Date: Thu, 18 Aug 2016 12:19:55 -0500 From: Linda Kateley <lkateley@kateley.com> To: juergen.gotteswinter@internetx.com, linda@kateley.com, Chris Watson <bsdunix44@gmail.com> Cc: freebsd-fs@freebsd.org Subject: Re: HAST + ZFS + NFS + CARP Message-ID: <4c34cbf9-84b5-5d42-e0b4-bf18aa1ef9a7@kateley.com> In-Reply-To: <7468cc18-85e8-3765-2b2b-a93ef73ca05a@internetx.com> References: <61283600-A41A-4A8A-92F9-7FAFF54DD175@ixsystems.com> <20160704183643.GI41276@mordor.lan> <AE372BF0-02BE-4BF3-9073-A05DB4E7FE34@ixsystems.com> <20160704193131.GJ41276@mordor.lan> <E7D42341-D324-41C7-B03A-2420DA7A7952@sarenet.es> <20160811091016.GI70364@mordor.lan> <1AA52221-9B04-4CF6-97A3-D2C2B330B7F9@sarenet.es> <472bc879-977f-8c4c-c91a-84cc61efcd86@internetx.com> <20160817085413.GE22506@mordor.lan> <465bdec5-45b7-8a1d-d580-329ab6d4881b@internetx.com> <20160817095222.GG22506@mordor.lan> <52d5b687-1351-9ec5-7b67-bfa0be1c8415@kateley.com> <92F4BE3D-E4C1-4E5C-B631-D8F124988A83@gmail.com> <6b866b6e-1ab3-bcc5-151b-653e401742bd@kateley.com> <7468cc18-85e8-3765-2b2b-a93ef73ca05a@internetx.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On 8/18/16 2:32 AM, InterNetX - Juergen Gotteswinter wrote: > > Am 17.08.2016 um 20:03 schrieb Linda Kateley: >> I just do consulting so I don't always get to see the end of the >> project. Although we are starting to do more ongoing support so we can >> see the progress.. >> >> I have worked with some of the guys from high-availability.com for maybe >> 20 years. RSF-1 is the cluster that is bundled with nexenta. Does work >> beautifully with omni/illumos. The one customer I have running it in >> prod is an isp in south america running openstack and zfs on freebsd as >> iscsi. Big boxes, 90+ drives per frame. If someone would like try it, i >> have some contacts there. Ping me offlist. > no offense, but it sounds a bit like marketing. > > here: running nexenta ha setup since several years with one catastrophic > failure due to split brain Just trying to say I don't see projects ongoing.. just at beginning > >> You do risk losing data if you batch zfs send. It is very hard to run >> that real time. > depends on how much data changes aka delta size > > > You have to take the snap then send the snap. Most >> people run in cron, even if it's not in cron, you would want one to >> finish before you started the next. > thats the reason why lock files where invented, tools like zrep handle > that themself via additional zfs properties > > or, if one does not trust a single layer > > -- snip -- > #!/bin/sh > if [ ! -f /var/run/replic ] ; then > touch /var/run/replic > /blah/path/zrep sync all >> /var/log/zfsrepli.log > rm -f /var/run/replic > fi > -- snip -- > > something like this, simple > > If you lose the sending host before >> the receive is complete you won't have a full copy. > if rsf fails, and you end up in split brain you loose way more. been > there, seen that. > > With zfs though you >> will probably still have the data on the sending host, however long it >> takes to bring it back up. RSF-1 runs in the zfs stack and send the >> writes to the second system. It's kind of pricey, but actually much less >> expensive than commercial alternatives. >> >> Anytime you run anything sync it adds latency but makes things safer.. > not surprising, it all depends on the usecase > >> There is also a cool tool I like, called zerto for vmware that sits in >> the hypervisor and sends a sync copy of a write locally and then an >> async remotely. It's pretty cool. Although I haven't run it myself, have >> a bunch of customers running it. I believe it works with proxmox too. >> >> Most people I run into (these days) don't mind losing 5 or even 30 >> minutes of data. Small shops. > you talk about minutes, what delta size are we talking here about? why > not using zrep in a loop for example > > They usually have a copy somewhere else. >> Or the cost of 5-30 minutes isn't that great. I used work as a >> datacenter architect for sun/oracle with only fortune 500. There