ADP architecture
Arenadata Prosperity (ADP) is a DBMS for efficient work with loads of different profiles (primarily OLTP — Online Transaction Processing). ADP allows you to work with various amounts of data and supports a wide range of data types (including JSON and custom types) with the ability to create data processing mechanisms.
The main components of the ADP cluster in the diagram:
-
ADP leader — the main host that has the right to perform write transactions.
-
ADP replica — additional hosts that are replicas of the leading ADP host.
-
etcd — distributed configuration storage.
-
Load balancer — distributes read/write transactions between leader host and replica hosts (HAProxy).
-
Connection pooler — the PgBouncer instances installed on each ADP node that manage the number of connections and reuse them.
ADP uses Patroni to build the cluster based on the PostgreSQL streaming replication and implement High Availability (HA). An ADP cluster that supports HA should contain several ADP nodes. Each node consists of a PostgreSQL server and a Patroni agent that serves the local PostgreSQL instance. One ADP node is the leader, others are replica instances. The leader serves read-write transactions (unless otherwise specified in load balancing settings), and replicas process only read-only requests. In case of the leader failure, Patroni elects a new leader from replicas and the ADP cluster continues to operate.
Patroni saves the cluster configuration in the Distributed Configuration Storage (DCS). ADP requires an etcd cluster to use it as Patroni DCS.
To implement load balancing, ADP uses HAProxy (High Availability Proxy). HAProxy is a software TCP/HTTP load balancer. It listens on ports in pairs: connections to one port from the pair are transferred to the leader (directly or throw PgBouncer), requests to another port are distributed across ADP nodes (directly or throw PgBouncer). For more information, see Load balancing.
PgBouncer is used as a connection pooler. An application can connect to PgBouncer as an ADP server, and PgBouncer creates a connection to the ADP server, or reuses one of its existing connections. PgBouncer is designed to reduce the performance impact of opening new connections. To avoid compromising transaction semantics, PgBouncer supports multiple types of pooling when rotating connections:
-
Session pooling — when a client connects, a server connection is assigned to it for the entire time the client is connected. When the client disconnects, the server connection returns to the pool. This is the default method.
-
Transaction pooling — a server connection is assigned to a client only during a transaction. When PgBouncer notices that the transaction is complete, the server connection is put back into the pool.
-
Statement pooling — the server connection is returned to the pool immediately after a query completes. Multi-statement transactions are prohibited in this mode, because they may be disrupted.
ADP cluster management is performed via ADCM. The following actions are available in the ADCM UI:
-
Cluster actions that allow you to manage the entire cluster with all services.
-
Service actions that manage a single service.