Currently if we have 32 connections open to a single backend, and it goes down, all 32 will go through their delay and retry phases at about the same time. This can exacerbate problems with backend applications where accepting connections is expensive (e.g. because they fork()). Once we've exhausted the backoff policy we will suddenly transition down to a single retrying connection. It would be nice if we tapered this transition off (e.g. by halving the number of parallel retries to a single backend at each backoff step)
Currently if we have 32 connections open to a single backend, and it goes down, all 32 will go through their delay and retry phases at about the same time. This can exacerbate problems with backend applications where accepting connections is expensive (e.g. because they fork()). Once we've exhausted the backoff policy we will suddenly transition down to a single retrying connection. It would be nice if we tapered this transition off (e.g. by halving the number of parallel retries to a single backend at each backoff step)