Re: [Tarantool-patches] [PATCH v14 2/6] qsync: update confirmed lsn on initial promote request

[Serge Petrenko via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

13.09.2021 01:25, Cyrill Gorcunov пишет:
> On Sun, Sep 12, 2021 at 05:44:04PM +0200, Vladislav Shpilevoy wrote:
>> Thanks for the patch!
>>
>> On 10.09.2021 17:29, Cyrill Gorcunov wrote:
>>> When promote request is handled we drop last confirmed
>>> lsn to zero because its value make sense for sync queue
>>> owner only. Still the case where we become queue owner
>>> for the first time is special - we need to fetch the
>>> obtained lsn from the request and remember it so we
>>> will be able to filter any next malformed requests
>>> with wrong lsn numbers (see queue filtering procedure
>>> in next patch).
>> I don't understand anything. Why isn't it needed always? And
>> how exactly will it help to filter stuff?
> This problem is revealed when run of ./test-run replication/gh-6034-qsync-limbo-ownership.test.lua
> with filteration turned on. The confirmed_lsn make sence in bound
> with limbo owner as far as I understand. And in test we have
> two nodes "default" and "replica". Initially default gets up, filled
> with some data into sync space and then we start up a replica node.
> The replica get subscribed and then we call box.promote() on it,
> since replica itself has not been writting anything its confirmed_lsn = 0,
> which we send back to the "default" inside promote request body. And
> it get rejected because "default" has non-zero confirmed_lsn. I've
> been talking to Serge a lot about this problem and if I'm not missing
> somthing obvious updating this filed on first promote arrival is
> mostly correct way to handle the issue. I presume we should get
> a meeting and talk again.
>
> Or maybe better via email. Serge could you please write the details here?

It would be easier to discuss this verbally, I think.


-- 
Serge Petrenko