Support connection load balancing in libpq

This adds support for load balancing connections with libpq using a
connection parameter: load_balance_hosts=<string>. When setting the
param to random, hosts and addresses will be connected to in random
order. This then results in load balancing across these addresses and
hosts when multiple clients or frequent connection setups are used.

The randomization employed performs two levels of shuffling:

  1. The given hosts are randomly shuffled, before resolving them
     one-by-one.
  2. Once a host its addresses get resolved, the returned addresses
     are shuffled, before trying to connect to them one-by-one.

Author: Jelte Fennema <[email protected]>
Reviewed-by: Aleksander Alekseev <[email protected]>
Reviewed-by: Michael Banck <[email protected]>
Reviewed-by: Andrey Borodin <[email protected]>
Discussion: https://postgr.es/m/PR3PR83MB04768E2FF04818EEB2179949F7A69@PR3PR83MB0476.EURPRD83.prod.outlook.

Author: danielgustafsson

.cirrus.yml

@@ -25,7 +25,7 @@ env:
   MTEST_ARGS: --print-errorlogs --no-rebuild -C build
   PGCTLTIMEOUT: 120 # avoids spurious failures during parallel tests
   TEMP_CONFIG: ${CIRRUS_WORKING_DIR}/src/tools/ci/pg_ci_base.conf
-  PG_TEST_EXTRA: kerberos ldap ssl
+  PG_TEST_EXTRA: kerberos ldap ssl load_balance
 
 
 # What files to preserve in case tests fail
@@ -313,6 +313,14 @@ task:
     mkdir -m 770 /tmp/cores
     chown root:postgres /tmp/cores
     sysctl kernel.core_pattern='/tmp/cores/%e-%s-%p.core'
+
+  setup_hosts_file_script: |
+    cat >> /etc/hosts <<-EOF
+      127.0.0.1 pg-loadbalancetest
+      127.0.0.2 pg-loadbalancetest
+      127.0.0.3 pg-loadbalancetest
+    EOF
+
   setup_additional_packages_script: |
     #apt-get update
     #DEBIAN_FRONTEND=noninteractive apt-get -y install ...
@@ -564,6 +572,12 @@ task:
   setup_additional_packages_script: |
     REM choco install -y --no-progress ...
 
+  setup_hosts_file_script: |
+    echo 127.0.0.1 pg-loadbalancetest >> c:\Windows\System32\Drivers\etc\hosts
+    echo 127.0.0.2 pg-loadbalancetest >> c:\Windows\System32\Drivers\etc\hosts
+    echo 127.0.0.3 pg-loadbalancetest >> c:\Windows\System32\Drivers\etc\hosts
+    type c:\Windows\System32\Drivers\etc\hosts
+
   # Use /DEBUG:FASTLINK to avoid high memory usage during linking
   configure_script: |
     vcvarsall x64

