On Fri, 5 Apr 2024 at 15:00, Andrii Ustymenko <andrii.ustyme...@adyen.com> wrote: > > Dear list! > > My name is Andrii. I work for Adyen. We are using haproxy as our main > software loadbalancer at quite large scale. > Af of now our main use-case for backends routing based on > server-template and dynamic dns from consul as service discovery. Below > is the example of simple backend configuration: > > ``` > backend example > balance roundrobin > server-template application 10 _application._tcp.consul resolvers > someresolvers init-addr last,libc,none resolve-opts allow-dup-ip > resolve-prefer ipv4 check ssl verify none > ``` > > and in global configuration > > ``` > resolvers someresolvers > nameserver ns1 10.10.10.10:53 > nameserver ns2 10.10.10.11:53 > ``` > > As we see haproxy will create internal table for backends with some > be_id and be_name=application and allocate 10 records for each server > with se_id from 1 to 10. Then those records get populated and updated > with the data from resolvers. > I would like to understand couple of things with regards to this > structure and how it works, which I could not figure out myself from the > source code: > > 1) In tcpdump for dns queries we see that haproxy asynchronously polls > all the nameservers simultaneously. For instance: > > ``` > 11:06:17.587798 eth2 Out ifindex 4 aa:aa:aa:aa:aa:aa ethertype IPv4 > (0x0800), length 108: 10.10.10.50.24050 > 10.10.10.10.53: 34307+ [1au] > SRV? _application._tcp.consul. (60) > 11:06:17.587802 eth2 Out ifindex 4 aa:aa:aa:aa:aa:aa ethertype IPv4 > (0x0800), length 108: 10.10.10.50.63155 > 10.10.10.11.53: 34307+ [1au] > SRV? _application._tcp.consul. (60) > 11:06:17.588097 eth2 In ifindex 4 ff:ff:ff:ff:ff:ff ethertype IPv4 > (0x0800), length 205: 10.10.10.10.53 > 10.10.10.50.24050: 2194 2/0/1 SRV > 0a5099e5.addr.consul.:25340 1 1, SRV 0a509934.addr.consul.:26010 1 1 (157) > 11:06:17.588097 eth2 In ifindex 4 ff:ff:ff:ff:ff:ff ethertype IPv4 > (0x0800), length 205: 10.10.10.11.53 > 10.10.10.50.63155: 2194 2/0/1 SRV > 0a5099e5.addr.consul.:25340 1 1, SRV 0a509934.addr.consul.:26010 1 1 (157) > ``` > > Both nameservers reply with the same response. But what if they are out > of sync? Let's say one says: server1, server2 and the second one says > server2, server3? So far testing this locally - I see sometimes the > reply overrides the table, but sometimes it seems to just gets merged > with the rest. >
Most service discovery systems employ "eventually consistency". Therefore, some instances of the DNS servers in front of that system may have "stale" data. The instance usually becomes consistent with milliseconds or seconds; it depends on the topology. I don't think haproxy performs any kind of merging, I believe it uses the 1st valid response. > 2) Each entry from SRV reply will be placed into the table under > specific se_id. Most of the times that placement won't change. So, for > the example above the most likely 0a5099e5.addr.consul. and > 0a509934.addr.consul. will have se_id 1 and 2 respectively. However > sometimes we have the following scenario: > > 1. We admistratively disable the server (drain traffic) with the next > command: > > ``` > echo "set server example/application1 state maint" | nc -U > /var/lib/haproxy/stats > ``` > > the MAINT flag will be added to the record with se_id 1 > > 2. Instance of application goes down and gets de-registered from consul, > so also evicted from srv replies and out of discovery of haproxy. > > 3. Instance of application goes up and gets registered by consul and > discovered by haproxy, but haproxy allocates different se_id for it. > Haproxy healthchecks will control the traffic to it in this case. > > 4. We will still have se_id 1 with MAINT flag and application instance > dns record placed into different se_id. > > The problem comes that any new discovered record which get placed into > se_id 1 will never be active until either command: > > ``` > echo "set server example/application1 state ready" | nc -U > /var/lib/haproxy/stats > ``` > > executed or haproxy gets reloaded without state file. With this pattern > we basically have persistent "records pollution" due to operations made > directly with control socket. > This situation could lead to minor incidents where most newly registered servers are assigned to se_ids that were previously in maintenance mode. So, you end up with a backend that has, let's say, 75% of servers in maintenance mode. > I am not sure is there anything to do about this. Maybe, if haproxy > could cache the state not only of se_id but also associated record with > that and then if that gets changed - re-schedule healtchecks. Or instead > of integer ids use some hashed ids based on dns/ip-addresses of > discovered records, in this case binding will happen exactly in the same > slot. > Another approach could be haproxy resetting the `srv_admin_state ` field for IDs for which an entry was de-registered. My 2 cents, Pavlos
