Fragmentation of REQUEST_STREAM and REQUEST_CHANNEL frames

[fiorix]

In case the REQUEST_STREAM or REQUEST_CHANNEL frames must be fragmented (containing the F flag), should it contain the Initial Request N field in all fragments? Seems like a waste of resources to xmit/recv the same field over and over.

It'd make more sense to only send the Initial Request N field on the first frame, but this approach implies having a state machine to handle reassembly.

Thoughts?

[yschimke]

I've been assuming that only the first frame is of type REQUEST_STREAM or CHANNEL, and all additional request fragments are PAYLOAD frames. why repeat the initial request header and allow for conflicting values? I don't think the protocol is clear enough. This part of the CHANNEL doc implies my thinking is correct, but its a weak assumption.

A requester MUST send only one REQUEST_CHANNEL frame. Subsequent messages from requester to responder MUST be sent as PAYLOAD frames.

[fiorix]

Maybe we should clarify that this is also true for other request* frames that must be fragmented?

The frames that allow the F(ollow) flag today are:

• REQUEST_RESPONSE (0x04)
• REQUEST_FNF (0x05)
• REQUEST_STREAM (0x06)
• REQUEST_CHANNEL (0x07)
• PAYLOAD (0x0A)

This allows the potential scenario of:

-- PAYLOAD frame 1
Frame type = One of the supported above: 0x04 - 0x07, 0x0A
Frame length = 16MB
(M)etadata present = 1
(F)ollows = 1 (fragments coming)
Metadata Length = 16MB

16MB of METADATA
0MB of Data

-- PAYLOAD frame 2
Frame type = PAYLOAD (0x0A)
Frame length = 16MB
(M)etadata present = 1
(F)ollows = 1 (fragments coming)
Metadata Length = 4MB

4MB of METADATA
12MB of Data

-- PAYLOAD frame 3
Frame type = PAYLOAD (0x0A)
Frame length = 13MB
(M)etadata present = 0
(F)ollows = 0

0MB of METADATA
13MB of Data

Also perhaps add @tmontgomery's comment on cancellation of fragmented streams? This means frames 2 or 3 above could be either PAYLOAD (0x0A), CANCEL (0x09), or ERROR (0x0B).

[benjchristensen]

I generally like this, though it feels like an exception in this spec to use PAYLOAD for anything other than the stream messages. But, I do agree it's cleaner than re-sending all other aspects of the initial request frame.

[benjchristensen]

@tmontgomery and @lehecka, what do you think?

[benjchristensen]

Proposed changes in #216

[benjchristensen]

merged.