Galera multimaster MySQL cluster for Zabbix?

[Tobias]

Dear Zabbix users,

a potential user of Galera Cluster for MySQL (our product, www.codership.com:)) is asking if we have any experience with Galera and Zabbix. Unfortunately we do not have that experience , but I would like to ask the community, if anyone of you have used Galera for high availability and scalability when using MySQL?

Galera is open source and we have been around since 2008, so I am wondering if we have any fans/users out there?

Your feedback would be appreciated.
Thanks.

[richlv]

hmm. i know about one deployment of zabbix with galera.
wondering whether it's the same or a different company, though

[IkRenzo]

I have setup a Galera XtraDB Cluster multi-master with Zabbix and Haproxy but with round-robin balance algorithm Zabbix encounter deadlocks :/
It works great with source balance algorithm in Haproxy but I would prefer round-robin and search a solution to avoid the deadlocks in this case.

[claytronic]

I have just started building out a concept 3 node Zabbix HA system using Corosync + mariadb-galera-server. My hope is to replace the Corosync + DRBD HA system we've been using.

[s3b0]

I have setup a Galera XtraDB Cluster multi-master with Zabbix and Haproxy but with round-robin balance algorithm Zabbix encounter deadlocks :/
It works great with source balance algorithm in Haproxy but I would prefer round-robin and search a solution to avoid the deadlocks in this case.

Hi,

I have the same problem - did you manage to solve this(the dedlock problem on galera) ?

[kloczek]

Hi,

I have the same problem - did you manage to solve this(the dedlock problem on galera) ?

Instead trying to fight with galera just buy descent hardware able to handle big number of inserts and updates.
Today for ~£700 is possible to buy 1.2TB i750 Inter NVMe SSD. Such SSD can handle up to 400k IO/s with avg latency below 10 microseconds (on using typical SAS/SATA SSD you will have not less than 2-3 microseconds .. so NVMe is almost 100-1000 times faster). If someone as even bigger needs those cards in Oracle X5 hardware scales linearly up to four such cards .. able to handle 1.6 mln IO/s.
On using such hardware with ZFS with lzjb compression typical zabbix data is possible to compress with 2-2.5 compression ratio so even not using those SSDs as L2ARC is possible to store 2-2.5 bigger database than raw storage size.
To handle big number of selects all what you need is a little more RAM.

To Secure content of zabbix DB all what you need is slave database and move all monitoring behind set of proxies. With few hours buffered data on proxies every reconnect will cause resync of the data buffered on proxy with those data already stored on main zabbix DB.

Well architected zabbix stack on main zabbix database should have read to write IOs like at last 1:50. On extreme size zabbix installation read IOs (and selects on DB layer) are never big deal so scaling those operations horizontally is pointless.

Using zabbix with multimaster DB engine is mistake and/or lack of understanding how zabbix interacts with DB backend.

[cdavid]

You can replace Haproxy by maxscale to resolve your problem with the deadlock

[s3b0]

Instead trying to fight with galera just buy descent hardware able to handle big number of inserts and updates.
Today for ~£700 is possible to buy 1.2TB i750 Inter NVMe SSD. Such SSD can handle up to 400k IO/s with avg latency below 10 microseconds (on using typical SAS/SATA SSD you will have not less than 2-3 microseconds .. so NVMe is almost 100-1000 times faster). If someone as even bigger needs those cards in Oracle X5 hardware scales linearly up to four such cards .. able to handle 1.6 mln IO/s.
On using such hardware with ZFS with lzjb compression typical zabbix data is possible to compress with 2-2.5 compression ratio so even not using those SSDs as L2ARC is possible to store 2-2.5 bigger database than raw storage size.
To handle big number of selects all what you need is a little more RAM.

To Secure content of zabbix DB all what you need is slave database and move all monitoring behind set of proxies. With few hours buffered data on proxies every reconnect will cause resync of the data buffered on proxy with those data already stored on main zabbix DB.

Well architected zabbix stack on main zabbix database should have read to write IOs like at last 1:50. On extreme size zabbix installation read IOs (and selects on DB layer) are never big deal so scaling those operations horizontally is pointless.

Using zabbix with multimaster DB engine is mistake and/or lack of understanding how zabbix interacts with DB backend.

Thx - but i don't need that kind of architecture, i'm playing with zabbix because of my hobby and I don't want to buy anything more then banana pi (i have 2 clusters with apache and mariadb on those machines).
I made some optimizations on OS level (my lovely gentoo ) and i don't see now any deadlocks (so why to buy any dedicated servers ?).

I raised timeouts on mariadbs and on haproxy for them

[kloczek]

Thx - but i don't need that kind of architecture, i'm playing with zabbix because of my hobby and I don't want to buy anything more then banana pi (i have 2 clusters with apache and mariadb on those machines).
I made some optimizations on OS level (my lovely gentoo ) and i don't see now any deadlocks (so why to buy any dedicated servers ?).

I raised timeouts on mariadbs and on haproxy for them

Actually .. ZFS has been ported to Pi as well
https://www.raspberrypi.org/forums/v...p?f=63&t=19062

It has been done by two guys (independently)
As well to start using ZFS you do need to buy anything :P