Allow setting a message max size

[awaited-hare]

Feature Request

Motivation

See tensorflow/serving#1382, if the message size is large, tonic currently does not support changing the max message size.

Proposal

Expose the API the change allowed message size.

[akiradeveloper]

@l1an0 @LucioFranco What is the maximum message size allowed now? As far as I understand, gRPC's default max size in spec is 4MB but tonic server seems to be capable of receive larger requests.

[LucioFranco]

I believe there is a cap with h2, its a bug. related issue is here for that @akiradeveloper #352

[akiradeveloper]

@LucioFranco Thanks for you reply but my question is why tonic doesn't have an upper limit as official implementations do

https://github.com/grpc/grpc/releases/tag/v1.29.0

A significant change was made in #22575 where gRPC Core will check whether the decompressed message fits in the configured limit for the maximum message length that the channel can receive. (The default limit is 4MB.)

Is that your intention not to comply with the official behaviors?

[LucioFranco]

@LucioFranco Thanks for you reply but my question is why tonic doesn't have an upper limit as official implementations do

I don't believe this is specified in the protocol definition so it was never added initially. Would be happy to add it if people want this behavior.

Is that your intention not to comply with the official behaviors?

Just because gRPC core does it, doesn't mean its "official" per-say. Each implemenation of gRPC has its own niche things. I would love to add this but I don't have the time currently. I would totally accept a PR though.

[davidpdrsn]

It seems this was fixed by #529. Just re-open if that is not the case.

[sfackler]

Unless I'm misreading the spec, the h2 maximum frame size is unrelated to the maximum gRPC message size: https://github.com/grpc/grpc/blob/master/doc/PROTOCOL-HTTP2.md#data-frames. From some quick testing, it appears that tonic will consume arbitrarily large messages.

[LucioFranco]

Yes, I believe the C++ implementation has a non-h2 limit as well. We could add this via Grpc types.