diff --git a/include/haproxy/server-t.h b/include/haproxy/server-t.h
index ff5faae6a..1ff7d18b1 100644
--- a/include/haproxy/server-t.h
+++ b/include/haproxy/server-t.h
@@ -212,6 +212,11 @@ struct srv_per_thread {
 	struct eb_root avail_conns;             /* Connections in use, but with still new streams available */
 };
 
+/* Each server will have one occurrence of this structure per thread group */
+struct srv_per_tgroup {
+	unsigned int next_takeover;             /* thread ID to try to steal connections from next time */
+};
+
 /* Configure the protocol selection for websocket */
 enum __attribute__((__packed__)) srv_ws_mode {
 	SRV_WS_AUTO = 0,
@@ -239,6 +244,7 @@ struct server {
 	const struct mux_proto_list *mux_proto;       /* the mux to use for all outgoing connections (specified by the "proto" keyword) */
 	unsigned maxconn, minconn;		/* max # of active sessions (0 = unlimited), min# for dynamic limit. */
 	struct srv_per_thread *per_thr;         /* array of per-thread stuff such as connections lists */
+	struct srv_per_tgroup *per_tgrp;        /* array of per-tgroup stuff such as idle conns */
 	unsigned int *curr_idle_thr;            /* Current number of orphan idling connections per thread */
 
 	unsigned int pool_purge_delay;          /* Delay before starting to purge the idle conns pool */
@@ -282,7 +288,6 @@ struct server {
 	unsigned int curr_used_conns;           /* Current number of used connections */
 	unsigned int max_used_conns;            /* Max number of used connections (the counter is reset at each connection purges */
 	unsigned int est_need_conns;            /* Estimate on the number of needed connections (max of curr and previous max_used) */
-	unsigned int next_takeover;             /* thread ID to try to steal connections from next time */
 
 	struct queue queue;			/* pending connections */
 
diff --git a/src/backend.c b/src/backend.c
index 4b1ce5ce3..bac8a08a5 100644
--- a/src/backend.c
+++ b/src/backend.c
@@ -1201,7 +1201,7 @@ static struct connection *conn_backend_get(struct stream *s, struct server *srv,
 	/* Lookup all other threads for an idle connection, starting from last
 	 * unvisited thread, but always staying in the same group.
 	 */
-	stop = srv->next_takeover;
+	stop = srv->per_tgrp[tgid - 1].next_takeover;
 	if (stop >= tg->count)
 		stop %= tg->count;
 
@@ -1246,7 +1246,7 @@ static struct connection *conn_backend_get(struct stream *s, struct server *srv,
 		conn = NULL;
  done:
 	if (conn) {
-		_HA_ATOMIC_STORE(&srv->next_takeover, (i + 1 == tg->base + tg->count) ? tg->base : i + 1);
+		_HA_ATOMIC_STORE(&srv->per_tgrp[tgid - 1].next_takeover, (i + 1 == tg->base + tg->count) ? tg->base : i + 1);
 
 		srv_use_conn(srv, conn);
 
diff --git a/src/server.c b/src/server.c
index 699113e88..51ca0cdf3 100644
--- a/src/server.c
+++ b/src/server.c
@@ -2381,6 +2381,7 @@ struct server *srv_drop(struct server *srv)
 	free(srv->hostname_dn);
 	free((char*)srv->conf.file);
 	free(srv->per_thr);
+	free(srv->per_tgrp);
 	free(srv->curr_idle_thr);
 	free(srv->resolvers_id);
 	free(srv->addr_node.key);
@@ -4650,7 +4651,8 @@ int srv_init_per_thr(struct server *srv)
 	int i;
 
 	srv->per_thr = calloc(global.nbthread, sizeof(*srv->per_thr));
-	if (!srv->per_thr)
+	srv->per_tgrp = calloc(global.nbtgroups, sizeof(*srv->per_tgrp));
+	if (!srv->per_thr || !srv->per_tgrp)
 		return -1;
 
 	for (i = 0; i < global.nbthread; i++) {