[Tarantool-patches] [PATCH v2 10/11] box: enrich ER_READONLY with new details

[Serge Petrenko]

13.11.2021 02:24, Vladislav Shpilevoy пишет:
> Hi! Thanks for the review!
>
>> 12.11.2021 02:54, Vladislav Shpilevoy пишет:
>>> ER_READONLY used not to have any details about the exact reason
>>> why the instance is read-only. The patch changes that by adding
>>> new fields into the error which explain why the error happened and
>>> even help to avoid it for next requests.
>>>
>> Thanks for the changes!
>>
>> Please, find two comments below.
>>
>> Sorry for coming up late with this one, but I think it'd be good
>> to report ro reason in box.info. Maybe box.info.ro_reason or
>> something similar. Only when box.info.ro is true, of course.
>>
>> Otherwise we help the user only partially. He sees what's wrong when receives
>> an error, but has to check every parameter manually when checks box.info.ro
> Sounds useful, I added it in a new commit on top of the branch. See a new
> email in the same thread.
>
>>> diff --git a/src/box/box.cc b/src/box/box.cc
>>> index 1ed1ce3f8..323982969 100644
>>> --- a/src/box/box.cc
>>> +++ b/src/box/box.cc
>>> @@ -177,16 +177,54 @@ box_update_ro_summary(void)
>>>    static int
>>>    box_check_writable(void)
>>>    {
>>> -    if (is_ro_summary) {
>>> +    if (!is_ro_summary)
>>> +        return 0;
>>> +    struct error *e = diag_set(ClientError, ER_READONLY);
>>> +    struct raft *raft = box_raft();
>>> +    /*
>>> +     * In case of multiple reasons at the same time only one is reported.
>>> +     * But the order is important. For example, if the instance has election
>>> +     * enabled, for the client it is better to see that it is a 'follower'
>>> +     * and who is the leader than just see cfg 'read_only' is true.
>>> +     */
>>> +    if (raft_is_ro(raft)) {
>>> +        error_set_str(e, "reason", "election");
>>> +        error_set_str(e, "state", raft_state_str(raft->state));
>>> +        error_set_uint(e, "term", raft->volatile_term);
>>> +        uint32_t id = raft->leader;
>>> +        if (id != REPLICA_ID_NIL) {
>>> +            error_set_uint(e, "leader_id", id);
>>> +            struct replica *r = replica_by_id(id);
>>> +            /*
>>> +             * XXX: when the leader is dropped from _cluster, it
>>> +             * is not reported to Raft.
>>> +             */
>>> +            if (r != NULL)
>>> +                error_set_uuid(e, "leader_uuid", &r->uuid);
>>> +        }
>>> +    } else if (txn_limbo_is_ro(&txn_limbo)) {
>>> +        error_set_str(e, "reason", "synchro");
>>> +        uint32_t id = txn_limbo.owner_id;
>>> +        error_set_uint(e, "queue_owner_id", id);
>>> +        error_set_uint(e, "term", raft->volatile_term);
>> I just noticed, we should report txn_limbo_greatest_term here, probably.
>>
>> This instance (which received ER_READONLY) is the leader, but hasn't claimed
>> the limbo yet.
>> This only makes sense when limbo term is behind raft's one.
> Hmm, you are probably right. Thanks for noticing! Fixed in the
> previous commit.
>
> ====================
> @@ -214,7 +214,7 @@ box_check_writable(void)
>   	} else if (txn_limbo_is_ro(&txn_limbo)) {
>   		error_set_str(e, "reason", "synchro");
>   		uint32_t id = txn_limbo.owner_id;
> -		uint64_t term = raft->volatile_term;
> +		uint64_t term = txn_limbo.promote_greatest_term;
>   		error_set_uint(e, "queue_owner_id", id);
>   		error_set_uint(e, "term", term);
> ====================

Thanks for the changes!

One final note:
I propose to move box_ro_reason() from the last commit here and use it
when setting error reason, so that error.reason would be the same as
box.info.ro_reason.


LGTM otherwise.

-- 
Serge Petrenko