On Fri, 11 Sep 2020 at 11:58, Fujii Masao <masao.fu...@oss.nttdata.com> wrote: > > > > On 2020/09/11 0:37, Masahiko Sawada wrote: > > On Tue, 8 Sep 2020 at 13:00, tsunakawa.ta...@fujitsu.com > > <tsunakawa.ta...@fujitsu.com> wrote: > >> > >> From: Amit Kapila <amit.kapil...@gmail.com> > >>> I intend to say that the global-visibility work can impact this in a > >>> major way and we have analyzed that to some extent during a discussion > >>> on the other thread. So, I think without having a complete > >>> design/solution that addresses both the 2PC and global-visibility, it > >>> is not apparent what is the right way to proceed. It seems to me that > >>> rather than working on individual (or smaller) parts one needs to come > >>> up with a bigger picture (or overall design) and then once we have > >>> figured that out correctly, it would be easier to decide which parts > >>> can go first. > >> > >> I'm really sorry I've been getting late and late and latex10 to publish > >> the revised scale-out design wiki to discuss the big picture! I don't > >> know why I'm taking this long time; I feel I were captive in a time prison > >> (yes, nobody is holding me captive; I'm just late.) Please wait a few > >> days. > >> > >> But to proceed with the development, let me comment on the atomic commit > >> and global visibility. > >> > >> * We have to hear from Andrey about their check on the possibility that > >> Clock-SI could be Microsoft's patent and if we can avoid it. > >> > >> * I have a feeling that we can adopt the algorithm used by Spanner, > >> CockroachDB, and YugabyteDB. That is, 2PC for multi-node atomic commit, > >> Paxos or Raft for replica synchronization (in the process of commit) to > >> make 2PC more highly available, and the timestamp-based global visibility. > >> However, the timestamp-based approach makes the database instance shut > >> down when the node's clock is distant from the other nodes. > >> > >> * Or, maybe we can use the following Commitment ordering that doesn't > >> require the timestamp or any other information to be transferred among the > >> cluster nodes. However, this seems to have to track the order of read and > >> write operations among concurrent transactions to ensure the correct > >> commit order, so I'm not sure about the performance. The MVCO paper seems > >> to present the information we need, but I haven't understood it well yet > >> (it's difficult.) Could you anybody kindly interpret this? > >> > >> Commitment ordering (CO) - yoavraz2 > >> https://sites.google.com/site/yoavraz2/the_principle_of_co > >> > >> > >> As for the Sawada-san's 2PC patch, which I find interesting purely as FDW > >> enhancement, I raised the following issues to be addressed: > >> > >> 1. Make FDW API implementable by other FDWs than postgres_fdw (this is > >> what Amit-san kindly pointed out.) I think oracle_fdw and jdbc_fdw would > >> be good examples to consider, while MySQL may not be good because it > >> exposes the XA feature as SQL statements, not C functions as defined in > >> the XA specification. > > > > I agree that we need to verify new FDW APIs will be suitable for other > > FDWs than postgres_fdw as well. > > > >> > >> 2. 2PC processing is queued and serialized in one background worker. That > >> severely subdues transaction throughput. Each backend should perform 2PC. > > > > Not sure it's safe that each backend perform PREPARE and COMMIT > > PREPARED since the current design is for not leading an inconsistency > > between the actual transaction result and the result the user sees. > > Can I check my understanding about why the resolver process is necessary? > > Firstly, you think that issuing COMMIT PREPARED command to the foreign server > can cause an error, for example, because of connection error, OOM, etc. On > the other hand, only waiting for other process to issue the command is less > likely to cause an error. Right? > > If an error occurs in backend process after commit record is WAL-logged, the > error would be reported to the client and it may misunderstand that the > transaction failed even though commit record was already flushed. So you > think that each backend should not issue COMMIT PREPARED command to avoid > that inconsistency. To avoid that, it's better to make other process, the > resolver, issue the command and just make each backend wait for that to > completed. Right? > > Also using the resolver process has another merit; when there are unresolved > foreign transactions but the corresponding backend exits, the resolver can > try to resolve them. If something like this automatic resolution is > necessary, the process like the resolver would be necessary. Right? > > To the contrary, if we don't need such automatic resolution (i.e., unresolved > foreign transactions always need to be resolved manually) and we can prevent > the code to issue COMMIT PREPARED command from causing an error (not sure if > that's possible, though...), probably we don't need the resolver process. > Right?
Yes, I'm on the same page about all the above explanations. The resolver process has two functionalities: resolving foreign transactions automatically when the user issues COMMIT (the case you described in the second paragraph), and resolving foreign transaction when the corresponding backend no longer exist or when the server crashes during in the middle of 2PC (described in the third paragraph). Considering the design without the resolver process, I think we can easily replace the latter with the manual resolution. OTOH, it's not easy for the former. I have no idea about better design for now, although, as you described, if we could ensure that the process doesn't raise an error during resolving foreign transactions after committing the local transaction we would not need the resolver process. Or the second idea would be that the backend commits only the local transaction then returns the acknowledgment of COMMIT to the user without resolving foreign transactions. Then the user manually resolves the foreign transactions by, for example, using the SQL function pg_resolve_foreign_xact() within a separate transaction. That way, even if an error occurred during resolving foreign transactions (i.g., executing COMMIT PREPARED), it’s okay as the user is already aware of the local transaction having been committed and can retry to resolve the unresolved foreign transaction. So we won't need the resolver process while avoiding such inconsistency. But a drawback would be that the transaction commit doesn't ensure that all foreign transactions are completed. The subsequent transactions would need to check if the previous distributed transaction is completed to see its results. I’m not sure it’s a good design in terms of usability. Regards, -- Masahiko Sawada http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
