hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jonathan Gray <jl...@streamy.com>
Subject Re: TTL, Versions and storing long history
Date Tue, 30 Jun 2009 02:46:34 GMT

Prior to 0.20, I would definitely recommend moving the time component to 
the keys, columns, and values.  Even after 0.20, I recommend doing that 
if you want complete control.  My personal philosophy is that versions 
are for versioning, and if you are really using them as a time dimension 
of individual data points, you should consider not using versions.

However, the API and server-side implementation for versions is greatly 
improved.  You can specify stamps manually and you can query for any 
range you want, gets and scans.

There is not currently a way to keep versions < x weeks old but always 
keep the latest version.  If you wanted to enforce something like that, 
you could always write a MapReduce job that ran periodically and 
enforced what you wanted.

If you want to keep history forever, the idea is to use the "big enough" 
values.  In practice, only since HBase 0.20 have we been able to handle 
millions of versions of a single column (Integer.MAX_VALUE is >2 
billion, far beyond the capabilities of HBase).  The same goes for 
TTL... 2 billion seconds is over 60 years.  Could also move everything 
to Long which would ensure there would never be an issue.  Will dig more 
and let you know.

In any case, you'll need 0.20 to fully take advantage of versions.

Hope that helps.


Jon Schutz wrote:
> How do TTL and Versions specifications interact?  I'm guessing that the
> first limit reached applies, i.e. if TTL is 1 week and versions is 3,
> adding a fourth update to a data record would cause the first to be
> bumped even if it is less than a week old?  And if I only have 2
> versions but one is 2 weeks old, the expired one gets bumped even though
> the versions limit has not been reached?
> Is there a way to say "Keep versions < x weeks old, but always keep at
> least the latest version, no matter how old?"
> Suppose I want to keep the history about a particular object forever.
> Looks like TTL can be set to 'Forever' (-1) but Versions has no
> 'infinite' setting - I guess that's OK as in practice MAXINT is "big
> enough".  Would it be wise to use Hbase like this to maintain a history,
> or should I be adding a time component into the key and storing multiple
> records?  Can anyone help outline the pros and cons?
> Thanks,

View raw message