When ran with TSAN/ALUBSAN etc. processes fork another process to resolve their symbols from addresses. These sub-processes are quite memory hungry depending on build model. ArangoDB cluster tests launch about half a dozen processes, and if each of those starts it's own symbolizer, then these add up to a larger amount effectively draining the host out of resources during CI-Runs or even local runs on bigger machines. However, these forked sub processes all work on similar if not the same binaries, hence could we only have one of these resource hoggers instead of many? Along comes the symbolizer server and client solution - one symbolizer is spawned by the server process, communicating via pipes. This server then runs a simple Python HTTP server to process POST requests from the clients:

./utils/llvm-symbolizer-server.py

Now only all the other processes need to be told that they should only launch a tiny python client (llvm-symbolizer-client.py) instead of running their own symbolizer by the respective environment variables. These client processes implement the same stdin/stdout protocol as the llvm-symbolizer, but then forward the request via HTTP to the server process: export TSAN_OPTIONS="external_symbolizer_path=$(pwd)/utils/llvm-symbolizer-client.py"

The resulting python processes will not be resource intense in comparison to the instrumented processes, hence no further optimisation like implementing it in GO were untertaken.