Description
Looking at the logs of RBH<>WBH complex relay (4 nodes connected, 4 messages race relays, 2 on-demand header relays, 2 on-demand parachain relays, version guards, metrics), we're doing ~30 RPC requests per second. We don't care much about relay performance, but it looks too much :) Most of requests are method=chain_getBlockHash and method=chain_getHeader. That's because every relay (and we have 8 relays running) are polling nodes for new headers. We may have some cache for these results.
Ideally we'll need to traitify the whole Substrate Client. This would give us possibility to use it in tests => allow to test client-specific code. And when we talk about this issue, we may have some CachedClient that nests the regular Client and maintains some results cache for some (or all) calls.