doc/src/sgml/libpq.sgml

@@ -2115,6 +2115,67 @@ postgresql://%2Fvar%2Flib%2Fpostgresql/dbname
        </para>
       </listitem>
      </varlistentry>
+
+     <varlistentry id="libpq-connect-load-balance-hosts" xreflabel="load_balance_hosts">
+      <term><literal>load_balance_hosts</literal></term>
+      <listitem>
+       <para>
+        Controls the order in which the client tries to connect to the available
+        hosts and addresses. Once a connection attempt is successful no other
+        hosts and addresses will be tried. This parameter is typically used in
+        combination with multiple host names or a DNS record that returns
+        multiple IPs. This parameter can be used in combination with
+        <xref linkend="libpq-connect-target-session-attrs"/>
+        to, for example, load balance over standby servers only. Once successfully
+        connected, subsequent queries on the returned connection will all be
+        sent to the same server. There are currently two modes:
+        <variablelist>
+         <varlistentry>
+          <term><literal>disable</literal> (default)</term>
+          <listitem>
+           <para>
+            No load balancing across hosts is performed.  Hosts are tried in
+            the order in which they are provided and addresses are tried in
+            the order they are received from DNS or a hosts file.
+           </para>
+          </listitem>
+         </varlistentry>
+
+         <varlistentry>
+          <term><literal>random</literal></term>
+          <listitem>
+           <para>
+            Hosts and addresses are tried in random order. This value is mostly
+            useful when opening multiple connections at the same time, possibly
+            from different machines. This way connections can be load balanced
+            across multiple <productname>PostgreSQL</productname> servers.
+           </para>
+           <para>
+            While random load balancing, due to its random nature, will almost
+            never result in a completely uniform distribution, it statistically
+            gets quite close. One important aspect here is that this algorithm
+            uses two levels of random choices: First the hosts
+            will be resolved in random order. Then secondly, before resolving
+            the next host, all resolved addresses for the current host will be
+            tried in random order. This behaviour can skew the amount of
+            connections each node gets greatly in certain cases, for instance
+            when some hosts resolve to more addresses than others. But such a
+            skew can also be used on purpose, e.g. to increase the number of
+            connections a larger server gets by providing its hostname multiple
+            times in the host string.
+           </para>
+           <para>
+            When using this value it's recommended to also configure a reasonable
+            value for <xref linkend="libpq-connect-connect-timeout"/>. Because then,
+            if one of the nodes that are used for load balancing is not responding,
+            a new node will be tried.
+           </para>
+          </listitem>
+         </varlistentry>
+        </variablelist>
+       </para>
+      </listitem>
+     </varlistentry>
     </variablelist>
    </para>
   </sect2>

doc/src/sgml/regress.sgml

@@ -256,7 +256,7 @@ make check-world -j8 >/dev/null
    <varname>PG_TEST_EXTRA</varname> to a whitespace-separated list, for
    example:
 <programlisting>
-make check-world PG_TEST_EXTRA='kerberos ldap ssl'
+make check-world PG_TEST_EXTRA='kerberos ldap ssl load_balance'
 </programlisting>
    The following values are currently supported:
    <variablelist>
@@ -290,6 +290,17 @@ make check-world PG_TEST_EXTRA='kerberos ldap ssl'
      </listitem>
     </varlistentry>
 
+    <varlistentry>
+     <term><literal>load_balance</literal></term>
+     <listitem>
+      <para>
+       Runs the test <filename>src/interfaces/libpq/t/004_load_balance_dns.pl</filename>.
+       This requires editing the system <filename>hosts</filename> file and
+       opens TCP/IP listen sockets.
+      </para>
+     </listitem>
+    </varlistentry>
+
     <varlistentry>
      <term><literal>wal_consistency_checking</literal></term>
      <listitem>

src/interfaces/libpq/fe-connect.c

@@ -123,6 +123,7 @@ static int	ldapServiceLookup(const char *purl, PQconninfoOption *options,
 #define DefaultChannelBinding	"disable"
 #endif
 #define DefaultTargetSessionAttrs	"any"
+#define DefaultLoadBalanceHosts	"disable"
 #ifdef USE_SSL
 #define DefaultSSLMode "prefer"
 #define DefaultSSLCertMode "allow"
@@ -351,6 +352,11 @@ static const internalPQconninfoOption PQconninfoOptions[] = {
 		"Target-Session-Attrs", "", 15, /* sizeof("prefer-standby") = 15 */
 	offsetof(struct pg_conn, target_session_attrs)},
 
