[ https://issues.apache.org/jira/browse/THRIFT-4591?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16533561#comment-16533561 ]
James E. King III commented on THRIFT-4591: ------------------------------------------- Also I found this in the python implementation: {noformat} # N.B.: Doing this string concatenation is WAY cheaper than making # two separate calls to the underlying socket object. Socket writes in # Python turn out to be REALLY expensive, but it seems to do a pretty # good job of managing string buffer operations without excessive copies {noformat} So you may be right in that two socket writes is more expensive computationally. However, unless the language is really smart about it, adding two buffers usually results in a new buffer, which means you need twice the memory to write the data and that isn't good. > Incompatibility using non-blocking server and frame transport on C++ side? > -------------------------------------------------------------------------- > > Key: THRIFT-4591 > URL: https://issues.apache.org/jira/browse/THRIFT-4591 > Project: Thrift > Issue Type: Bug > Components: C++ - Library > Affects Versions: 0.11.0 > Reporter: allen_lee > Assignee: James E. King III > Priority: Blocker > Attachments: 9090.pcap, 9090_1.pcap > > Original Estimate: 4h > Remaining Estimate: 4h > > 1) realize thrift server with TNonblockingServer via c++; > 2) realize thrift client via lua lib and choose frame transport. > 3) call remote interface failed with "TTransportException:0: Default > (unknown)" print, and the server show "TConnection::workSocket(): > THRIFT_EAGAIN (unavailable resources)" error. > 4)investigate this fault with tcpdump tool, attachment 9090.pcap show the > frame msg doesnot contains frame size field, the rifht situation of > attachment 9090_1.pcap show the frame msg contains 4 bytes (00 00 00 25) > before protocol id field. > 5) dig into the fault and tried to find root cause, then i found there is an > fault in TFramedTransport:flush function in TFramedTransport.lua file. the > original realization is: > ----- > function TFramedTransport:flush() > if self.doWrite == false then > return self.trans:flush() > end > -- If the write fails we still want wBuf to be clear > local tmp = self.wBuf > self.wBuf = '' > local frame_len_buf = libluabpack.bpack("i", string.len(tmp)) > self.trans:write(frame_len_buf) > self.trans:write(tmp) > self.trans:flush() > end > ----- > which send frame size file and reset msg content independently. -- This message was sent by Atlassian JIRA (v7.6.3#76005)