Hi Baptiste
It doesn't matter. :-)
When the weight of server is set to 0 with the balance roundrobin
algorithm, srv->eweight is update to 0 and
fwrr_update_server_weight() (lb_fwrr.c) will be called as below:
static void fwrr_update_server_weight(struct server *srv)
{
...
old_state = srv_is_usable(srv->prev_state, srv->prev_eweight);
new_state = srv_is_usable(srv->state, srv->eweight);
if (!old_state && !new_state) {
srv->prev_state = srv->state;
srv->prev_eweight = srv->eweight;
return;
}
else if (!old_state && new_state) {
fwrr_set_server_status_up(srv);
return;
}
else if (old_state && !new_state) {
fwrr_set_server_status_down(srv);
return;
}
...
}
Since srv->eweight is 0, new_state should be also 0, then
fwrr_set_server_status_down() will be called.
At the end the server will be remove from weight tree by
fwrr_dequeue_srv() and lb_tree by fwrr_remove_from_tree().
But the srv->state has not been updated, still keeps in SRV_RUNNING. As
a result, when the same server is selected by
sticking rules, it will be used again.
AH, sorry, my mistake.
I read your mail too quickly.
Baptiste
On Wed, May 29, 2013 at 9:18 AM, Godbach <[email protected]> wrote:
Hi Baptiste
Thanks for your replying.
I am using the balance roundrobin algorithm and sticking on src, not the the
balance source algorithm. The configuration has been presented in my first
mail as below:
backend pool
balance roundrobin.
stick-table type ip size 200k expire 600s
stick on src
server 1 10.128.7.1:80 id 1 cookie srv1 weight 1 maxconn 0
slowstart 0s
server 2 10.128.7.2:80 id 2 cookie srv2 weight 1 maxconn 0
slowstart 0s
Best Regards,
Godbach
On 2013/5/29 13:35, Baptiste wrote:
Hi Godbach,
Before reading HAProxy source code, it worths reading its
configuration guide for the options you use.
IE, the balance source algorithm would tell you that:
"This algorithm is static by default, which means that changing a
server's weight on the fly will have no effect, but this can be
changed using "hash-type"."
Please update your configuration following the recommandation above
and let us know your feedback.
Baptiste
On Wed, May 29, 2013 at 5:22 AM, Godbach <[email protected]> wrote:
Hi, all
It is expected that new http request will not get response from the
server of which weight was changed to 0. It cannot work well with
persistence on src but work well without the persistence in lastest
snapshot.
There are two servers in my backend, and persistence on src ip has been
enabled in backend. The configuration in backend as below:
backend pool
balance roundrobin.
stick-table type ip size 200k expire 600s
stick on src
server 1 10.128.7.1:80 id 1 cookie srv1 weight 1 maxconn 0
slowstart 0s
server 2 10.128.7.2:80 id 2 cookie srv2 weight 1 maxconn 0
slowstart 0s
During continuous http requset with the same client, the stick table as
below:
# table: pool, type: ip, size:204800, used:1
0x17d2284: key=172.22.16.250 use=0 exp=599095 server_id=1
Then I set weight of server 1 to 0 use command as below:
set weight pool/1 0
And I get the weight of server 1 with command:
get weight pool/1
The result is
0 (initial 1)
So I think I have set the weight of sever1 to 0 successfully. But the
response still comes from server 1 which server 2 is expected. And the
stick table keeps the same.
I review the code of process_sticking_rules() in session.c. The codes
when server is found as below:
1403 ptr = stktable_data_ptr(rule->table.t, ts,
STKTABLE_DT_SERVER_ID);
1404 node =
eb32_lookup(&px->conf.used_server_id, stktable_data_cast(ptr,
server_id));
1405 if (node) {
1406 struct server *srv;
1407
1408 srv = container_of(node, struct server,
conf.id);
1409 if ((srv->state & SRV_RUNNING) ||
1410 (px->options & PR_O_PERSIST) ||
1411 (s->flags & SN_FORCE_PRST)) {
1412 s->flags |= SN_DIRECT | SN_ASSIGNED;
1413 s->target = &srv->obj_type;
1414 }
1415 }
1416 }
1417 stktable_touch(rule->table.t, ts, 1);
Line 1409 used (srv->state & SRV_RUNNING) to check the server status.
If I used srv_is_usable() to instead such as below:
- if ((srv->state & SRV_RUNNING) ||
+ if (srv_is_usable(srv->state, srv->eweight) ||
The new request will get response from server 2 once the weight of
server 1 is set to 0. But it seems to be just a workaround.
Since the manual of haproxy about 'set weight' says that:
A typical usage of this command is to disable a server during an update
by setting its weight to zero.
I am wondering that whether the flag SRV_RUNNING of server should be
cleared or not when its weight is set to 0.
--
Best Regards,
Godbach
--
Best Regards,
Godbach
