I think what makes the most sense is to limit the number of connections to another host. A host only has so many CPU resources, and beyond a certain point throughput would start to suffer anyway (and then only make the problem worse). It also makes sense in that a client could generate documents faster than we can index them (either for a short period of time, or on average) and having flow control to prevent unlimited buffering (which is essentially what this is) makes sense.
Nick - when you switched to HttpSolrServer, things worked because this added an explicit flow control mechanism. A single request (i.e. an add with one or more documents) is fully indexed to all endpoints before the response is returned. Hence if you have 10 indexing threads and are adding documents in batches of 100, there can be only 1000 documents buffered in the system at any one time. -Yonik http://lucidimagination.com
