The following bug has been logged on the website:
Bug reference: 13844
Logged by: Jarred Ward
Email address: jarred@webriots.com
PostgreSQL version: 9.4.5
Operating system: Linux
Description:
This is a pretty serious logical decoding bug that bit us pretty hard
in production.
Run the following code:
----------------------------------------------------------------------------
CREATE TABLE foo(a SERIAL PRIMARY KEY, b TEXT);
SELECT pg_create_logical_replication_slot('regression_slot',
'test_decoding');
INSERT INTO foo (b) VALUES ('john');
BEGIN;
SELECT * FROM foo WHERE b = 'john' FOR UPDATE;
SAVEPOINT sp1;
INSERT INTO foo (b) VALUES ('jane');
ROLLBACK TO SAVEPOINT sp1;
INSERT INTO foo (b) VALUES ('doe');
END;
SELECT data FROM pg_logical_slot_get_changes('regression_slot', NULL,
NULL);
----------------------------------------------------------------------------
The last select above should return the decoded WAL, but instead it
returns:
----------------------------------------------------------------------------
ERROR: subxact logged without previous toplevel record
----------------------------------------------------------------------------
And the process is disconnected with no way to recover the replication
slot without dropping and recreating.
If SELECT statement in the transaction is changed the following:
----------------------------------------------------------------------------
SELECT * FROM foo WHERE b = 'john' FOR UPDATE;
----------------------------------------------------------------------------
to remove the row level locking the correct decoded change set is
returned.