication identity (usually a primary key). We use the term logical in contrast to physical replication, which uses exact block addresses and byte-by-byte replication. Index changes aren't replicated, thereby avoiding write amplification and reducing bandwidth.
Logical replication starts by copying a snapshot of the data from the source node. Once that's done, later commits are sent to other nodes as they occur in real time. Changes are replicated without executing SQL again, so the exact data written is replicated quickly and accurately.
Nodes apply data in the order in which commits were made on the source node, ensuring transactional consistency is guaranteed for the changes from any single node. Changes from different nodes are applied independently of other nodes to ensure the rapid replication of changes.
Replicated data is sent in binary form when it's safe to do so.
Connection management
Connection management leverages consensus-driven quorum to determine the correct connection endpoint in a semi-exclusive manner to prevent unintended multi-node writes from an application. This approach reduces the potential for data conflicts. The node selected as the correct connection endpoint at any point in time is referred to as the write leader.
PGD Proxy is the tool for application connection management provided as part of EDB Postgres Distributed.
High availability
Each master node can be protected by one or more standby nodes, so any node that goes down can be quickly replaced and continue. Each standby node can be either a logical or a physical standby node.
Replication continues between currently connected nodes even if one or more nodes are currently unavailable. When the node recovers, replication can restart from where it left off without missing any changes.
Nodes can run different release levels, negotiating the required protocols to communicate. As a result, EDB Postgres Distributed clusters can use rolling upgrades, even for major versions of database software.
DDL is replicated across nodes by default. DDL execution can be user controlled to allow rolling application upgrades, if desired.
Architectural options and performance
Always On architectures
A number of different architectures can be configured, each of which has different performance and scalability characteristics.
The group is the basic building block consisting of 2+ nodes (servers). In a group, each node is in a different availability zone, with dedicated router and backup, giving immediate switchover and high availability. Each group has a dedicated replication set defined on it. If the group loses a node, you can easily repair or replace it by copying an existing node from the group.
The Always On architectures are built from either one group in a single location or two groups in two separate locations. Each group provides high availability. When two groups are leveraged in remote locations, they together also provide disaster recovery (DR).
Tables are created across both groups, so any change goes to all nodes, not just to nodes in the local group.
One node in each group is selected as the group's write leader. The write leader is the target for the main application's writes and queries. All other nodes are described as shadow nodes (or "read-write replica"), waiting to take over when needed. If a node loses contact, processing switches immediately to a shadow node to continue processing. If a group fails, processing can switch to the other group. Scalability isn't the goal of this architecture.
Since writes are mainly to only one node, the possibility of contention between nodes is reduced to almost zero. As a result, performance impact is much reduced.
Secondary applications might execute against the shadow nodes, although these are reduced or interrupted if the main application begins using that node.
In the future, one node will be elected as the main replicator to other groups, limiting CPU overhead of replication as the cluster grows and minimizing the bandwidth to other groups.
Supported Postgres database servers
PGD is compatible with PostgreSQL, EDB Postgres Extended Server and EDB Postgres Advanced Server and is deployed as a standard Postgres extension named BDR. See the Compatibility matrix for details of supported version combinations.
Some key PGD features depend on certain core capabilities being available in the target Postgres database server. Therefore, PGD users must also adopt the Postgres database server distribution that's best suited to their business needs. For example, if having the PGD feature Commit At Most Once (CAMO) is mission critical to your use case, don't adopt the community PostgreSQL distribution. It doesn't have the core capability required to handle CAMO. See the full feature matrix compatibility in Choosing a Postgres distribution.
PGD offers close-to-native Postgres compatibility. However, some access patterns don't necessarily work as well in multi-node setup as they do on a single instance. There are also some limitations in what you can safely replicate in a multi-node setting. Application usage goes into detail about how PGD behaves from an application development perspective.
Characteristics affecting performance
By default, PGD keeps one copy of each table on each node in the group, and any changes propagate to all nodes in the group.
Since copies of data are everywhere, SELECTs need only ever access the local node. On a read-only cluster, performance on any one node isn't affected by the number of nodes and is immune to replication conflicts on other nodes caused by long-running SELECT queries. Thus, adding nodes increases linearly the total possible SELECT throughput.
If an INSERT, UPDATE, and DELETE (DML) is performed locally, then the changes propagate to all nodes in the group. The overhead of DML apply is less than the original execution. So if you run a pure write workload on multiple nodes concurrently, a multi-node cluster can handle more TPS than a single node.
Conflict handling has a cost that acts to reduce the throughput. The throughput then depends on how much contention the application displays in practice. Applications with very low contention perform better than a single node. Applications with high contention can perform worse than a single node. These results are consistent with any multimaster technology and aren't particular to PGD.
Synchronous replication options can send changes concurrently to multiple nodes so that the replication lag is minimized. Adding more nodes means using more CPU for replication, so peak TPS reduces slightly as each node is added.
If the workload tries to use all CPU resources, then this resource constrains replication, which can then affect the replication lag.
In summary, adding more master nodes to a PGD group doesn't result in significant write throughput increase when most tables are replicated because all the writes are replayed on all nodes. Because PGD writes are in general more effective than writes coming from Postgres clients by way of SQL, you can increase performance. Read throughput generally scales linearly with the number of nodes.