At Wed, 5 Apr 2017 11:33:51 -0400, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote in
<5401fef6-c0c0-7e8a-d8b1-169e30cbd854@2ndquadrant.com>
> After further thinking, I prefer the alternative approach of using
> pq_sendcountedtext() as is and sticking the trailing zero byte on on the
> receiving side. This is a more localized change, and keeps the logical
> replication protocol consistent with the main FE/BE protocol. (Also, we
> don't need to send a useless byte around.)
I'm not sure about the significance of the trailing zero in the
the logical replication protocol. Anyway the patch works.
> Patch attached, and also a test case.
The problem was revealed when a string is shortened by encoding
conversion. The test covers the situation.
- The patches appliy on the master cleanly.
- The patch works for the UTF-8 => EUC_JP case.
- The test seems proper.
By the way, an untranslatable character on the publisher table
stops walsender with the following error.
> ERROR: character with byte sequence 0xe6 0xbc 0xa2 in encoding "UTF8" has no equivalent in encoding "LATIN1"
> STATEMENT: COPY public.t TO STDOUT
> LOG: could not send data to client: Broken pipe
> FATAL: connection to client lost
walreceiver stops on the opposite side with the following
complaint.
> ERROR: could not receive data from WAL stream: ERROR: character with byte sequence 0xe6 0xbc 0xa2 in encoding
"UTF8"has no equivalent in encoding "LATIN1"
> CONTEXT: COPY t, line 1: ""
> LOG: worker process: logical replication worker for subscription 16391 sync 16384 (PID 26915) exited with exit code
1
After this, walreceiver repeats reconnecting to master with no
wait. Maybe walreceiver had better refrain from reconnection
after certain kinds of faiure but it is not an urgent issue.
regards,
--
Kyotaro Horiguchi
NTT Open Source Software Center