Just a Theory

Black lives matter

Mocking Serialization Failures

I’ve been hacking on the forthcoming Bucardo 5 code base the last couple weeks, as we’re going to start using it pretty extensively at work, and it needed a little love to get it closer to release. The biggest issue I fixed was the handling of serialization failures.

When copying deltas from one database to another, Bucardo sets the transaction isolation to “Serializable”. As of PostgreSQL 9.1, this is true serializable isolation. However, there were no tests for it in Bucardo. And since pervious versions of PostgreSQL had poorer isolation (retained in 9.1 as “Repeatable Read”), I don’t think anyone really noticed it much. As I’m doing all my testing against 9.2, I was getting the serialization failures about half the time I ran the test suite. It took me a good week to chase down the issue. Once I did, I posted to the Bucardo mail list pointing out that Bucardo was not attempting to run a transaction again after failure, and at any rate, the model for how it thought to do so was a little wonky: it let the replicating process die, on the assumption that a new process would pick up where it left off. It did not.

Bucardo maintainer Greg Sabino Mullane proposed that we let the replicating process try again on its own. So I went and made it do that. And then the tests started passing every time. Yay!

Returning to the point of this post, I felt that there ought to be tests for serialization failures in the Bucardo test suite, so that we can ensure that this continues to work. My first thought was to use PL/pgSQL in 8.4 and higher to mock a serialization failure. Observe:

david=# \set VERBOSITY verbose
david=# DO $$BEGIN RAISE EXCEPTION 'Serialization error'
       USING ERRCODE = 'serialization_failure'; END $$;
ERROR:  40001: Serialization error
LOCATION:  exec_stmt_raise, pl_exec.c:2840

Cool, right? Well, the trick is to get this to run on the replication target, but only once. When Bucardo retries, we want it to succeed, thus properly demonstrating the COPY/SERIALIZATION FAIL/ROLLBACK/COPY/SUCCESS pattern. Furthermore, when it copies deltas to a target, Bucardo disables all triggers and rules. So how to get something trigger-like to run on a target table and throw the serialization error?

Studying the Bucardo source code, I discovered that Bucardo itself does not disable triggers and rules. Rather, it sets the session_replica_role GUC to “replica”. This causes PostgreSQL to disable the triggers and rules — except for those that have been set to ENABLE REPLICA. The PostgreSQL ALTER TABLE docs:

The trigger firing mechanism is also affected by the configuration variable session_replication_role. Simply enabled triggers will fire when the replication role is “origin” (the default) or “local”. Triggers configured as ENABLE REPLICA will only fire if the session is in “replica” mode, and triggers configured as ENABLE ALWAYS will fire regardless of the current replication mode.

Well how cool is that? So all I needed to do was plug in a replica trigger and have it throw an exception once but not twice. Via email, Kevin Grittner pointed out that a sequence might work, and indeed it does. Because sequence values are non-transactional, sequences return different values every time they’re access.

Here’s what I came up with:


) RETURNS trigger LANGUAGE plpgsql AS $_$
    IF nextval('serial_seq') % 2 = 0 THEN RETURN NEW; END IF;
    RAISE EXCEPTION 'Serialization error'
            USING ERRCODE = 'serialization_failure';

CREATE TRIGGER mock_serial_fail AFTER INSERT ON bucardo_test2
    FOR EACH ROW EXECUTE PROCEDURE mock_serial_fail();
ALTER TABLE bucardo_test2 ENABLE REPLICA TRIGGER mock_serial_fail;

The first INSERT (or, in Bucardo’s case, COPY) to bucardo_test2 will die with the serialization error. The second INSERT (or COPY) succeeds. This worked great, and I was able to write test in a few hours and get them committed. And now we can be reasonably sure that Bucardo will always properly handle serialization failures.

Looking for the comments? Try the old layout.