[Date Prev][Date Next]
Re: new admin guide draft
At 03:39 PM 9/24/2003, Quanah Gibson-Mount wrote:
>--On Thursday, September 18, 2003 12:20 AM -0400 Jonghyuk Choi <email@example.com> wrote:
>>>I've also seen the non-recovery -- On my systems, it also sent slapd
>>>into a mode where it was eating up the CPU, memory, and swap (I've
>>>filed an ITS on that). Do you see this over time? It took a little
>>>while for it to really ramp up.
>>More info on the setup is required in assessing the case.
>>Looks like this setup is larger than what I experimented with.
>>Was this observed only with the refreshAndPersist mode ?
>Okay, with refreshAndPersist, what I see is the master's slapd spike when the replica is first started (up to 20% of the CPU), and then drop back down to near 0 once it is done sending everything (about 5 minutes), so that by itself isn't bad. I do wonder what happens in a case like ours, where we shutdown all the servers nightly, then restart -- I imagine having 9 replica's start hammering the master within minutes of each other could have a negative effect.
I think you can resolve this by disabling updates on the
master until all the slaves are persisting. (If there are
no changes, then the refresh part of the refreshAndPersist
will have zero updates plus zero deletes.)
How to temporarily disable updates I leave as an exercise. :-)