Re: [Tarantool-patches] [PATCH vshard 6/6] router: update master using a hint from storage

[Oleg Babin via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Thanks for your answers.

I have following questions:

   - After changes in previous patch I think it could be better to use 
"master_search_wakeup" instead of master_search_fiber:wakeup()

   - Should we document somehow `update_master` method for replicaset 
object? Or maybe it even shouldn't be a part of public API?

On 03.07.2021 00:36, Vladislav Shpilevoy wrote:
> Thanks for the review!
>
>>> diff --git a/vshard/error.lua b/vshard/error.lua
>>> index e2d1a31..fa7bdaa 100644
>>> --- a/vshard/error.lua
>>> +++ b/vshard/error.lua
>>> @@ -20,7 +20,7 @@ local error_message_template = {
>>>        [2] = {
>>>            name = 'NON_MASTER',
>>>            msg = 'Replica %s is not a master for replicaset %s anymore',
>>> -        args = {'replica_uuid', 'replicaset_uuid'}
>>> +        args = {'replica_uuid', 'replicaset_uuid', 'master_uuid'}
>> Error format string still contains only 2 arguments. Is it ok?
> Yes. All the 'args' are remembered as the error object attributes.
> I can use it to attach information which I don't want to add to
> the message or simply can't.
>
> Here I can't, because master_uuid might be nil. If it would be in
> the format, it would be displayed as 'nil' sometimes, which would be
> confusing.
>
>>> diff --git a/vshard/replicaset.lua b/vshard/replicaset.lua
>>> index 7747258..660c786 100644
>>> --- a/vshard/replicaset.lua
>>> +++ b/vshard/replicaset.lua
>>> @@ -570,6 +570,70 @@ local function rebind_replicasets(replicasets, old_replicasets)
>>>        end
>>>    end
>>>    +--
>>> +-- Let the replicaset know @a old_master_uuid is not a master anymore, should
>>> +-- use @a candidate_uuid instead.
>>> +-- Returns whether the request, which brought this information, can be retried.
>>> +--
>>> +local function replicaset_update_master(replicaset, old_master_uuid,
>>> +                                        candidate_uuid)
>>> +    local is_auto = replicaset.is_auto_master
>>> +    local replicaset_uuid = replicaset.uuid
>>> +    if old_master_uuid == candidate_uuid then
>>> +        -- It should not happen ever, but be ready to everything.
>>> +        log.warn('Replica %s in replicaset %s reports self as both master '..
>>> +                 'and not master', master_uuid, replicaset_uuid)
>>> +        return is_auto
>>> +    end
>>> +    local master = replicaset.master
>>> +    if not master then
>>> +        if not is_auto or not candidate_uuid then
>>> +            return is_auto
>>> +        end
>>> +        local candidate = replicaset.replicas[candidate_uuid]
>> AFAIU it means that candidate_uuid doesn't belong to replicaset.
>>
>> Why is it true?
> It can happen if router config is different from the storage's due
> to any reason. If the master is not in router's config, it won't be
> able to find it by UUID.