accumulo-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Josh Elser <josh.el...@gmail.com>
Subject Re: Determining tablets assigned to table splits, and the number of rows in each tablet
Date Sat, 04 Oct 2014 14:33:44 GMT
For N splits in a table, you will always have N+1 tablets.

As you put below, the tablets are defined as (-Inf,c], (c,g], (g,w], 
(w,Inf) as long as I'm remembering the inclusivity correctly.

The number of unique rows within a tablet are not explicitly tracked.

I'd encourage you to submit some suggestions about new methods that we 
could add that would fill any gaps about distribution of tablets. 
Letting clients have some notion of where data is coming from may be 
useful (even if it is subject to change at any moment -- table balancer, 
servers dying, etc).

Dylan Hutchison wrote:
> This is for Accumulo 1.6.  Suppose we have the table splits
>
>     c
>
>     g
>
>     w
>
>
> Does anyone know how to determine
>
>  1. *the number of tablets assigned to each table split range? *
>     For this example, this is the number of tablets in the ranges
>     (-Inf,c), (c,g), (g,w), (w,Inf).  Or is the design 1-1, that is, for
>     each table split range there is exactly one tablet?
>  2. *the number of rows inside all the tablets occupying a table split
>     range? *
>     For this example, this is the total number of rows among all tablets
>     in the ranges (-Inf,c), (c,g), (g,w), (w,Inf).
>
> We use this count to verify how well manually set table splits are load
> balancing in the tables.
>
> Some context: I wrote functions that found these numbers two years ago
> working on D4M in Accumulo 1.5.  I took the dark route of using
> non-public Accumulo API to get TabletServer information, get TabletStats
> information, and find the matchings to a table's splits by scanning the
> extents listed in the METATABLE.  I can share the code if anyone is
> curious.  It's not pretty, but it did the job.
>
> Moving forward as we aim to upgrade to Accumulo 1.6, we should determine
> the tablet split information the right way, not by reverse engineering
> Accumulo.  Any suggestions?
>
> Thanks,
> Dylan Hutchison
>
> --
> www.cs.stevens.edu/~dhutchis <http://www.cs.stevens.edu/~dhutchis>

Mime
View raw message