On 09.11.2026 14:47, Dumitru Ceara wrote: > Внимание: ВНЕШНИЙ отправитель! > > > Будьте осторожны с вложениями и ссылками. > > > On 11/9/26 12:44 PM, Rukomoinikova Aleksandra wrote: >> On 09.11.2026 14:04, Dumitru Ceara wrote: >>> On 11/9/26 12:01 PM, Dumitru Ceara wrote: >>>> On 11/9/26 10:17 AM, Rukomoinikova Aleksandra wrote: >>>>> On 07.11.2026 00:33, Dumitru Ceara wrote: >>>>>> On 9/28/26 11:45 AM, Alexandra Rukomoinikova via dev wrote: >>>>>>> Several ovn-controller instances may run with the same chassis name and >>>>>>> the same encapsulation IP, an anycast address that is reachable over the >>>>>>> underlay through ECMP. OVN then sees the whole group of nodes as a >>>>>>> single chassis: a single Chassis record, a single Encap record and a >>>>>>> single binding for every port placed on it, while the underlay decides >>>>>>> which node of the group terminates a given tunnel flow. >>>>>>> >>>>>>> There is no coordination and no leader election between the nodes of the >>>>>>> group while writing to SBDB: every ``ovn-controller`` independently >>>>>>> does the >>>>>>> full bookkeeping for that one chassis, and it is the idempotence of >>>>>>> those >>>>>>> writes that keeps them from fighting each other. >>>>>>> >>>>>>> The flip side is that the nodes must be configured identically in >>>>>>> chassis-related params, and that nothing which relies on conntrack can >>>>>>> be >>>>>>> used, because the group shares no runtime state. >>>>>>> >>>>>>> Document all of this, and add a multinode test that deploys two >>>>>>> gateway nodes as one replicated gateway, gives a hypervisor and a node >>>>>>> that >>>>>>> does not run OVN an underlay ECMP route to the anycast encapsulation IP, >>>>>>> checks the north-south connectivity through each of the two nodes and >>>>>>> checks that the controllers do not fight over the shared records. >>>>>>> >>>>>>> Signed-off-by: Alexandra Rukomoinikova <[email protected]> >>>>>>> --- >>>>>> Hi Alexandra, >>>>>> >>>>>> I have a few minor comments on this patch, please see below. >>>>>> >>>>>> I was thinking of taking care of them myself and squash a small fixup >>>>>> patch in when pushing yours to main: >>>>>> >>>>>> https://github.com/dceara/ovn/commit/42b64d5 >>>>>> >>>>>> I didn't test this locally but CI is running in my fork: >>>>>> >>>>>> https://github.com/dceara/ovn/actions/runs/42295260684 >>>>>> >>>>>> Assuming that passes would you be OK with this incremental change? >>>>>> >>>>>> Regards, >>>>>> Dumitru >>>> Hi Sasha, >>>> >>>>> Hi Dumitru! I took a fresh look at this patch, and I’d like to remove >>>>> all references to ECMP → anycast IP and drop the mention of using BGP >>>>> for announcements - this seems unnecessary. Is it still okay to make >>>> I actually liked the ECMP reference, maybe it's an idea to still mention >>>> that and BGP as an example? I'm not against removing it though. >>>> >>> While at it, maybe it's worth also mentioning that other (stateful) >>> features might not work as expected in such deployments. E.g., IGMP. >>> >>>>> those changes rn? I can take the fixup based on fixups in your fork, >>>>> make the changes on top of it, and send you the updated patch. 🙂 >>>>> >>>> I haven't merged your patch yet because CI didn't pass, one of the lanes >>>> still hangs and I didn't have time to look into it: >>>> >>>> https://github.com/dceara/ovn/actions/runs/42295260684/job/129055547103 >>>> >>>> So if you could take the fixup from my fork and work on a v2 that'd be >>>> great. >>>> >>>> Thanks, >>>> Dumitru >> I’ll look into why the test is hanging and send a v2. >> >> Regarding BGP, I meant this part: >> >>>> /+ In a real deployment the route is normally learned from the fabric, >> with the />>/+ gateway nodes advertising the anycast prefix over BGP./ >> >> I’d like to fix this part and remove the mention of BGP from it. >> >> If you’re liked with the ECMP part, I’ll leave it as it is. >> >> In that case, I’ll just go through the test and remove BGP from the >> example above. >> > Sounds good to me, thanks! Hi Dumitru! The test issue was that OVS was stopping before ovn-controller. With INLINE_BR_EX=yes, that cut off the node’s connection to the SB database, so the deletion of the old chassis record have not reached database in right time. controller was then terminated with signal. When it came back up, it would usually finish committing its pending changes and delete the record. Sometimes that didn’t happen, likely because the controller received updates from the server first.
The test would then hang waiting for hv_cfg to sync because the chassis record from the old default controller was still there under its original name, from before the rename to replicated. I made a separate commit that changes the shutdown order in the tests so ovn-controller stops before OVS. I ran the test 10 times in my fork just to check everything is ok, and it passed every time. https://github.com/Sashhkaa/ovn/actions/runs/42721235978/job/129518340143 I also made a separate fixup commit that removes the BGP references from the docs and replaces the remaining vgv names in the tests with replicated. I’m including links to the fork and the fixup commit in case it’s easier to review the changes there than in v2 that i have sended. https://github.com/Sashhkaa/ovn/commit/6a9bc030c74ee048bbc3d6227d8797f90e834402 Thanks a lot for the review! thanks, sasha >> thanks, sasha >> > Внимание: ВНЕШНИЙ отправитель! > > > Будьте осторожны с вложениями и ссылками. _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
