On Tue, Sep 15, 2009 at 3:00 AM, Noah Slater <nsla...@apache.org> wrote: > Hey, > > I plan on cutting the 0.10 release for a vote today. > > I would like to collect any objections now before I go through the motions. >
This morning I reopened http://issues.apache.org/jira/browse/COUCHDB-449 by committing a default configuration that maintains the setting that 0.9 ships with. The config option (delayed_commits) has been sitting in trunk and 0.10.x for a while, set to "slow and as safe as possible" instead of "fast and still pretty safe". I think fast-by-default is the winning strategy, especially as the safety we gain is only marginal and depends on the underlying operating system, and comes at roughly a 50x speed cost for serial doc inserts (the first thing new users try when benchmarking). The issue of slow vs fast is discussed in more depth on the ticket URL above. Sorry about the backwards order I operated in (commit first ask questions later) basically I'd been setting aside worries about the new slow default as abstract, until this thread surfaced, which made me hit the panic button. There is also a documentation hole here - we need to make it clear to users requiring absolute transactional safety, that they should switch the default (and also not run on operating systems that don't have a real fsync()). Chris > Thanks, > > -- > Noah Slater, http://tumbolia.org/nslater > -- Chris Anderson http://jchrisa.net http://couch.io