I have been evaluating the logical replication feature in the JDBC driver v42.0.0, against postgres v9.6. One scenario I'm testing is streaming a high volume of transactions, where each transaction contains multiple INSERTs.
Sometimes, two transaction COMMIT messages follow each other, one after the other, in the DB transaction log, and so the transactions arrive one after the other in the logical replication stream. This is expected behaviour.
I apply and flush the LastReceivedLSN and force an update to the server after every COMMIT message.
To stress-test the recovery behaviour, I'm also disconnecting and reconnecting on every fourth INSERT received.
In most cases I see the following expected behaviour:
The WAL might have (with made-up LSNs for illustrative purposes):
The above behaviour makes sense since the replication slot's confirmed_flush_lsn=99 upon reconnect.
My issue: occasionally after reconnecting, I observe that the INSERTs for TXN 2 are not resent and instead the stream moves on to TXN 3.
With the same WAL as above, the stream looks like:
BEGIN (TXN 1) LastReceivedLSN=19 INSERT (TXN 1) LastReceivedLSN=39 INSERT (TXN 1) LastReceivedLSN=59 INSERT (TXN 1) LastReceivedLSN=79 COMMIT (TXN 1) LastReceivedLSN=109 <-- This LSN is wrong! <JDBC APPLY & FLUSH 109, FORCE UPDATE> BEGIN (TXN 2) LastReceivedLSN=29 INSERT (TXN 2) LastReceivedLSN=49 <JDBC RECONNECT> BEGIN (TXN 3) ...
Debugging the JDBC driver and comparing it with the WAL (via pg_xlogdump), it seems that on these occasions COMMIT (TXN 1) arrives with the correct payload but with StartLSN set to the StartLSN of COMMIT (TXN 2)! The JDBC driver computes COMMIT (TXN 1) LastReceivedLSN = COMMIT (TXN 2) StartLSN + length of COMMIT (TXN 1) payload. This causes us to unwittingly set confirmed_flush_lsn to the end WAL position of COMMIT (TXN 2) before disconnecting, meaning I don't decode TXN 2 after reconnection.
Is this a known issue? Is it caused by the JDBC driver or something in the server?