[
https://issues.apache.org/jira/browse/AMQ-7082?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gary Tully resolved AMQ-7082.
-----------------------------
Resolution: Fixed
The recovery is now back in the start phase, if it completes we are good.
Otherwise we try again the next time.
[~jgoodyear] - this is a take on the parallel approach, I think it makes good
sense.
[~alanprot] - there is still a good case for checkpointing which will reduce
the full replay window, but the perf impact will the be key determinant on that
I think.
It would be good to gauge the impact of the second reader in your case over
NFS, there may be need to slow down the recovery thread such that it does not
hog the disk or cpu.
> KahaDB index, recover free pages in parallel with start
> -------------------------------------------------------
>
> Key: AMQ-7082
> URL: https://issues.apache.org/jira/browse/AMQ-7082
> Project: ActiveMQ
> Issue Type: Bug
> Components: KahaDB
> Affects Versions: 5.15.0
> Reporter: Gary Tully
> Assignee: Gary Tully
> Priority: Major
> Fix For: 5.16.0
>
>
> AMQ-6590 fixes free page loss through recovery. The recover process can be
> timely, which prevents fast failover, doing recovery on shutdown is
> preferable, but it is still not ideal b/c it will hold onto the kahadb lock.
> It also can stall shutdown unexpectedly.
> AMQ-7080 is going to tackle checkpointing the free list. This should help
> avoid the need for recovery but it may still be necessary. If the perf hit is
> significant this may need to be optional.
> There will still be the need to walk the index to find the free list.
> It is possible to run with no free list and grow, and we can do that while we
> recover the free list in parallel, then merge the two at a safe point. This
> we can do at startup.
> In cases where the disk is the bottleneck this won't help much, but it will
> help failover and it will help shutdown, with a bit of luck the recovery will
> complete before we stop.
>
> Initially I thought this would be too complex, but if we concede some growth
> while we recover, ie: start with an empty free list, it is should be straight
> forward to merge with a recovered one.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
