For a relatively sane look at the tradeoff's we're talking about, this is a good resource:
http://developer.postgresql.org/pgdocs/postgres/runtime-config-wal.html I wish it was simple to write a heuristic which would detect single serialized client workloads, and delay commits, but I don't think it is. I lean (slightly) toward leaving delayed_commits = true because the worst case scenario, even in the case of a crash, isn't data corruption, just lost data from the most recent activity. It is also worth noting that there is an API to ensure_full_commit aside from the configuration value, so if you have high-value data you are writing, you can call ensure_full_commit (or use a header value to make the last PUT or POST operation force full commit) I think this is worth discussing. I'm not strongly in favor of the delayed_commit=true setting, but I do think it is slightly more user-friendly... Chris On Jul 5, 2010, at 10:02 AM, Mikeal Rogers wrote: > For the concurrent performance tests I wrote in relaximation it's actually > better to run with delayed_commits off because it measures the roundtrip > time of all the concurrent clients. > > The reason it's enabled by default is because of apache-bench and other > single writer performance test tools. From what I've seen, it doesn't > actually improve write performance under concurrent load and leads to a kind > of blocking behavior when you start throwing too many writes at it than it > can fsync in a second. The degradation in performance is pretty huge with > this "blocking" in my concurrent tests. > > I don't know of a lot of good concurrent performance test tools which is why > I went and wrote one. But, it only tests CouchDB and people love to pick up > one of these tools that tests a bunch of other dbs (poorly) and be like > "CouchDB is slow" because they are using a single writer. > > But, IMHO it's better to ship with more guarantees about consistency than > optimized for crappy perf tools. > > -Mikeal > > On Mon, Jul 5, 2010 at 8:49 AM, Volker Mische <[email protected]>wrote: > >> Hi All, >> >> delayed_commits were enabled to have better performance especially for >> single writers. The price you pay for is that you potentially lose up to one >> second of writes in case of a crash. >> >> Such a setting makes sense, though in my opinion it shouldn't be enabled by >> default. I expect* that people running into performance issues at least take >> a look at the README or a FAQ section somewhere. There the delayed_commit >> setting could be pointed out. >> >> I'd like to be able to say that on a vanilla CouchDB it's hard to lose >> data, but I can't atm. I'm also well aware that there will be plenty of >> performance tests when 1.0 is released and people will complain (if >> delayed_commits would be set to false by default) that it is horrible slow. >> Though safety of the data is more important for me. >> >> If the only reason why delayed_commits is true by default are the >> performance tests of some noobs, I really don't think it's a price worth >> paying. >> >> *I know that in reality people don't >> >> I would like to see delayed_commits=false for 1.0 >> >> Cheers, >> Volker >>
