Actually this is a bit inaccurate. _Some_ implementations are not implemented as a sink.

Also, you can in fact instantiate the sinks yourself as well, as in

   readings.addSink(new CassandraTupleSink(<query>, <builder>);


On 22.11.2016 09:30, Chesnay Schepler wrote:
Hello,

the CassandraSink is not implemented as a sink but as a special operator, so you wouldn't be able to use the
addSink() method. (I can't remember the actual method being used.)

There are also several different implementations for various types (tuples, pojo's, scala case classes) but we did not want the user to be aware of it. This has the neat property that we can change the underlying classes any way we want (like modifying the constructor) without breaking anything.

Regards,
Chesnay

On 22.11.2016 08:06, Stephan Epping wrote:
Hello,

I wondered why the cassandra connector has such an unusual interface:
CassandraSink<Reading> csink = CassandraSink.addSink(readings)
while all other sinks seem to look like

RMQSink<Reading> sink =new RMQSink<Reading>(cfg, "readings_persist_out", new 
JSONReadingSchema());
readings.addSink(sink);
best,
Stephan





Reply via email to