[Pgpool-general] replication_mode safety

[Tatsuo Ishii]

Hi Jeff,

It was nice to see you in Ottawa.

> I have a few questions about the safety of replication_mode in pgpool-
> II. There are two cases that I'm thinking about:
> 
> (1)
> 
> The docs say about replication_strict:
> "When set to true, pgpool-II waits for the Master DB for all queries to
> be processed before moving on to the next backend. This option
> eliminates the risk of deadlocks, but the performance may degrade due to
> the lack of a parallelism."
> 
> Does this mean that all writes are serialized through pgpool, preventing
> any concurrent writes at all, even from different clients?

No.

> Or does it just mean that, for a given client, the queries are executed
> on the underlying replicants serially?

Yes.

> (2)
> 
> If the pgpool server fails suddenly, is it possible that some
> transactions may have committed on some replicants and not others?

Yes, it could happen.

> If
> so, is there a way to re-synchronize them?

Currently no. There is an ongoing development to re-sync nodes. It
will act very much similar to PGCluster does.

> It's possible 2PC would be useful here (?), but it seems like it would
> be difficult to change client queries into 2PC transactions.

Yes, I think 2PC is usefull for this case.
--
Tatsuo Ishii
SRA OSS, Inc. Japan