It seems we need to use the upstream ID instead of type as the prefix?

key => upstream ID + source
version => tostring(up_conf)

I remember that upstream IDs conflict with discovery?

when upstream changed in etcd, upstream changed in worker process like this: upstream-A-v1 -> upstream-A-v2 -> upstream-A-v3 -> upstream-A-v4 upstream-B-v1 -> upstream-B-v2 -> upstream-B-v3 -> upstream-B-v4 v1 ,v2, v3 have been GC，Is it possible that upstream-B-v4 address use the upstream-A-v1 address?

That's might be possible but what's the influence? Old v1, v2 and v3 objects are GCed, it means that we never use it, in such a case, a new object reuses the old object's address looks fine?

when upstream-B-v4 address use the upstream-A-v1 address, if picker's lrucache like this:
upstream-A-v1 address ---> upstream-A-v1 picker
the traffic proxied to upstream-B-v4,will use the upstream-A-v1 picker ,this will occur error