Обсуждение: Re: [PERFORM] how to plan for vacuum?
Jim C. Nasby wrote: > I'll generally start with a cost delay of 20ms and adjust based on IO > utilization. I've been considering set a default autovacuum cost delay to 10ms; does this sound reasonable? -- Alvaro Herrera http://www.CommandPrompt.com/ PostgreSQL Replication, Consulting, Custom Development, 24x7 support
On Thu, Jan 25, 2007 at 12:52:02AM -0300, Alvaro Herrera wrote: > Jim C. Nasby wrote: > > > I'll generally start with a cost delay of 20ms and adjust based on IO > > utilization. > > I've been considering set a default autovacuum cost delay to 10ms; does > this sound reasonable? For a lightly loaded system, sure. For a heavier load that might be too much, but of course that's very dependent on not only your hardware, but how much you want vacuum to interfere with normal operations. Though, I'd say as a default it's probably better to be more aggressive rather than less. Also, it might be better to only set autovac_cost_delay by default; presumably if someone's running vacuum by hand they want it done pronto. -- Jim Nasby jim@nasby.net EnterpriseDB http://enterprisedb.com 512.569.9461 (cell)
Jim C. Nasby wrote: > On Thu, Jan 25, 2007 at 12:52:02AM -0300, Alvaro Herrera wrote: >> Jim C. Nasby wrote: >> >>> I'll generally start with a cost delay of 20ms and adjust based on IO >>> utilization. >> I've been considering set a default autovacuum cost delay to 10ms; does >> this sound reasonable? The problem in here is that we can not easily find a direct relation between Cost delay <-> CPU/IO utilization <--> real performance (response time in peak hour) It is very hard for any normal user to set this correctly. I think the experience / trial-and-error approach is awful for the user, every DBA need to be an expert of vacuum to keep the system stable. For vacuum is still a big threat to the performance, a more intelligent way is needed. A lot of efforts have contributed to make vacuum to be a more lightweight operation, but I think we should still need more efforts on how to make it can be used easily and safely. So I have proposed the "vacuum in time" feature in previous; just let vacuum know how long can it runs, and then it will minimize the impact in the time span for you. Some argue that it should not have the maintenance window assumption, but the most safely way is to run in the maintenance window.
Alvaro Herrera wrote: > Jim C. Nasby wrote: > >> I'll generally start with a cost delay of 20ms and adjust based on IO >> utilization. > > I've been considering set a default autovacuum cost delay to 10ms; does > this sound reasonable? It really depends on the system. Most of our systems run anywhere from 10-25ms. I find that any more than that, Vacuum takes too long. Joshua D. Drake -- === The PostgreSQL Company: Command Prompt, Inc. === Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240 Providing the most comprehensive PostgreSQL solutions since 1997 http://www.commandprompt.com/ Donate to the PostgreSQL Project: http://www.postgresql.org/about/donate PostgreSQL Replication: http://www.commandprompt.com/products/
On Thu, Jan 25, 2007 at 07:52:50PM +0900, Galy Lee wrote: > It is very hard for any normal user to set this correctly. I think the > experience / trial-and-error approach is awful for the user, every DBA > need to be an expert of vacuum to keep the system stable. For vacuum is > still a big threat to the performance, a more intelligent way is needed. Agreed. > So I have proposed the "vacuum in time" feature in previous; just let > vacuum know how long can it runs, and then it will minimize the impact > in the time span for you. Some argue that it should not have the > maintenance window assumption, but the most safely way is to run in the > maintenance window. Most systems I work on don't have a maintenance window. For those that do, the window is at best once a day, and that's nowhere near often enough to be vacuuming any database I've run across. I'm not saying they don't exist, but effort put into restricting vacuums to a maintenance window would serve very few people. It'd be much better to put effort into things like piggyback vacuum. -- Jim Nasby jim@nasby.net EnterpriseDB http://enterprisedb.com 512.569.9461 (cell)
On Thu, Jan 25, 2007 at 08:04:49AM -0800, Joshua D. Drake wrote: > > It really depends on the system. Most of our systems run anywhere from > 10-25ms. I find that any more than that, Vacuum takes too long. How do you measure the impact of setting it to 12 as opposed to 15?
On Jan 25, 2007, at 10:33 AM, Ray Stell wrote: > On Thu, Jan 25, 2007 at 08:04:49AM -0800, Joshua D. Drake wrote: >> >> It really depends on the system. Most of our systems run anywhere >> from >> 10-25ms. I find that any more than that, Vacuum takes too long. > > > How do you measure the impact of setting it to 12 as opposed to 15? If you've got a tool that will report disk utilization as a percentage it's very easy; I'll decrease the setting until I'm at about 90% utilization with the system's normal workload (leaving some room for spikes, etc). Sometimes I'll also tune the costs if reads vs. writes are a concern. -- Jim Nasby jim@nasby.net EnterpriseDB http://enterprisedb.com 512.569.9461 (cell)