Skip site navigation (1) Skip section navigation (2)

NOLOGGING option, or ?

From: Simon Riggs <simon(at)2ndquadrant(dot)com>
To: pgsql-hackers(at)postgresql(dot)org
Subject: NOLOGGING option, or ?
Date: 2005-05-31 23:01:43
Message-ID: 1117580503.3844.785.camel@localhost.localdomain (view raw or flat)
Thread:
Lists: pgsql-hackers
Recent test results have shown a substantial performance improvement
(+25%) if WAL logging is disabled for large COPY statements. This is to
be expected, though has a price attached: losing the ability to crash
recover data loaded in this manner.

There are two parts to this proposal. First, when and whether to do this
at all. Second, syntax and invocation.

Why?

Performance.

The performance gain has a price and so should only be enabled if
requested explicitly by the user. It is up to the user whether they
accept this price, since in many useful cases it is a small price
against a huge saving.

The price is that if a crash occurs, then any table that was not empty
to begin with would not be in a transactionally consistent state
following crash recovery. It may have data in it, but it would be up to
the user to determine whether that was satisfactory or not. It could be
possible to sense what to do in this situation automatically, by putting
the table into a needs-recovery type state... I don't propose to handle
this *at this stage*.

Syntax and invocation:

Previously I had discussed adding a NOLOGGING option onto both COPY and
CREATE TABLE AS SELECT that would bypass the creation of wal logging
data. That is still a possibility, but would require manual code changes
to much of the SQL submitted.

Now, I would like to discuss adding an enable_logging USERSET GUC, that
would apply *only* to COPY and CREATE TABLE AS SELECT. The default of
this would be false.

How can we gain this performance benefit for those willing to accept the
restrictions imposed?

Your comments are sought and are most welcome.

Best Regards, Simon Riggs


Responses

pgsql-hackers by date

Next:From: Jonah H. HarrisDate: 2005-05-31 23:06:40
Subject: Re: Tablespace-level Block Size Definitions
Previous:From: Marc G. FournierDate: 2005-05-31 22:57:22
Subject: Re: Deadlock with tsearch2 index ...

Privacy Policy | About PostgreSQL
Copyright © 1996-2014 The PostgreSQL Global Development Group