DRBD-like something for FreeBSD

Dmitriy Kirhlarov dimma at higis.ru
Thu Aug 4 07:41:48 GMT 2005


Hi Anton!

On Wed, 03 Aug 2005, Anton Nikiforov wrote:

> not see the way to promote second node as prymary in case of real 
> prymary crashed. Or i should test it with vrrpd (for example) and 
> restart secondary as ggated instead of being ggatec? Did not clear yet. 
> If you have the solution or an idea - could you please drop me a line?

Yes.
/usr/ports/net/freevrrpd

> algorythm looks the following for me:
> We have two computers:
> stage 0 (up and running claster)
> 1. da0 is exported via ggated over the network.
> This computer runing postgresql (for example) that storing data on da0.
> 2. da1 is being imported over network via ggatec as da1. Local da0 and 
> remote da1 are mirrored via ggated/geom_mirror.ko
> and there is no service activity on this device. just a mirroring of data.

No.
When you work with ggate{c,d} you get devices ggate{0,1,2,...}. Not with da1.

> stage 1 (slave failed)
> 1. working as before
> 2. down
> Requests to update remote data failed. There should be no problem.
> 
> stage 2 (slave is coming up)
> 1. working as before
> 2. booting and starting to mirror date from the master
> in case secondary goes down and up again there are no troubles. it just 
> boots itself in the stage0 config. that is it

When slave is start you can rebuild mirror on the fly.
But it get many time (in my case) -- 54Gb slice rebuilded slow :(

> stage 3 (master is down)
> 1. down
> 2. should consider master is down and stop ggatec, then start ggated and 
> export the da0. Then mount da0 and start PostgreSQL.

when master down, then freevrrpd can run script on slave machine.

> stage 4 (master coming up)
> 1. Should understand it is not the master anymore and start ggatec and 
> mirror remote data to the local drive.
> 2. running as a master. with services moved/started at stage 3
> Here troubles begins. When bouth become synchronized the service that 
> was running on this disk shoul be moved from 2 to 1. to do that i'll 
> have to restart services in stage 0. But how? using vrrpd? or some other 
> utility? How can i understand that the system is ready to get the 
> Postgresql back to 1 and stop it on 2?

Yes. vrrp -- your friend. But moving from node2 -> node1, when node1 coming up bad idea. Service must continue to run on node2.

> stage 5 (bouth down, master coming up first)
> 
> stage 6 (bouth down, slave is coming up first)

My algoritm.
No difference between master and slave in start time.
We have:
/etc/rc.conf -- this file identical on both nodes. Many parametes in it and 2 includes:
/etc/rc.clh -- specific addreses and specific params for every node.
/etc/rc.cluster -- simlink dinamically switching between:
/etc/rc.custer.master or /etc/rc.cluster.slave

If we can ping shared IP, when starting -- we start as slave. In not -- master.
We start freevrrpd with master or slave config (see /etc/rc.d/freevrrpd).
/etc/rc.d/freevrrpd run after network start, but before all daemons.
It work in RELENG_5* with new style rc-scripts.

/boot/loader.conf have different autoboot_delay on different nodes (bad guarantee what important start-ping not run in one time on both machines)

> This is main confuse. Maybe i did not read mutch, but it is not clear 
> for me, how to inderstand which drive contain the last data?

Yes. It's a problem in one case:
reboot node0 and reboot node1, when mirror not fully synchronize.

See attached scripts -- this my experiments. It have some known and, my be unknown issues:
 - in /usr/local/ifstated/bin/ggatec-gmirror.sh script I can't get unit name with one line:
ggate=`/sbin/ggatec ${ggatec_flags}`
 - when I restart /etc/rc.d/freevrrpd many time I get many ggate devices.
 - parametert for ggate{c,d} may be not ideal, but it work on gigabit interconnect.
 - I can't undestand what I must open on firewall for freevrrpd.

By.
Dmitriy 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cluster.tbz2
Type: application/octet-stream
Size: 3982 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-cluster/attachments/20050804/c2172224/cluster.obj


More information about the freebsd-cluster mailing list