Skip to content

Latest commit

 

History

History
195 lines (150 loc) · 5.33 KB

DESIGN.md

File metadata and controls

195 lines (150 loc) · 5.33 KB
Raw

pgrouter - A Postgres Router

pgrouter provides a replication-aware smart router that can sit in front of a replicated PostgreSQL cluster and intelligently direct queries to the appropriate node based on type, replication lag and backend availability.

Moving Parts

pgrouter is a multi-threaded server that uses a shared dataset for configuration and backend state. Threads use mutexes and semaphores to safely synchronize access to the shared memory.

A SUPERVISOR thread handles all UNIX signals, including SIGHUP (reload) and SIGTERM (shutdown).

A WATCHER thread checks the health of the backend nodes, and update the shared dataset with availability and replication lag statistics.

A MONITOR thread listens for inbound connections on a (loopback) interface and port. When a client connects, it dumps monitoring and state data (in a custom, human-readable format) to the client and closes the connection. This can be used for validating the health of a pgrouter installation.

Multiple WORKER threads listen on a TCP port for inbound client connections and dispatch them to an appropriate, healthy backend.

Some Bits of Code

/* Status of backends */
#define BACKEND_IS_OK        0  /* backend is up and running    */
#define BACKEND_IS_STARTING  1  /* backend is still connecting  */
#define BACKEND_IS_FAILED    2  /* backend won't connect        */

/* SSL/TLS behaviors */
#define BACKEND_TLS_OFF      0  /* don't do SSL/TLS             */
#define BACKEND_TLS_VERIFY   1  /* do SSL/TLS; verify certs.    */
#define BACKEND_TLS_NOVERIFY 2  /* do SSL/TLS; skip verif.      */

typedef lag_t unsigned long long int;

typedef struct {
	pthread_rwlock_t rwlock;    /* read/write lock for sync.    */
	char *hostname;             /* hostname/ip to connect to    */
	int port;                   /* port to connect to           */

	int tls;                    /* a BACKEND_TLS_* constant     */

	int master;                 /* 1 = master, 0 = slave        */
	int status;                 /* a BACKEND_IS_* constant      */
	int weight;                 /* weight, range [0, RAND_MAX]  */

	struct {
		const char *database;   /* database to connect to       */
		const char *username;   /* username to connect with     */
		const char *password;   /* password to auth. with       */

		lag_t lag;              /* replication lag, in bytes    */
		lag_t threshold;        /* threshold for lag (bytes)    */
	} health;
} BACKEND;

typedef struct {
	pthread_rwlock_t rwlock;    /* read/write lock for sync.    */

	int workers;                /* how many WORKER threads      */
	int loglevel;               /* what messages to log         */

	char *hbafile;              /* path to the HBA/ACL config.  */
	char *pidfile;              /* path to the daemon pidfile   */

	struct {
		int interval;           /* how often to check backends  */
		int timeout;            /* total timeout in seconds     */

		char *database;         /* database to connect to       */
		char *username;         /* username to connect with     */
		char *password;         /* password to auth. with       */
	} health;

	struct {
		char *listen_host;      /* host/ip to bind              */
		int   listen_port;      /* port to bind                 */

		char *monitor_host;     /* host/ip to bind              */
		int   monitor_port;     /* port to bind                 */

		char *tls_ciphers;      /* SSL/TLS ciphers to advert.   */
		char *tls_certfile;     /* path to SSL/TLS certificate  */
		char *tls_keyfile;      /* path to SSL/TLS key          */

		char *user;             /* user to run as               */
		char *group;            /* group to run as              */
		int daemonize;          /* to daemonize or not          */
	} startup;

	int num_backends;           /* how many backends are there? */
	BACKEND *backends;          /* the backends -- epic         */
} CONTEXT;

Monitoring Data

When connecting to the monitoring port, clients should recceive something that looks like this:

backends 3/5\n
workers 64\n
clients 32\n
connections 28\n
10.0.0.3:6432 master OK 0\n
10.0.0.1:6432 slave OK 200\n
10.0.0.2:6432 slave OK 208\n
10.0.0.4:6432 slave FAILED\n
10.0.0.5:6432 slave STARTING\n

The lines for OK backends contain:

  1. host:port
  2. type (master or slave)
  3. status (OK)
  4. lag (0 on master, bytes on slaves)

Usage

Start the daemon using the standard configuration file, and daemonize into the background:

pgrouter

Run in the foregound, with a custom configuration file, and lots of logging (to stderr):

pgrouter -Fvvvvv -c /path/to/config

Parse the configuration and exit 0 if everything checks out. Errors should be reported to stderr and cause an exit 1:

pgrouter -t

Configuration

# pgrouter.conf

listen *:5432
monitor 127.0.0.1:9881
workers 64
hba /path/to/hba

tls {
  ciphers ALL:!EXP:!LOW
  cert /path/to/pem
  key  /path/to/pem
}

user  vcap
group vcap
pidfile /var/vcap/sys/run/pgrouter/pgrouter.pid
log INFO

health {
  timeout 5s
  check 15s
  database postgres
  username pgtest
  password "sekrit squirrel"
}

backend default {
  tls skipverify;
  lag 200
  weight 100
}
backend 10.0.0.5:6432 { lag 800; }
backend 10.0.0.6:6432

backend default { tls off }
backend 10.0.0.7:6432 { weight 125; lag 500b }
backend 10.0.0.8:6432 { weight 250 }