losing >> 1 sec could put large companies out of business. I worked with banks and >> exchanges. > again, usecase. i bet 99% on this list are not operating fortune 500 > bank filers > > They couldn't ever lose a single transaction. Most people >> nowadays do the replication/availability in the application though and >> don't care about underlying hardware, especially disk. >> >> >> On 8/17/16 11:55 AM, Chris Watson wrote: >>> Of course, if you are willing to accept some amount of data loss that >>> opens up a lot more options. :) >>> >>> Some may find that acceptable though. Like turning off fsync with >>> PostgreSQL to get much higher throughput. As little no as you are made >>> *very* aware of the risks. >>> >>> It's good to have input in this thread from one with more experience >>> with RSF-1 than the rest of us. You confirm what others have that said >>> about RSF-1, that it's stable and works well. What were you deploying >>> it on? >>> >>> Chris >>> >>> Sent from my iPhone 5 >>> >>> On Aug 17, 2016, at 11:18 AM, Linda Kateley <lkateley@kateley.com >>> <mailto:lkateley@kateley.com>> wrote: >>> >>>> The question I always ask, as an architect, is "can you lose 1 minute >>>> worth of data?" If you can, then batched replication is perfect. If >>>> you can't.. then HA. Every place I have positioned it, rsf-1 has >>>> worked extremely well. If i remember right, it works at the dmu. I >>>> would suggest try it. They have been trying to have a full freebsd >>>> solution, I have several customers running it well. >>>> >>>> linda >>>> >>>> >>>> On 8/17/16 4:52 AM, Julien Cigar wrote: >>>>> On Wed, Aug 17, 2016 at 11:05:46AM +0200, InterNetX - Juergen >>>>> Gotteswinter wrote: >>>>>> Am 17.08.2016 um 10:54 schrieb Julien Cigar: >>>>>>> On Wed, Aug 17, 2016 at 09:25:30AM +0200, InterNetX - Juergen >>>>>>> Gotteswinter wrote: >>>>>>>> Am 11.08.2016 um 11:24 schrieb Borja Marcos: >>>>>>>>>> On 11 Aug 2016, at 11:10, Julien Cigar <julien@perdition.city >>>>>>>>>> <mailto:julien@perdition.city>> wrote: >>>>>>>>>> >>>>>>>>>> As I said in a previous post I tested the zfs send/receive >>>>>>>>>> approach (with >>>>>>>>>> zrep) and it works (more or less) perfectly.. so I concur in >>>>>>>>>> all what you >>>>>>>>>> said, especially about off-site replicate and synchronous >>>>>>>>>> replication. >>>>>>>>>> >>>>>>>>>> Out of curiosity I'm also testing a ZFS + iSCSI + CARP at the >>>>>>>>>> moment, >>>>>>>>>> I'm in the early tests, haven't done any heavy writes yet, but >>>>>>>>>> ATM it >>>>>>>>>> works as expected, I havent' managed to corrupt the zpool. >>>>>>>>> I must be too old school, but I don’t quite like the idea of >>>>>>>>> using an essentially unreliable transport >>>>>>>>> (Ethernet) for low-level filesystem operations. >>>>>>>>> >>>>>>>>> In case something went wrong, that approach could risk >>>>>>>>> corrupting a pool. Although, frankly, >>>>>>>>> ZFS is extremely resilient. One of mine even survived a SAS HBA >>>>>>>>> problem that caused some >>>>>>>>> silent corruption. >>>>>>>> try dual split import :D i mean, zpool -f import on 2 machines >>>>>>>> hooked up >>>>>>>> to the same disk chassis. >>>>>>> Yes this is the first thing on the list to avoid .. :) >>>>>>> >>>>>>> I'm still busy to test the whole setup here, including the >>>>>>> MASTER -> BACKUP failover script (CARP), but I think you can prevent >>>>>>> that thanks to: >>>>>>> >>>>>>> - As long as ctld is running on the BACKUP the disks are locked >>>>>>> and you can't import the pool (even with -f) for ex (filer2 is the >>>>>>> BACKUP): >>>>>>> https://gist.github.com/silenius/f9536e081d473ba4fddd50f59c56b58f >>>>>>> >>>>>>> - The shared pool should not be mounted at boot, and you should >>>>>>> ensure >>>>>>> that the failover script is not executed during boot time too: >>>>>>> this is >>>>>>> to handle the case wherein both machines turn off and/or re-ignite at >>>>>>> the same time. Indeed, the CARP interface can "flip" it's status >>>>>>> if both >>>>>>> machines are powered on