That last sentence is the key. How many of us have actually written a good encoder? For instance, our sparse vectors don't use Golomb-delta encoding of indexes. They don't have a special case for binary data. They don't check the stats to see if zig-zag encoding of integers would help with the values.
The average encoding of data in Avro will be smaller than the average encoding of data as Writable simply due to naivete and laziness. I include myself in the class of people with both of those defects. On Sun, May 29, 2011 at 5:08 AM, Sean Owen <[email protected]> wrote: > Versus... Writable? no, the receiver has to know the Writable class in > advance and therefore knows how to decode. It's not embedded in the > serialization. Writable is nothing if not compact -- if you write a good > encoder that is. > > On Sun, May 29, 2011 at 12:42 PM, Grant Ingersoll <[email protected] > >wrote: > > > > > - data knows its own schema and will interoperate with any other > > compatible > > > schema > > > > > > > Isn't it also more compact due to it doesn't have to send encoding info? > > > > > > >
