Just a guess: the HTTP connection to Mailgun probably gets closed while in the pool. E.g. because it wasn’t used for a long time. And it probably happens to many connections in that pool, so it takes that many retries to find one that’s still connected.
