Re: [PATCH] Add `truncate` option to subscription commands
От | David Christensen |
---|---|
Тема | Re: [PATCH] Add `truncate` option to subscription commands |
Дата | |
Msg-id | C88FC8CB-2EC9-41CD-AA25-AAF9C61266D7@endpoint.com обсуждение исходный текст |
Ответ на | Re: [PATCH] Add `truncate` option to subscription commands (Amit Kapila <amit.kapila16@gmail.com>) |
Ответы |
Re: [PATCH] Add `truncate` option to subscription commands
(Rahila Syed <rahilasyed90@gmail.com>)
|
Список | pgsql-hackers |
> On Oct 11, 2020, at 10:00 PM, Amit Kapila <amit.kapila16@gmail.com> wrote: > > On Mon, Oct 12, 2020 at 3:44 AM David Christensen <david@endpoint.com> wrote: >> >> >> On Oct 11, 2020, at 1:14 PM, Euler Taveira <euler.taveira@2ndquadrant.com> wrote: >> >> >> On Fri, 9 Oct 2020 at 15:54, David Christensen <david@endpoint.com> wrote: >>> >>> >>> Enclosed find a patch to add a “truncate” option to subscription commands. >>> >>> When adding new tables to a subscription (either via `CREATE SUBSCRIPTION` or `REFRESH PUBLICATION`), tables on the targetwhich are being newly subscribed will be truncated before the data copy step. This saves explicit coordination ofa manual `TRUNCATE` on the target tables and allows the results of the initial data sync to be the same as on the publisherat the time of sync. >>> >>> To preserve compatibility with existing behavior, the default value for this parameter is `false`. >>> >> >> Truncate will fail for tables whose foreign keys refer to it. If such a feature cannot handle foreign keys, the usefulnesswill be restricted. >> >> >> This is true for existing “truncate” with FKs, so doesn’t seem to be any different to me. >> > > What would happen if there are multiple tables and truncate on only > one of them failed due to FK check? Does it give an error in such a > case, if so will the other tables be truncated? Currently each SyncRep relation is sync’d separately in its own worker process; we are doing the truncate at the initializationstep of this, so it’s inherently in its own transaction. I think if we are going to do any sort of validationon this, it would have to be at the point of the CREATE SUBSCRIPTION/REFRESH PUBLICATION where we have the relationlist and can do sanity-checking there. Obviously if someone changes the schema at some point between when it does this and when relation syncs start there is arace condition, but the same issue would affect other data sync things, so I don’t care to solve that as part of this patch. >> Hypothetically if you checked all new tables and could verify if there were FK cycles only already in the new tables beingadded then “truncate cascade” would be fine. Arguably if they had existing tables that were part of an FK that wasn’tfully replicated they were already operating brokenly. >> > > I think if both PK_table and FK_table are part of the same > subscription then there should be a problem as both them first get > truncated? If they are part of a different subscription (or if they > are not subscribed due to whatever reason) then probably we need to > deal such cases carefully. You mean “should not be a problem” here? If so, I agree with that. Obviously if we determine this features is only usefulwith this support we’d have to chase the entire dependency graph and make sure that is all contained in the set ofnewly-subscribed tables (or at least FK referents). I have not considered tables that are part of more than one subscription (is that possible?); we presumably should errorout if any table exists already in a separate subscription, as we’d want to avoid truncating tables already part ofan existing subscription. While I’m happy to take a stab at fixing some of the FK/PK issues, it seems easy to go down a rabbit hole. I’m not convincedthat we couldn’t just detect FK issues and choose to not handle this case without decreasing the utility for atleast some cases. (Perhaps we could give a hint as to the issues detected to point someone in the right direction.) Anyway,glad to keep discussing potential implications, etc. Best, David -- David Christensen Senior Software and Database Engineer End Point Corporation david@endpoint.com 785-727-1171
Вложения
В списке pgsql-hackers по дате отправления:
Предыдущее
От: Justin PryzbyДата:
Сообщение: Re: pg_upgrade: fail early if a tablespace dir already exists for new cluster version
Следующее
От: Tom LaneДата:
Сообщение: Re: Feature improvement: can we add queryId for pg_catalog.pg_stat_activity view?