The trick was to run many HTTP requests in parallel, using different subdomains so they wouldn't be subject to the browser limit on number of parallel connections to a domain, a dummy first request per subdomain so the interesting messages went over established persistent connections, and decoupling outbound from inbound messages to make a set of "short-long-polling" connections.

That resulted in real-time message throughput close to the game frame rate, bidirectional, and no connection setup latency per message, just the transmission latency even over TLS (SSL back then).