hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From James Taylor <jtay...@salesforce.com>
Subject Re: Help in designing row key
Date Wed, 03 Jul 2013 10:33:48 GMT
Hi Flavio,
Have you had a look at Phoenix (https://github.com/forcedotcom/phoenix)? 
It will allow you to model your multi-part row key like this:

CREATE TABLE flavio.analytics (
     source INTEGER,
     type INTEGER,
     qual VARCHAR,
     hash VARCHAR,
     ts DATE
     CONSTRAINT pk PRIMARY KEY (source, type, qual, hash, ts) // Defines 
columns that make up the row key
)

Then you can issue SQL queries like this (to query for the last 7 days 
worth of data):
SELECT * FROM flavio.analytics WHERE source IN (1,2,5) AND type IN 
(55,66) AND ts > CURRENT_DATE() - 7

This will internally take advantage of our SkipScan 
(http://phoenix-hbase.blogspot.com/2013/05/demystifying-skip-scan-in-phoenix.html) 
to jump through your key space similar to FuzzyRowFilter, but in 
parallel from the client taking into account your region boundaries.

Or do more complex GROUP BY queries like this (to aggregate over the 
last 30 days worth of data, bucketized by day):
SELECT type,COUNT(*) FROM flavio.analytics WHERE ts > CURRENT_DATE() - 
30 GROUP BY type,TRUNCATE(ts,'DAY')

No need to worry about lexicographical sort order, flipping sign bits, 
normalizing/padding integer values, and all the other nuances of working 
with an API that works at the level of bytes. No need to write and 
manage installation of your own coprocessors to make aggregation 
efficient, perform topN queries, etc.

HTH.

