[Tarantool-patches] [PATCH 3/7] txn_limbo: fix promote term filtering

[Serge Petrenko]

19.06.2021 01:49, Vladislav Shpilevoy пишет:
> Hi! Thanks for the fixes!
>
>> 15.06.2021 23:57, Vladislav Shpilevoy пишет:
>>> Thanks for the patch!
>>>
>>> On 10.06.2021 15:32, Serge Petrenko via Tarantool-patches wrote:
>>>> txn_limbo_process() used to filter out promote requests whose term was
>>>> equal to the greatest term seen. This wasn't correct for PROMOTE entries
>>>> with term 1.
>>>>
>>>> Such entries appear after box.ctl.promote() is issued on an instance
>>>> with disabled elections. Every PROMOTE entry from such an instance has
>>>> term 1, but should still be applied. Fix this in the patch.
>>> Didn't we agree that PROMOTE should bump the term always? I see no purpose
>>> for the PROMOTE which never bumps the term except for the tests, but for
>>> such occasion it would be better to have internal.make_leader() or something
>>> which bypasses everything. Although the best option would to bump the term
>>> always.
>>>
>>> I see you did something in the last commit about that, but the part
>>> about `limbo->promote_greatest_term > 1` still remains. It looks very
>>> illegal to point out some terms as "special".
>> What about old instances?
>> They may issue multiple promotes for the same term when elections are off.
> What if they issue multiple promotes for the same term > 1? They could have
> elections enabled, then the term was bumped a few times, then elections were
> turned off again. Next promotes will use the current term which is > 1,
> won't they?

Yes, they will.

But such promotes must be filtered out because we can't  be sure if they 
were
issued with or without elections.

I still think we need some compatibility with old instances. At least 
for the case
when no elections were used at all.


>> Previous behaviour (wrong, because we forgot to filter out too old promotes at all)
>> made the instance apply all such promotes.
>>
>> I thought we should allow multiple promotes, at least for the default term (1), which
>> happen only when elections are turned off and have never been turned on.
>>
>> Otherwise, having a new instance together with the old one in the same cluster,
>> you'll have one of them apply all the promotes for one term, and have one dataset,
>> and the other ignore all but the first promote, and have possibly different data.> I did fix the issue in the last commit. All the new instances will issue promotes with
>> increasing term each time, and the term will be >= 2.
>>
>> So all the promotes with term == 1 surely will be from the old instances.

-- 
Serge Petrenko