On 06/05/17 19:15, Tom Lane wrote:
> Petr Jelinek <petr.jelinek@2ndquadrant.com> writes:
>> On 06/05/17 18:16, Tom Lane wrote:
>>> Hmm, I'm hoping for something more user-oriented. Is the corruption
>>> time-limited? What's an "exported snapshot" anyway, is it the same
>>> thing as pg_export_snapshot(), and if so what's that got to do with
>>> logical replication?
>
>> Okay to explain what's happening. When you create logical replication
>> slot via walsender, it exports snapshot (like the one exported by
>> pg_export_snapshot() indeed) which corresponds to exact point in time
>> where the decoding will start for the slot. You can import this snapshot
>> to another backend and use it to copy any existing data before starting
>> the replication. The bugs cause that these snapshots would be corrupted
>> and you'd have inconsistent data (some tuples missing). One of them
>> would export snapshot that's only valid for system catalogs but not user
>> tables (the ondisk snapshot, this needs at least one preexisting slot).
>> The other would not guarantee that tuples needed by the snapshot weren't
>> removed by vacuum of HOT pruning (the window being only the time it took
>> to generate the snapshot).
>
> OK, that's better. But I'm still not really seeing a reason to treat
> these as two separate items for release-note purposes: I don't think users
Sure my main issue was that the text combined the info from two commits
in a way that made the statement no longer correct.
> will care. Now that I've read section 31.4, I'd suggest that we phrase
> the release notes in the terms it uses. How about saying something like
> "The initial snapshot created for a logical replication slot was
> incorrect. This could allow the apply process to copy incomplete or
> inconsistent data. This was more likely to happen if the source server
> was busy at the time of slot creation, or if two slots were created
> concurrently" ?
>
> (Or, wait a minute. That documentation only applies to v10, but we
> need to be writing this relnote for 9.6 users. What terminology should
> we be using anyway?)
Yeah we need to somehow mention that it only affects 3rd party tools
using logical decoding.
"The initial snapshot created for a logical decoding slot was
potentially incorrect. This could allow the 3rd party tools using
the logical decoding to copy incomplete existing(?) data. This was
more likely to happen if the source server was busy at the time of
slot creation, or if two slots were created concurrently."
>
> Also, do we need to recommend that people not trust any logical replicas
> at this point, but recreate them after installing the update?
>
Yes, but only if there was preexisting data *and* there was concurrent
activity on the server when the "replication" was setup.
-- Petr Jelinek http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training &
Services