Regards,
James
@JamesPlusPlus

On 07/03/2013 02:58 AM, Anoop John wrote:
> When you make the RK and convert the int parts into byte[] ( Use
> org.apache.hadoop.hbase.util.Bytes#toBytes(*int) *)  it will give 4 bytes
> for every byte..  Be careful about the ordering...   When u convert a +ve
> and -ve integer into byte[] and u do Lexiographical compare (as done in
> HBase) u will see -ve number being greater than +ve..  If you dont have to
> do deal with -ve numbers no issues  :)
>
> Well when all the parts of the RK is of fixed width u will need any
> seperator??
>
> -Anoop-
>
> On Wed, Jul 3, 2013 at 2:44 PM, Flavio Pompermaier <pompermaier@okkam.it>wrote:
>
>> Yeah, I was thinking to use a normalization step in order to allow the use
>> of FuzzyRowFilter but what is not clear to me is if integers must also be
>> normalized or not.
>> I will explain myself better. Suppose that i follow your advice and I
>> produce keys like:
>>   - 1|1|somehash|sometimestamp
>>   - 55|555|somehash|sometimestamp
>>
>> Whould they match the same pattern or do I have to normalize them to the
>> following?
>>   - 001|001|somehash|sometimestamp
>>   - 055|555|somehash|sometimestamp
>>
>> Moreover, I noticed that you used dots ('.') to separate things instead of
>> pipe ('|')..is there a reason for that (maybe performance or whatever) or
>> is just your favourite separator?
>>
>> Best,
>> Flavio
>>
>>
>> On Wed, Jul 3, 2013 at 10:12 AM, Mike Axiak <mike@axiak.net> wrote:
>>
>>> I'm not sure if you're eliding this fact or not, but you'd be much
>>> better off if you used a fixed-width format for your keys. So in your
>>> example, you'd have:
>>>
>>> PATTERN: source(4-byte-int).type(4-byte-int or smaller).fixed 128-bit
>>> hash.8-byte timestamp
>>>
>>> Example: \x00\x00\x00\x01\x00\x00\x02\x03....
>>>
>>> The advantage of this is not only that it's significantly less data
>>> (remember your key is stored on each KeyValue), but also you can now
>>> use FuzzyRowFilter and other techniques to quickly perform scans. The
>>> disadvantage is that you have to normalize the source-> integer but I
>>> find I can either store that in an enum or cache it for a long time so
>>> it's not a big issue.
>>>
>>> -Mike
>>>
>>> On Wed, Jul 3, 2013 at 4:05 AM, Flavio Pompermaier <pompermaier@okkam.it
>>>
>>> wrote:
>>>> Thank you very much for the great support!
>>>> This is how I thought to design my key:
>>>>
>>>> PATTERN: source|type|qualifier|hash(name)|timestamp
>>>> EXAMPLE:
>>>> google|appliance|oven|be9173589a7471a7179e928adc1a86f7|1372837702753
>>>>
>>>> Do you think my key could be good for my scope (my search will be
>>>> essentially by source or source|type)?
>>>> Another point is that initially I will not have so many sources, so I
>>> will
>>>> probably have only google|* but in the next phases there could be more
>>>> sources..
>>>>
>>>> Best,
>>>> Flavio
>>>>
>>>> On Tue, Jul 2, 2013 at 7:53 PM, Ted Yu <yuzhihong@gmail.com> wrote:
>>>>
>>>>> For #1, yes - the client receives less data after filtering.
>>>>>
>>>>> For #2, please take a look at TestMultiVersions
>>>>> (./src/test/java/org/apache/hadoop/hbase/TestMultiVersions.java in
>> 0.94)
>>>>> for time range:
>>>>>
>>>>>      scan = new Scan();
>>>>>
>>>>>      scan.setTimeRange(1000L, Long.MAX_VALUE);
>>>>> For row key selection, you need a filter. Take a look at
>>>>> FuzzyRowFilter.java
>>>>>
>>>>> Cheers
>>>>>
>>>>> On Tue, Jul 2, 2013 at 10:35 AM, Flavio Pompermaier <
>>> pompermaier@okkam.it
>>>>>> wrote:
>>>>>>   Thanks for the reply! I thus have two questions more:
>>>>>>
>>>>>> 1) is it true that filtering on timestamps doesn't affect
>>> performance..?
>>>>>> 2) could you send me a little snippet of how you would do such a
>>> filter
>>>>> (by
>>>>>> row key + timestamps)? For example get all rows whose key starts
>> with
>>>>>> 'someid-' and whose timestamps is greater than some timestamp?
>>>>>>
>>>>>> Best,
>>>>>> Flavio
>>>>>>
>>>>>>
>>>>>> On Tue, Jul 2, 2013 at 6:25 PM, Ted Yu <yuzhihong@gmail.com>
wrote:
>>>>>>
>>>>>>> bq. Using timestamp in row-keys is discouraged
>>>>>>>
>>>>>>> The above is true.
>>>>>>> Prefixing row key with timestamp would create hot region.
>>>>>>>
>>>>>>> bq. should I filter by a simpler row-key plus a filter on
>> timestamp?
>>>>>>> You can do the above.
>>>>>>>
>>>>>>> On Tue, Jul 2, 2013 at 9:13 AM, Flavio Pompermaier <
>>>>> pompermaier@okkam.it
>>>>>>>> wrote:
>>>>>>>> Hi to everybody,
>>>>>>>>
>>>>>>>> in my use case I have to perform batch analysis skipping
old
>> data.
>>>>>>>> For example, I want to process all rows created after a certain
>>>>>>> timestamp,
>>>>>>>> passed as parameter.
>>>>>>>>
>>>>>>>> What is the most effective way to do this?
>>>>>>>> Should I design my row-key to embed timestamp?
>>>>>>>> Or just filtering by timestamp of the row is fast as well?
Or
>> what
>>>>>> else?
>>>>>>>> Initially I was thinking to compose my key as:
>>>>>>>> timestamp|source|title|type
>>>>>>>>
>>>>>>>> but:
>>>>>>>>
>>>>>>>> 1) Using timestamp in row-keys is discouraged
>>>>>>>> 2) If this design is ok, using this approach I still have
>> problems
>>>>>>>> filtering by timestamp because I cannot found a way to
>> numerically
>>>>>> filer
>>>>>>>> (instead of alphanumerically/by string). Example:
>>>>>>>> 1372776400441|something has timestamp lesser
>>>>>>>> than 1372778470913|somethingelse but I cannot filter all
row
>> whose
>>>>> key
>>>>>> is
>>>>>>>> "numerically" greater than 1372776400441. Is it possible
to
>>> overcome
>>>>>> this
>>>>>>>> issue?
>>>>>>>> 3) If this design is not ok, should I filter by a simpler
>> row-key
>>>>> plus
>>>>>> a
>>>>>>>> filter on timestamp? Or what else?
>>>>>>>>
>>>>>>>> Best,
>>>>>>>> Flavio
>>>>>>>>


Mime
View raw message