CAPI replication fails repeatedly because of incomplete http response

When running CAPI against elastic search, I noticed that replication failed repeatedly. After XDCR sends http requests to target cluster to update documents, it tries to read http response. Sometimes it gets only the first half of the response. Next time it reads the response, the second half of the previous response is prepended to the second response, resulting in an invalid http response and causing XDCR to fail eventually.

The buffer that XDCR allocates to hold the response is too small. After increasing it to the correct size, the frequency of the problem was greatly reduced. Strangely, it did not go away completely.

I tried several different ways of getting http response and none worked. The incomplete response is not caused by buffer being too small or read time out being too small. Need some more investigation.

The easiest way to work around this problem is to let CAPI reset tcp connection when an incomplete response is seen.