Our compute node is basically a usual postgres where we intercept WAL write and page read streams. In other words we don't do any compute sharding since we want

Our compute node is basically a usual postgres where we intercept WAL write and page read streams. In other words we don't do any compute sharding since we want to preserve vanilla postgres compatibility and any sharded solution will be riddled with a lot of issues like you've mentioned.