+	{"load_balance_hosts", "PGLOADBALANCEHOSTS",
+		DefaultLoadBalanceHosts, NULL,
+		"Load-Balance-Hosts", "", 8,	/* sizeof("disable") = 8 */
+	offsetof(struct pg_conn, load_balance_hosts)},
+
 	/* Terminating entry --- MUST BE LAST */
 	{NULL, NULL, NULL, NULL,
 	NULL, NULL, 0}
@@ -435,6 +441,8 @@ static void pgpassfileWarning(PGconn *conn);
 static void default_threadlock(int acquire);
 static bool sslVerifyProtocolVersion(const char *version);
 static bool sslVerifyProtocolRange(const char *min, const char *max);
+static bool parse_int_param(const char *value, int *result, PGconn *conn,
+							const char *context);
 
 
 /* global variable because fe-auth.c needs to access it */
@@ -1020,6 +1028,31 @@ parse_comma_separated_list(char **startptr, bool *more)
 	return p;
 }
 
+/*
+ * Initializes the prng_state field of the connection. We want something
+ * unpredictable, so if possible, use high-quality random bits for the
+ * seed. Otherwise, fall back to a seed based on the connection address,
+ * timestamp and PID.
+ */
+static void
+libpq_prng_init(PGconn *conn)
+{
+	uint64		rseed;
+	struct timeval tval = {0};
+
+	if (pg_prng_strong_seed(&conn->prng_state))
+		return;
+
+	gettimeofday(&tval, NULL);
+
+	rseed = ((uint64) conn) ^
+			((uint64) getpid()) ^
+			((uint64) tval.tv_usec) ^
+			((uint64) tval.tv_sec);
+
+	pg_prng_seed(&conn->prng_state, rseed);
+}
+
 /*
  *		connectOptions2
  *
@@ -1619,6 +1652,49 @@ connectOptions2(PGconn *conn)
 	else
 		conn->target_server_type = SERVER_TYPE_ANY;
 
+	/*
+	 * validate load_balance_hosts option, and set load_balance_type
+	 */
+	if (conn->load_balance_hosts)
+	{
+		if (strcmp(conn->load_balance_hosts, "disable") == 0)
+			conn->load_balance_type = LOAD_BALANCE_DISABLE;
+		else if (strcmp(conn->load_balance_hosts, "random") == 0)
+			conn->load_balance_type = LOAD_BALANCE_RANDOM;
+		else
+		{
+			conn->status = CONNECTION_BAD;
+			libpq_append_conn_error(conn, "invalid %s value: \"%s\"",
+									"load_balance_hosts",
+									conn->load_balance_hosts);
+			return false;
+		}
+	}
+	else
+		conn->load_balance_type = LOAD_BALANCE_DISABLE;
+
+	if (conn->load_balance_type == LOAD_BALANCE_RANDOM)
+	{
+		libpq_prng_init(conn);
+
+		/*
+		 * This is the "inside-out" variant of the Fisher-Yates shuffle
+		 * algorithm. Notionally, we append each new value to the array and
+		 * then swap it with a randomly-chosen array element (possibly
+		 * including itself, else we fail to generate permutations with the
+		 * last integer last).  The swap step can be optimized by combining it
+		 * with the insertion.
+		 */
+		for (i = 1; i < conn->nconnhost; i++)
+		{
+			int			j = pg_prng_uint64_range(&conn->prng_state, 0, i);
+			pg_conn_host temp = conn->connhost[j];
+
+			conn->connhost[j] = conn->connhost[i];
+			conn->connhost[i] = temp;
+		}
+	}
+
 	/*
 	 * Resolve special "auto" client_encoding from the locale
 	 */
@@ -2626,6 +2702,32 @@ PQconnectPoll(PGconn *conn)
 		if (ret)
 			goto error_return;	/* message already logged */
 
