kudu-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alexandre Fouché <afou...@onfocus.io>
Subject Re: stripes, JBOD: Assignment and rebalance ?
Date Thu, 09 Mar 2017 14:31:06 GMT
Great, thank you Dan !

2017-03-02 18:51 GMT+01:00 Dan Burkert <danburkert@apache.org>:

> Individual tablets aren't assigned to a specific disk, they are spread
> across all of the data disks.  So individual disks should fill up evenly.
>
> - Dan
>
> On Thu, Mar 2, 2017 at 9:38 AM, Alexandre Fouché <afouche@onfocus.io>
> wrote:
>
>> Hi Dan
>> So will Kudu rebalance tablets on different disks dynamically when some
>> gets bigger than other, in order not to fill a disk while the other disk
>> space could remain mostly free ?
>>
>> 2017-03-02 18:26 GMT+01:00 Dan Burkert <danburkert@apache.org>:
>>
>>> Hi Alexandre,
>>>
>>> responses inline
>>>
>>> On Thu, Mar 2, 2017 at 9:18 AM, Alexandre Fouché <afouche@onfocus.io>
>>> wrote:
>>>>
>>>>
>>>> When storing data on multiple JBOD disks, will Kudu assign data for
>>>> tablets efficiently as far as tablet sizes or activity are concerned, or
>>>> will it simply try to assign roughly the same number of tablets on each
>>>> disk, regarless of tablets true size or activity (we have many empty
>>>> tablets at this time). And will it rebalance tablets to one disk or another
>>>> automatically ? Or is it still better to expose one RAID0 volume ?
>>>>
>>>
>>> Kudu will evenly spread data from all tablets across all disks.  This
>>> allows Kudu to get good write throughput and balancing, but similarly to
>>> RAID 0 it means that if one drive fails, all tablets on that tablet server
>>> will become unavailable. Kudu will automatically recover by re-replicating
>>> the tablets to a different tablet server as long as a majority of replicas
>>> are still available.  So, RAID 0 should provide no benefit for Kudu.  It's
>>> on the roadmap to make multi-disk configuration more flexible so that if a
>>> single disk dies only a subset of the tablets will become unavailable, but
>>> I don't have a timeline on that feature (no one is working on it to my
>>> knowledge).
>>>
>>>
>>>> Will using JBOD disks better than RAID stripes ? It seems from Bug
>>>> reports that when WAL disk fails, or one of the JBOD data disks, Kudu is
>>>> still unable to recover and keep or migrate good tablets. In that case, it
>>>> shows no improvement over a failed disk on a RAID0 where in both cases the
>>>> only recover option is to delete the whole Kudu data and WAL and let it
>>>> resync from other nodes ?
>>>>
>>>
>>> I think what I wrote previously answers this, if not I can clarify.
>>>
>>> I found comments that WAL can only be one one disk, is it still the
>>>> case, or is this info obsolete ?
>>>>
>>>
>>> This is currently the case. If you have many disks it's often
>>> advantageous to put the WAL on it's own disk (ideally an SSD if it's
>>> available). The WAL workload is more latency sensitive than the data
>>> workload.
>>>
>>> - Dan
>>>
>>
>>
>

Mime
View raw message