Hi,

Actually no, I have not used it in production, I wrote it for fun :)
But I know that it's used in TempoDB. You might want to contact them and
ask -
http://blog.tempo-db.com/post/42318820124/estimating-percentiles-on-streams-of-data

On Thu Nov 14 2013 at 6:17:15 AM, Matt Abrams <[email protected]> wrote:

> Great!  I have not used Q digest in production yet but I believe
> Eugene, the author of stream-lib's Q digest implementation, has.
> Eugene, can you comment on how it performs in practice?
>
> Matt
>
>
> On Wed, Nov 13, 2013 at 4:45 PM, Ted Dunning <[email protected]>
> wrote:
> >
> >
> > As soon as it is for sure done.  I have one more significant improvement
> to make so that it works on sequential values.  I will hand the code to
> suneel who will be packaging it for mahout. You can def have it at the same
> time.
> >
> > I would love a review from you guys when I am ready. The theory doc is
> nearly to that point.  Would you like I start there?  Also, can I get some
> info from you about how q digests work in practice?
> >
> > Sent from my iPhone
> >
> > On Nov 13, 2013, at 20:46, Matt Abrams <[email protected]> wrote:
> >
> >> Ted -
> >>
> >> Any chance we can add your quantile estimator to stream-lib?
> >>
> >> Matt
> >>
> >> On Wed, Nov 13, 2013 at 5:38 AM, Ted Dunning <[email protected]>
> wrote:
> >>> I also have a new quantile estimator that dominates all other
> >>> implementations that I know of on speed and accuracy (10us per point
> added,
> >>> 8K data size to get a few ppm accuracy for high or low quantiles and
> about
> >>> 0.05% accuracy on middle quantiles like the median).
> >>>
> >>>
> >>>
> >>>
> >>> On Wed, Nov 13, 2013 at 8:53 AM, Dmitriy Ryaboy <[email protected]>
> wrote:
> >>>
> >>>> Summingbird uses algebird. I think Stripe might also have a library,
> Avi
> >>>> Bryant was toying with this for a while.
> >>>>
> >>>> Algebird has some nice features like not doing approximation at all
> for
> >>>> small sets (just use the real values), etc. we also recently did a
> bunch of
> >>>> work to make sure we can serialize all approximate structures so they
> can
> >>>> be correctly reused by different computations, sent across the wire,
> etc.
> >>>>
> >>>> I don't recall doing speed comparisons and the like, it would be
> >>>> interesting to see them if you guys are choosing what library to use.
> >>>>
> >>>> On Nov 13, 2013, at 12:33 AM, Ted Dunning <[email protected]>
> wrote:
> >>>>
> >>>>> stream-lib is used quite widely and is generally high quality.
> >>>>>
> >>>>> The other competitive library is Brick House from Klout.
> >>>> http://engineering.klout.com/2013/01/introducing-
> brickhouse-major-open-source-release-from-klout/
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>> On Tue, Nov 12, 2013 at 7:28 PM, Timothy Chen <[email protected]>
> wrote:
> >>>>>
> >>>>>> Just saw this library today and thought it's something we can
> >>>> potentially
> >>>>>> leverage:
> >>>>>>
> >>>>>> https://github.com/addthis/stream-lib
> >>>>>>
> >>>>>> It has a number of algo for approximation streams and has code for
> >>>>>> cardinality estimation (HyperLogLog) and others.
> >>>>>>
> >>>>>> Looks like Twitter's SummingBird uses this library too.
> >>>>>>
> >>>>>> Tim
> >>>>
>

Reply via email to