Re: Global snapshots

On 2020/10/23 11:58, Masahiko Sawada wrote:
> On Thu, 15 Oct 2020 at 01:41, Fujii Masao <masao.fujii@oss.nttdata.com> wrote:
>>
>>
>>
>> On 2020/09/17 15:56, Amit Kapila wrote:
>>> On Thu, Sep 10, 2020 at 4:20 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:
>>>>
>>>>>> One alternative is to add only hooks into PostgreSQL core so that we can
>>>>>> implement the global transaction management outside. This idea was
>>>>>> discussed before as the title "eXtensible Transaction Manager API".
>>>>>
>>>>> Yeah, I read that discussion.  And I remember Robert Haas and Postgres Pro people said it's not good...
>>>>
>>>> But it may be worth revisiting this idea if we cannot avoid the patent issue.
>>>>
>>>
>>> It is not very clear what exactly we can do about the point raised by
>>> Tsunakawa-San related to patent in this technology as I haven't seen
>>> that discussed during other development but maybe we can try to study
>>> a bit. One more thing I would like to bring here is that it seems to
>>> be there have been some concerns about this idea when originally
>>> discussed [1]. It is not very clear to me if all the concerns are
>>> addressed or not. If one can summarize the concerns discussed and how
>>> the latest patch is able to address those then it will be great.
>>
>> I have one concern about Clock-SI (sorry if this concern was already
>> discussed in the past). As far as I read the paper about Clock-SI, ISTM that
>> Tx2 that starts after Tx1's commit can fail to see the results by Tx1,
>> due to the clock skew. Please see the following example;
>>
>> 1. Tx1 starts at the server A.
>>
>> 2. Tx1 writes some records at the server A.
>>
>> 3. Tx1 gets the local clock 20, uses 20 as CommitTime, then completes
>>        the commit at the server A.
>>        This means that Tx1 is the local transaction, not distributed one.
>>
>> 4. Tx2 starts at the server B, i.e., the server B works as
>>        the coordinator node for Tx2.
>>
>> 5. Tx2 gets the local clock 10 (i.e., it's delayed behind the server A
>>        due to clock skew) and uses 10 as SnapshotTime at the server B.
>>
>> 6. Tx2 starts the remote transaction at the server A with SnapshotTime 10.
>>
>> 7. Tx2 doesn't need to wait due to clock skew because the imported
>>        SnapshotTime 10 is smaller than the local clock at the server A.
>>
>> 8. Tx2 fails to see the records written by Tx1 at the server A because
>>        Tx1's CommitTime 20 is larger than SnapshotTime 10.
>>
>> So Tx1 was successfully committed before Tx2 starts. But, at the above example,
>> the subsequent transaction Tx2 fails to see the committed results.
>>
>> The single PostgreSQL instance seems to guarantee that linearizability of
>> the transactions, but Clock-SI doesn't in the distributed env. Is this my
>> understanding right? Or am I missing something?
>>
>> If my understanding is right, shouldn't we address that issue when using
>> Clock-SI? Or the patch has already addressed the issue?
> 
> As far as I read the paper, the above scenario can happen. I could
> reproduce the above scenario with the patch. Moreover, a stale read
> could happen even if Tx1 was initiated at server B (i.g., both
> transactions started at the same server in sequence). In this case,
> Tx1's commit timestamp would be 20 taken from server A's local clock
> whereas Tx2's snapshot timestamp would be 10 same as the above case.
> Therefore, in spite of both transactions were initiated at the same
> server the linearizability is not provided.

Yeah, so if we need to guarantee the transaction linearizability even
in distributed env (probably this is yes. Right?), using only Clock-SI
is not enough. We would need to implement something more
in addition to Clock-SI or adopt the different approach other than Clock-SI.
Thought?

Regards,

-- 
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION