[openstack-dev] [Neutron] DB: transaction isolation and related questions

Eugene Nikanorov enikanorov at mirantis.com
Fri Nov 21 08:24:58 UTC 2014


Comments inline:

On Thu, Nov 20, 2014 at 4:34 AM, Jay Pipes <jaypipes at gmail.com> wrote:

>
>>
> So while the SELECTs may return different data on successive calls when
> you use the READ COMMITTED isolation level, the UPDATE statements will
> continue to return 0 rows affected **if they attempt to change rows that
> have been changed since the start of the transaction**
>
> The reason that changing the isolation level to READ COMMITTED appears to
> work for the code in question:
>
> https://github.com/openstack/neutron/blob/master/neutron/
> plugins/ml2/drivers/helpers.py#L98
>
> is likely because the SELECT ... LIMIT 1 query is returning a different
> row on successive attempts (though since there is no ORDER BY on the query,
> the returned row of the query is entirely unpredictable (line 112)). Since
> data from that returned row is used in the UPDATE statement (line 118 and
> 124), *different* rows are actually being changed by successive UPDATE
> statements.
>

Not really, we're updating the same row we've selected. It's ensured
by 'raw_segment'
which actually contains 'gre_id' (or similar) attribute.
So in each iteration we're working with the same row, but in different
iterations READ COMMITTED allows us to see different data and hence work
with a different row.


>
> What this means is that for this *very particular* case, setting the
> transaction isolation level to READ COMMITTTED will work presumably most of
> the time on MySQL, but it's not an appropriate solution for the generalized
> problem domain of the SELECT FOR UPDATE. If you need to issue a SELECT and
> an UPDATE in a retry loop, and you are attempting to update the same row or
> rows (for instance, in the quota reservation or resource allocation
> scenarios), this solution will not work, even with READ COMMITTED. This is
> why I say it's not really appropriate, and a better general solution is to
> use separate transactions for each loop in the retry mechanic.


By saying 'this solution will not work' what issues do you mean what
exactly?
Btw, I agree on using separate transaction for each loop, the problem is
that transaction is usually not 'local' to the method where the retry loop
resides.



> The issue is about doing the retry within a single transaction. That's not
> what I recommend doing. I recommend instead doing short separate
> transactions instead of long-lived, multi-statement transactions and
> relying on the behaviour of the DB's isolation level (default or otherwise)
> to "solve" the problem of reading changes to a record that you intend to
> update.
>

" instead of long-lived, multi-statement transactions" - that's really what
would require quite large code redesign.
So far finding a way to bring retry logic upper to the stack of nesting
transactions seems more appropriate.

Thanks,
Eugene.

>
> Cheers,
> -jay
>
>  Also, thanks Clint for clarification about example scenario described by
>> Mike Bayer.
>> Initially the issue was discovered with concurrent tests on multi master
>> environment with galera as a DB backend.
>>
>> Thanks,
>> Eugene
>>
>> On Thu, Nov 20, 2014 at 12:20 AM, Mike Bayer <mbayer at redhat.com
>> <mailto:mbayer at redhat.com>> wrote:
>>
>>
>>      On Nov 19, 2014, at 3:47 PM, Ryan Moats <rmoats at us.ibm.com
>>>     <mailto:rmoats at us.ibm.com>> wrote:
>>>
>>>     >
>>>     BTW, I view your examples from oslo as helping make my argument for
>>>     me (and I don't think that was your intent :) )
>>>
>>>
>>     I disagree with that as IMHO the differences in producing MM in the
>>     app layer against arbitrary backends (Postgresql vs. DB2 vs. MariaDB
>>     vs. ???)  will incur a lot more “bifurcation” than a system that
>>     targets only a handful of existing MM solutions.  The example I
>>     referred to in oslo.db is dealing with distinct, non MM backends.
>>     That level of DB-specific code and more is a given if we are
>>     building a MM system against multiple backends generically.
>>
>>     It’s not possible to say which approach would be better or worse at
>>     the level of “how much database specific application logic do we
>>     need”, though in my experience, no matter what one is trying to do,
>>     the answer is always, “tons”; we’re dealing not just with databases
>>     but also Python drivers that have a vast amount of differences in
>>     behaviors, at every level.    On top of all of that, hand-rolled MM
>>     adds just that much more application code to be developed and
>>     maintained, which also claims it will do a better job than mature
>>     (ish?) database systems designed to do the same job against a
>>     specific backend.
>>
>>
>>
>>
>>>     > > My reason for asking this question here is that if the community
>>>     > > wants to consider #2, then these problems are the place to start
>>>     > > crafting that solution - if we solve the conflicts inherent
>>>     with the
>>>     > > two conncurrent thread scenarios, then I think we will find that
>>>     > > we've solved the multi-master problem essentially "for free”.
>>>     >
>>>     > Maybe I’m missing something, if we learn how to write out a row
>>> such
>>>     > that a concurrent transaction against the same row doesn’t throw us
>>>     > off, where is the part where that data is replicated to databases
>>>     > running concurrently on other IP numbers in a way that is atomic
>>>     > come out of that effort “for free” ?   A home-rolled “multi master”
>>>     > scenario would have to start with a system that has multiple
>>>     > create_engine() calls, since we need to communicate directly to
>>>     > multiple database servers. From there it gets really crazy.
>>>     Where’sall that ?
>>>
>>>     Boiled down, what you are talking about here w.r.t. concurrent
>>>     transactions is really conflict resolution, which is the hardest
>>>     part of implementing multi-master (as a side note, using locking in
>>>     this case is the equivalent of option #1).
>>>
>>>     All I wished to point out is that there are other ways to solve the
>>>     conflict resolution that could then be leveraged into a multi-master
>>>     scenario.
>>>
>>>     As for the parts that I glossed over, once conflict resolution is
>>>     separated out, replication turns into a much simpler problem with
>>>     well understood patterns and so I view that part as coming
>>>     "for free."
>>>
>>>     Ryan
>>>
>>>     _______________________________________________
>>>     OpenStack-dev mailing list
>>>     OpenStack-dev at lists.openstack.org
>>>     <mailto:OpenStack-dev at lists.openstack.org>
>>>     http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>
>>
>>
>>     _______________________________________________
>>     OpenStack-dev mailing list
>>     OpenStack-dev at lists.openstack.org
>>     <mailto:OpenStack-dev at lists.openstack.org>
>>     http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>
>>
>>
>>
>> _______________________________________________
>> OpenStack-dev mailing list
>> OpenStack-dev at lists.openstack.org
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>
>>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20141121/e0ad1920/attachment.html>


More information about the OpenStack-dev mailing list