Description:
See #2063 and the 10k testcase. In that testcase, with #2358 applied, the 30 second startup time is dominated by regex code associated with tags.

I did a little debugging to try to understand the tags, and I still don't. But my naive observation is that we are applying a bunch of regexes to stat names to do this transform (in Counter& ThreadLocalStoreImpl::ScopeImpl::counter():

input:
    name = "http.admin.downstream_cx_http1_total"
outputs:
   tag_extracted_name = "http.downstream_cx_http1_total"
   tags = vector {{
      name_ = "envoy.http_conn_manager_prefix", 
      value_ = "admin"
   }}

My feeling is that this is a structural transformation that could be done much faster if the name structure is retained in the stat structures, rather than flattening the string in stats and then rediscovering. But I don't have enough of a feel for whether this alone would make startup fast, or whether the effort to fix would be worth instant startups on huge numbers of clusters.