Per-tuple memory leak in 9.0
While testing triggers, I came across the following memory leak.
Here's a simple test case:
CREATE TABLE foo(a int);
CREATE OR REPLACE FUNCTION trig_fn() RETURNS trigger AS
$$
BEGIN
RETURN NEW;
END;
$$
LANGUAGE plpgsql;
CREATE TRIGGER ins_trig BEFORE INSERT ON foo
FOR EACH ROW EXECUTE PROCEDURE trig_fn();
INSERT INTO foo SELECT g
FROM generate_series(1, 5000000) AS g;
Memory usage goes up by around 100 bytes per row for the duration of the query.
The problem is that the trigger code assumes that anything it
allocates in the per-tuple memory context will be freed per-tuple
processed, which used to be the case because the loop in ExecutePlan()
calls ResetPerTupleExprContext() once each time round the loop, and
that used to correspond to once per tuple.
However, with the refactoring of that code out to nodeModifyTable.c,
this is no longer the case because the ModifyTable node processes all
the tuples from the subquery before returning, so I guess that the
loop in ExecModifyTable() needs to call ResetPerTupleExprContext()
each time round.
Regards,
Dean
Sorry, I meant to cc this to -bugs as well as -hackers
---------- Forwarded message ----------
From: Dean Rasheed <dean.a.rasheed@gmail.com>
Date: 18 August 2010 18:29
Subject: Per-tuple memory leak in 9.0
To: pgsql-hackers@postgresql.org
While testing triggers, I came across the following memory leak.
Here's a simple test case:
CREATE TABLE foo(a int);
CREATE OR REPLACE FUNCTION trig_fn() RETURNS trigger AS
$$
BEGIN
RETURN NEW;
END;
$$
LANGUAGE plpgsql;
CREATE TRIGGER ins_trig BEFORE INSERT ON foo
FOR EACH ROW EXECUTE PROCEDURE trig_fn();
INSERT INTO foo SELECT g
FROM generate_series(1, 5000000) AS g;
Memory usage goes up by around 100 bytes per row for the duration of the query.
The problem is that the trigger code assumes that anything it
allocates in the per-tuple memory context will be freed per-tuple
processed, which used to be the case because the loop in ExecutePlan()
calls ResetPerTupleExprContext() once each time round the loop, and
that used to correspond to once per tuple.
However, with the refactoring of that code out to nodeModifyTable.c,
this is no longer the case because the ModifyTable node processes all
the tuples from the subquery before returning, so I guess that the
loop in ExecModifyTable() needs to call ResetPerTupleExprContext()
each time round.
Regards,
Dean
Dean Rasheed <dean.a.rasheed@gmail.com> writes:
The problem is that the trigger code assumes that anything it
allocates in the per-tuple memory context will be freed per-tuple
processed, which used to be the case because the loop in ExecutePlan()
calls ResetPerTupleExprContext() once each time round the loop, and
that used to correspond to once per tuple.
However, with the refactoring of that code out to nodeModifyTable.c,
this is no longer the case because the ModifyTable node processes all
the tuples from the subquery before returning, so I guess that the
loop in ExecModifyTable() needs to call ResetPerTupleExprContext()
each time round.
Hmmm ... it seems a bit unclean to be resetting the output-tuple
exprcontext at a level below the top of the plan. I agree that that's
probably the sanest fix at the moment, but I fear we may need to revisit
this in connection with writable CTEs. We might need a separate output
tuple context for each ModifyTable node, or something like that.
regards, tom lane