phoenix-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Simon Wang <simon.w...@airbnb.com>
Subject Re: Index tables at scale
Date Tue, 12 Jul 2016 00:11:32 GMT
Thanks Mujtaba. This is good to know. It is possible manipulate the key bit to avoid the hot-spotting,
so we are probably trying unsalted table out. 

Still, it would be nice if combined indexes in a single table is possible. 


> On Jul 11, 2016, at 2:41 PM, Mujtaba Chohan <mujtaba@apache.org> wrote:
> 
> FYI if you keys are not written in order i.e. you are not concerned about write hot-spotting/write
throughput then try writing your data to an un-salted table. Read performance for un-salted
table can be comparable or better to salted one with stats <https://phoenix.apache.org/update_statistics.html>.
> 
> On Mon, Jul 11, 2016 at 2:31 PM, Simon Wang <simon.wang@airbnb.com <mailto:simon.wang@airbnb.com>>
wrote:
> This indexes will be salted indeed. (so is the data table). If all indexes reside in
the same table, there will be only 512 regions in total (256 for data table, 256 for the combined
index table). Indeed the combined index table will be 12x large as a single index table. But
it doesn’t cover all columns so it should be fine.
> 
>> On Jul 11, 2016, at 2:26 PM, James Taylor <jamestaylor@apache.org <mailto:jamestaylor@apache.org>>
wrote:
>> 
>> Will the index be salted (and that's why it's 256 regions per table)? If not, how
many regions would there be if all indexes are in the same table (assuming the table is 12x
bigger than one index table)?
>> 
>> On Monday, July 11, 2016, Simon Wang <simon.wang@airbnb.com <mailto:simon.wang@airbnb.com>>
wrote:
>> Thanks, Mujtaba. What you wrote is exactly what I meant. While not all our tables
needs these many regions and indexes, the num of regions/region server can grow quickly.
>> 
>> -Simon
>> 
>>> On Jul 11, 2016, at 2:17 PM, Mujtaba Chohan <mujtaba@apache.org <>>
wrote:
>>> 
>>> 12 index tables * 256 region per table = ~3K regions for index tables assuming
we are talking of covered index which implies 200+ regions/region server on a 15 node cluster.
>>> 
>>> On Mon, Jul 11, 2016 at 1:58 PM, James Taylor <jamestaylor@apache.org <>>
wrote:
>>> Hi Simon,
>>> 
>>> I might be missing something, but with 12 separate index tables or 1 index table,
the amount of data will be the same. Won't there be the same number of regions either way?
>>> 
>>> Thanks,
>>> James
>>> 
>>> On Sun, Jul 10, 2016 at 10:50 PM, Simon Wang <simon.wang@airbnb.com <>>
wrote:
>>> Hi James,
>>> 
>>> Thanks for the response.
>>> 
>>> In our use case, there is a 256 region table, and we want to build ~12 indexes
on it. We have 15 region servers. If each index is in its own table, that would be a total
of 221 regions per region server of this single table. I think the extra write time cost is
okay. But the number of regions is too high for us.
>>> 
>>> Best,
>>> Simon
>>> 
>>> 
>>>> On Jul 9, 2016, at 1:18 AM, James Taylor <jamestaylor@apache.org <>>
wrote:
>>>> 
>>>> Hi Simon,
>>>> The reason we've taken this approach with views is that it's possible with
multi-tenancy that the number of views would grow unbounded since you might end up with a
view per tenant (100K or 1M views or more - clearly too many for HBase to handle as separate
tables).
>>>> 
>>>> With secondary indexes directly on physical tables, you're somewhat bounded
by the hit you're willing to take on the write side, as the cost of maintaining the index
is similar to the cost of the write to the data table. So the extra number of physical tables
for indexes seems within the bounds of what HBase could handle. 
>>>> 
>>>> How many secondary indexes are you creating and are you ok with the extra
write-time cost?
>>>> 
>>>> From a code consistency standpoint, using the same approach across local,
global, and view indexes might simplify things, though. Please file a JIRA with a bit more
detail on your use case.
>>>> 
>>>> Thanks,
>>>> James
>>>> 
>>>> 
>>>> 
>>>> On Fri, Jul 8, 2016 at 8:59 PM, Simon Wang <simon.wang@airbnb.com <>>
wrote:
>>>> Hi all,
>>>> 
>>>> I am writing to ask if there is a way to let Phoenix store all indexes on
a single table in the same HBase table. If each index must be stored in a separate table,
creating more than a few indexes on table with a large number of regions will not scale well.
>>>> 
>>>> From what I have learned, when Phoenix builds indexes on a view, it stores
all indexes in a table associated with the underlying table of the view. e.g. if V1 is a view
of T1, all indexes on V1 will be stored in _IDX_T1. It would be great if this behavior can
be optionally turned on for indexes on tables.
>>>> 
>>>> Best,
>>>> Simon
>>>> 
>>> 
>>> 
>>> 
>> 
> 
> 


Mime
View raw message