Handle case where server doesn't support pipelining? about undici

[ronag]

Possibly an edge case not worth fixing... I think this mostly happens with very old servers.

[ronag]

I mean we could just explicitly state in the README that persistent connection and pipelining must be supported by the target host. Is that good enough?

[ronag]

Yup. What happens otherwise? Does it throw when there already requests queued?

Everything inflight fails except for the first request. So with pipelining=1, every other request would fail (if the queue is full).

[mcollina]

How would the server communicate that back to the client?

[ronag]

The server may send a "Connection: Close" header after the first request is sent indicating it does not want to use a persistent connection.

[ronag]

So basically the logic here would be that the client does not pipeline until the first request has succeeded with a persistent connection in order to indicate that the server seems to support pipelining.

[mcollina]

That would be nice. However I fear it's a lot of complexity.
How are other clients doing it, if at all? Isn't this part of HTTP/1.1 that server has to support?

[ronag]

Probably, I’ll take a look. This might be another one of those edge cases.

Sorry to bring up every possible edge case. Just trying to be throughout.

[ronag]

How are other clients doing it, if at all? Isn't this part of HTTP/1.1 that server has to support?

@szmarczak do you have any input?

[szmarczak]

Isn't this part of HTTP/1.1 that server has to support?

No. Pipelining happens only on a persistent (keep-alive) connection and is optional. Unfortunately the HTTP/1.1 RFC does not specify how to verify whether the server supports pipelining or not and how many concurrent requests at maximum it should have.

[ronag]

Unfortunately the HTTP/1.1 RFC does not specify how to verify whether the server supports pipelining or not and how many concurrent requests at maximum it should have.

I mean... you can kind of detect it by sending a request to the server and see what it does.

So basically the logic here would be that the client does not pipeline until the first request has succeeded with a persistent connection in order to indicate that the server seems to support pipelining.

[szmarczak]

you can kind of detect it by sending a request to the server and see what it does.

One request is not sufficient. Because the connection is keep-alive it doesn't mean that it supports pipelining (according to the RFC). But I too think that most servers do support pipelining. To check if the server supports pipelining, you need to make two requests.

I see your point. I say let's assume that every keep-alive server does support pipelining, then only one request is enough to make the check.

[szmarczak]

Yup. What happens otherwise? Does it throw when there already requests queued?

[szmarczak]

Seems to be fine :)