+		/*
+		 * If random load balancing is enabled we shuffle the addresses.
+		 */
+		if (conn->load_balance_type == LOAD_BALANCE_RANDOM)
+		{
+			/*
+			 * This is the "inside-out" variant of the Fisher-Yates shuffle
+			 * algorithm. Notionally, we append each new value to the array
+			 * and then swap it with a randomly-chosen array element (possibly
+			 * including itself, else we fail to generate permutations with
+			 * the last integer last).  The swap step can be optimized by
+			 * combining it with the insertion.
+			 *
+			 * We don't need to initialize conn->prng_state here, because that
+			 * already happened in connectOptions2.
+			 */
+			for (int i = 1; i < conn->naddr; i++)
+			{
+				int			j = pg_prng_uint64_range(&conn->prng_state, 0, i);
+				AddrInfo	temp = conn->addr[j];
+
+				conn->addr[j] = conn->addr[i];
+				conn->addr[i] = temp;
+			}
+		}
+
 		reset_connection_state_machine = true;
 		conn->try_next_host = false;
 	}
@@ -4320,6 +4422,7 @@ freePGconn(PGconn *conn)
 	free(conn->outBuffer);
 	free(conn->rowBuf);
 	free(conn->target_session_attrs);
+	free(conn->load_balance_hosts);
 	termPQExpBuffer(&conn->errorMessage);
 	termPQExpBuffer(&conn->workBuffer);
 

src/interfaces/libpq/libpq-int.h

@@ -26,7 +26,8 @@
 #include <netdb.h>
 #include <sys/socket.h>
 #include <time.h>
-#ifndef WIN32
+/* MinGW has sys/time.h, but MSVC doesn't */
+#ifndef _MSC_VER
 #include <sys/time.h>
 #endif
 
@@ -82,6 +83,8 @@ typedef struct
 #endif
 #endif							/* USE_OPENSSL */
 
+#include "common/pg_prng.h"
+
 /*
  * POSTGRES backend dependent Constants.
  */
@@ -242,6 +245,13 @@ typedef enum
 	SERVER_TYPE_PREFER_STANDBY_PASS2	/* second pass - behaves same as ANY */
 } PGTargetServerType;
 
+/* Target server type (decoded value of load_balance_hosts) */
+typedef enum
+{
+	LOAD_BALANCE_DISABLE = 0,	/* Use the existing host order (default) */
+	LOAD_BALANCE_RANDOM,		/* Randomly shuffle the hosts */
+} PGLoadBalanceType;
+
 /* Boolean value plus a not-known state, for GUCs we might have to fetch */
 typedef enum
 {
@@ -398,6 +408,7 @@ struct pg_conn
 	char	   *ssl_max_protocol_version;	/* maximum TLS protocol version */
 	char	   *target_session_attrs;	/* desired session properties */
 	char	   *require_auth;	/* name of the expected auth method */
+	char	   *load_balance_hosts; /* load balance over hosts */
 
 	/* Optional file to write trace info to */
 	FILE	   *Pfdebug;
@@ -469,6 +480,8 @@ struct pg_conn
 
 	/* Transient state needed while establishing connection */
 	PGTargetServerType target_server_type;	/* desired session properties */
+	PGLoadBalanceType load_balance_type;	/* desired load balancing
+											 * algorithm */
 	bool		try_next_addr;	/* time to advance to next address/host? */
 	bool		try_next_host;	/* time to advance to next connhost[]? */
 	int			naddr;			/* number of addresses returned by getaddrinfo */
@@ -488,6 +501,8 @@ struct pg_conn
 	PGVerbosity verbosity;		/* error/notice message verbosity */
 	PGContextVisibility show_context;	/* whether to show CONTEXT field */
 	PGlobjfuncs *lobjfuncs;		/* private state for large-object access fns */
+	pg_prng_state prng_state;	/* prng state for load balancing connections */
+
 
 	/* Buffer for data received from backend and not yet processed */
 	char	   *inBuffer;		/* currently allocated buffer */

src/interfaces/libpq/meson.build

@@ -116,6 +116,8 @@ tests += {
     'tests': [
       't/001_uri.pl',
       't/002_api.pl',
+      't/003_load_balance_host_list.pl',
+      't/004_load_balance_dns.pl',
     ],
     'env': {'with_ssl': ssl_library},
   },