I mean…the cloud message queues that use HTTP are not good examples of quality software. They all end up being mediocre to poor on every axis: they’re not generalizable enough to be high quality low level components in complex data routing (e.g. SQS’s design basically precludes rapid redelivery on client failure, and is resistant to heterogenous workloads by requiring an up-front redelivery/deadletter timeout); simultaneously, HTTP’s statelessness at the client makes extremely basic use cases flaky since e.g. consumer acknowledgment/“pop” failures are hard to differentiate as server-side issues, incorrect client behavior, or conceptual partitions in the consume transaction network link…”conceptual” because that connection doesn’t actually exist, leading to all these problems. Transactionality between stream operations, too, is either a hell-no or a hella-slow (requiring all the hoop-jumping mentioned in TFA for clients’ operations to find the server session that “owns” the transaction’s pseudo connection) if built on top of HTTP.

In other words, you can’t emulate a stateful connection on top of stateless RPC—well, you can, but nobody does because it’d be slow and require complicated clients. Instead, they staple a few headers on top of RPC and assert that it’s just as good as a socket. Dear reader: it is not.

This isn’t an endorsement of AMQP 0.9 and the like or anything. The true messaging/streaming protocols have plenty of their own issues. But at least they don’t build on a completely self-sabotaged foundation.

Like, I get it. HTTP is popular and a lot of client ecosystems balk at more complex protocols. But in the case of stateful duplex communication (of which queueing is a subset), you don’t save on complexity by building on HTTP. You just move the complexity into the reliability domain rather than the implementation domain.