at the same time, for ex: >>>>>>> https://gist.github.com/silenius/344c3e998a1889f988fdfc3ceba57aaf and >>>>>>> you will have a split-brain scenario >>>>>>> >>>>>>> - Sometimes you'll need to reboot the MASTER for some $reasons >>>>>>> (freebsd-update, etc) and the MASTER -> BACKUP switch should not >>>>>>> happen, this can be handled with a trigger file or something like >>>>>>> that >>>>>>> >>>>>>> - I've still have to check if the order is OK, but I think that as >>>>>>> long >>>>>>> as you shutdown the replication interface and that you adapt the >>>>>>> advskew (including the config file) of the CARP interface before the >>>>>>> zpool import -f in the failover script you can be relatively >>>>>>> confident >>>>>>> that nothing will be written on the iSCSI targets >>>>>>> >>>>>>> - A zpool scrub should be run at regular intervals >>>>>>> >>>>>>> This is my MASTER -> BACKUP CARP script ATM >>>>>>> https://gist.github.com/silenius/7f6ee8030eb6b923affb655a259bfef7 >>>>>>> >>>>>>> Julien >>>>>>> >>>>>> 100€ question without detailed looking at that script. yes from a >>>>>> first >>>>>> view its super simple, but: why are solutions like rsf-1 such more >>>>>> powerful / featurerich. Theres a reason for, which is that they try to >>>>>> cover every possible situation (which makes more than sense for this). >>>>> I've never used "rsf-1" so I can't say much more about it, but I have >>>>> no doubts about it's ability to handle "complex situations", where >>>>> multiple nodes / networks are involved. >>>>> >>>>>> That script works for sure, within very limited cases imho >>>>>> >>>>>>>> kaboom, really ugly kaboom. thats what is very likely to happen >>>>>>>> sooner >>>>>>>> or later especially when it comes to homegrown automatism solutions. >>>>>>>> even the commercial parts where much more time/work goes into such >>>>>>>> solutions fail in a regular manner >>>>>>>> >>>>>>>>> The advantage of ZFS send/receive of datasets is, however, that >>>>>>>>> you can consider it >>>>>>>>> essentially atomic. A transport corruption should not cause >>>>>>>>> trouble (apart from a failed >>>>>>>>> "zfs receive") and with snapshot retention you can even roll >>>>>>>>> back. You can’t roll back >>>>>>>>> zpool replications :) >>>>>>>>> >>>>>>>>> ZFS receive does a lot of sanity checks as well. As long as your >>>>>>>>> zfs receive doesn’t involve a rollback >>>>>>>>> to the latest snapshot, it won’t destroy anything by mistake. >>>>>>>>> Just make sure that your replica datasets >>>>>>>>> aren’t mounted and zfs receive won’t complain. >>>>>>>>> >>>>>>>>> >>>>>>>>> Cheers, >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> Borja. >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> _______________________________________________ >>>>>>>>> freebsd-fs@freebsd.org <mailto:freebsd-fs@freebsd.org> mailing list >>>>>>>>> https://lists.freebsd.org/mailman/listinfo/freebsd-fs >>>>>>>>> To unsubscribe, send any mail to >>>>>>>>> "freebsd-fs-unsubscribe@freebsd.org >>>>>>>>> <mailto:freebsd-fs-unsubscribe@freebsd.org>" >>>>>>>>> >>>>>>>> _______________________________________________ >>>>>>>> freebsd-fs@freebsd.org <mailto:freebsd-fs@freebsd.org> mailing list >>>>>>>> https://lists.freebsd.org/mailman/listinfo/freebsd-fs >>>>>>>> To unsubscribe, send any mail to >>>>>>>> "freebsd-fs-unsubscribe@freebsd.org >>>>>>>> <mailto:freebsd-fs-unsubscribe@freebsd.org>" >>>> _______________________________________________ >>>> freebsd-fs@freebsd.org <mailto:freebsd-fs@freebsd.org> mailing list >>>> https://lists.freebsd.org/mailman/listinfo/freebsd-fs >>>> To unsubscribe, send any mail to "freebsd-fs-unsubscribe@freebsd.org >>>> <mailto:freebsd-fs-unsubscribe@freebsd.org>" >> _______________________________________________ >> freebsd-fs@freebsd.org mailing list >> https://lists.freebsd.org/mailman/listinfo/freebsd-fs >> To unsubscribe, send any mail to "freebsd-fs-unsubscribe@freebsd.org"
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4c34cbf9-84b5-5d42-e0b4-bf18aa1ef9a7>