pgagroal architecture ​

Overview ​

pgagroal use a process model (fork()), where each process handles one connection to PostgreSQL. This was done such a potential crash on one connection won't take the entire pool down.

The main process is defined in main.c. When a client connects it is processed in its own process, which is handle in worker.h (worker.c).

Once the client disconnects the connection is put back in the pool, and the child process is terminated.

Shared memory ​

A memory segment (shmem.h) is shared among all processes which contains the pgagroal state containing the configuration of the pool, the list of servers and the state of each connection.

The configuration of pgagroal (struct configuration), the configuration of the servers (struct server) and the state of each connection (struct connection) is initialized in this shared memory segment. These structs are all defined in pgagroal.h.

The shared memory segment is created using the mmap() call.

Atomic operations ​

The atomic operation library is used to define the state of each of the connection, and move them around in the connection state diagram. The state diagram has the follow states

These state are defined in pgagroal.h.

Pool ​

The pgagroal pool API is defined in pool.h (pool.c).

This API defines the functionality of the pool such as getting a connection from the pool, and returning it. There is no ordering among processes, so a newly created process can obtain a connection before an older process.

The pool operates on the struct connection data type defined in pgagroal.h.

Network and messages ​

All communication is abstracted using the struct message data type defined in message.h.

Reading and writing messages are handled in the message.h (message.c) files.

Network operations are defined in network.h (network.c).

Memory ​

Each process uses a fixed memory block for its network communication, which is allocated upon startup of the worker.

That way we don't have to allocate memory for each network message, and more importantly free it after end of use.

The memory interface is defined in memory.h (memory.c).

Management ​

pgagroal has a management interface which defines the administrator abilities that can be performed when it is running. This include for example taking a backup. The pgagroal-cli program is used for these operations (cli.c).

The management interface is defined in management.h. The management interface uses its own protocol which uses JSON as its foundation.

Write ​

The client sends a single JSON string to the server,

The server sends a single JSON string to the client,

Read ​

The server sends a single JSON string to the client,

The client sends to the server a single JSON documents,

Remote management ​

The remote management functionality uses the same protocol as the standard management method.

However, before the management packet is sent the client has to authenticate using SCRAM-SHA-256 using the same message format that PostgreSQL uses, e.g. StartupMessage, AuthenticationSASL, AuthenticationSASLContinue, AuthenticationSASLFinal and AuthenticationOk. The SSLRequest message is supported.

The remote management interface is defined in remote.h (remote.c).

I/O layer ​

The I/O layer interface is primarily defined in ev.h (and implemented in ev.c).

These files contain the definition and implementation of the event loop for the three supported backends: io_uring, epoll, and kqueue.

The backend is defined during runtime and can be set with the configuration option ev_backend. Default is auto, which will select the first supported backend, considering the following order: io_uring, epoll, kqueue.

liburing was used for setup and usage io_uring instances.

Each process has its own event loop, such that the process only gets notified when data related only to that process is ready. The main loop handles the system wide "services" such as idle timeout checks and so on.

The I/O Layer works with a registered event watchers. Those can either be a watcher for I/O events (io_watcher), Timer events (periodic_watcher) and Signal events (signal_watcher).

The event interface provides ways to register and cancel watching events through the above watchers.

The events watched by the main loop is different from the events watched by the workers.

The main loop registers timers, signals and accept watchers.

The worker registers the client watcher (responsible for receiving the message from the client and bouncing it to the server), the server watcher (responsible for watching for a message from the server and bouncing it to the client) and one signal watcher.

The event loop system supports multiple execution contexts to handle different pgagroal components:

• Main Context (PGAGROAL_CONTEXT_MAIN): Used by the main pgagroal process for connection pooling and management operations
• Vault Context (PGAGROAL_CONTEXT_VAULT): Used by pgagroal-vault for HTTP server operations and management communication

Each context uses its own configuration structure and event backend settings. The context is set explicitly before event loop initialization to ensure the correct configuration is used for backend selection and setup.

Backend Selection ​

The event backend selection process varies by context:

For the main pgagroal process:

• Reads ev_backend setting from main configuration file
• Validates backend availability and TLS compatibility
• Falls back to supported alternatives if needed

For pgagroal-vault:

• Reads ev_backend setting from vault configuration file
• Uses the same validation and fallback logic as main process
• Supports all the same backends: io_uring, epoll, kqueue

Both contexts support the same configuration options:

• auto: Automatically selects the best available backend
• io_uring: Linux-specific, high-performance backend (not supported with TLS)
• epoll: Linux-specific, traditional event notification
• kqueue: BSD/macOS event notification mechanism

The implementation is done in ev.h and ev.c.

Pipeline ​

pgagroal has the concept of a pipeline that defines how communication is routed from the client through pgagroal to PostgreSQL. Likewise in the other direction.

A pipeline is defined by

struct pipeline
{
   initialize initialize;
   start start;
   callback client;
   callback server;
   stop stop;
   destroy destroy;
   periodic periodic;
};

struct pipeline
{
   initialize initialize;
   start start;
   callback client;
   callback server;
   stop stop;
   destroy destroy;
   periodic periodic;
};

in pipeline.h.

