We did things like incrementally rewrite a million-line binary from C++ into Java while it was running, or completely change the indexing system from batch processing to continuous updating, or grow the number of documents indexed from about 80B to 1T. There's a huge iceberg of development that you don't see, much of which has to do with scalability (you generally have to rewrite every time a key metric grows by a factor of 10) and much of which is experimental, trying out new features to see what resonates with the userbase.