The functions in the pipeline are defined as

The functions start, client, server and stop has access to the following information

struct worker_io
{
   struct io_watcher io; /* The base type for io operations */
   int client_fd;        /* The client descriptor */
   int server_fd;        /* The server descriptor */
   int slot;             /* The slot */
   SSL* client_ssl;      /* The client SSL context */
   SSL* server_ssl;      /* The server SSL context */
};

struct worker_io
{
   struct io_watcher io; /* The base type for io operations */
   int client_fd;        /* The client descriptor */
   int server_fd;        /* The server descriptor */
   int slot;             /* The slot */
   SSL* client_ssl;      /* The client SSL context */
   SSL* server_ssl;      /* The server SSL context */
};

defined in worker.h.

Performance pipeline ​

One of the goals for pgagroal is performance, so the performance pipeline will only look for the Terminate message from the client and act on that. Likewise the performance pipeline will only look for FATAL errors from the server. This makes the pipeline very fast, since there is a minimum overhead in the interaction.

The pipeline is defined in pipeline_perf.c in the functions

Session pipeline ​

The session pipeline works like the performance pipeline with the exception that it checks if a Transport Layer Security (TLS) transport should be used.

The pipeline is defined in pipeline_session.c in the functions

Transaction pipeline ​

The transaction pipeline will return the connection to the server after each transaction. The pipeline supports Transport Layer Security (TLS).

The pipeline uses the ReadyForQuery message to check the status of the transaction, and therefore needs to maintain track of the message headers.

The pipeline has a management interface in order to receive the socket descriptors from the parent process when a new connection is added to the pool. The pool will retry if the client in question doesn't consider the socket descriptor valid.

The pipeline is defined in pipeline_transaction.c in the functions

Signals ​

The main process of pgagroal supports the following signals SIGTERM, SIGINT and SIGALRM as a mechanism for shutting down. The SIGTRAP signal will put pgagroal into graceful shutdown, meaning that exisiting connections are allowed to finish their session. The SIGABRT is used to request a core dump (abort()). The SIGHUP signal will trigger a full reload of the configuration. When SIGHUP is received, pgagroal will re-read the configuration from the configuration files on disk and apply any changes that can be handled at runtime. This is the standard way to apply changes made to the configuration files.

In contrast, the SIGUSR1 signal will trigger a service reload, but does not re-read the configuration files. Instead, SIGUSR1 restarts sockets and listeners using the current in-memory configuration. This is useful for applying certain changes (such as re-opening sockets or refreshing listeners) without modifying or reloading the configuration from disk. Any changes made to the configuration files will not be picked up when using SIGUSR1; only the configuration already loaded in memory will be used.

Use SIGHUP when you want to apply changes from updated configuration files.
Use SIGUSR1 when you want to restart services without changing the current configuration.

The child processes support SIGQUIT as a mechanism to shutdown. This will not shutdown the pool itself.

It should not be needed to use SIGKILL for pgagroal. Please, consider using SIGABRT instead, and share the core dump and debug logs with the pgagroal community.

Reload ​

The SIGHUP signal will trigger a reload of the configuration.

However, some configuration settings requires a full restart of pgagroal in order to take effect. These are

• hugepage
• ev_backend
• log_path
• log_type
• max_connections
• pipeline
• unix_socket_dir
• pidfile
• Limit rules defined by pgagroal_databases.conf
• TLS rules defined by server section

The configuration can also be reloaded using pgagroal-cli -c pgagroal.conf conf reload. The command is only supported over the local interface, and hence doesn't work remotely.

Prometheus ​

pgagroal has support for Prometheus when the metrics port is specified.

Note: It is crucial to carefully initialize Prometheus memory in any program files for example functions like pgagroal_init_prometheus() and pgagroal_init_prometheus_cache() should only be invoked if metrics is greater than 0.

The module serves two endpoints

• / - Overview of the functionality (text/html)
• /metrics - The metrics (text/plain)

All other URLs will result in a 403 response.

The metrics endpoint supports Transfer-Encoding: chunked to account for a large amount of data.

The implementation is done in prometheus.h and prometheus.c.

Failover support ​

pgagroal can failover a PostgreSQL instance if clients can't write to it.

This is done using an external script provided by the user.

The implementation is done in server.h and server.c.

Logging ​

Simple logging implementation based on a atomic_schar lock.

The implementation is done in logging.h and logging.c.

Protocol ​

The protocol interactions can be debugged using Wireshark or pgprtdbg.

Database Alias ​

A database alias in pgagroal allows clients to connect using an alternative name for a configured database. This is useful for scenarios such as application migrations, multi-tenancy, or providing user-friendly names without exposing the actual backend database name.

How it works ​

• Each database entry in the limits configuration (pgagroal_databases.conf) can specify up to eight aliases.
• When a client connects using an alias, pgagroal transparently maps the alias to the real database name before establishing or reusing a backend connection.
• Aliases are resolved during both pooled and unpooled connection handling, ensuring that connections are matched and authenticated against the correct backend database.