jayant

thanks again for this, which I have now read about N times 

I think your approach and argument is sound -- but I do agree that getting in front of some users would help greatly.  to start with: will you be at kubecon?

alexis

On Tue, Oct 25, 2016 at 9:26 PM, Jayant Kolhe <jkolhe@...> wrote:

Thanks Alexis.

> Can you please list which ones are notable?  The main one seems to be "streaming" replies.  

Streaming is certainly notable but multiplexing and flow-control are also very important here. Without http2 flow-control features, deployments can easily experience OOMs, we have seen this in practice and so have community users. This is particularly important for infrastructural APIs. Header compression and multiplexing (i.e avoiding additional TLS negotiation) are very important for mobile clients and embedded devices.

> This shouldn't prevent someone building a http 1 compatibility layer, if that what helps people with adoption.

We are not opposed to it. We have not built it yet. We definitely considered it. However, a compatibility layer has significant complexity costs. The matrix of testing and verifying across all language implementations while having good performance seemed very expensive when we looked at it. It is not just http1.x you need to consider if seamless compatibility with existing ecosystem is needed. You have to consider what features of http 1.x work well in existing ecosystem. For example: to build streaming solution, gRPC protocol relies on trailers. While it is http 1.x feature, many proxies/libraries do not support trailers. To make it compatible with existing ecosystem, we need to then consider alternate schemes and have multiple schemes that work in different scenarios.

While implementing gRPC on top of HTTP1.1 downgrade is feasible, such implementation would lose many of gRPC advantages of efficient usage of network resources and high performance.

How much does that matter to real world users for the practical cases that such an implementation would facilitate?

When we talked to users, their biggest concern is migration from their existing system. That existing system has different payloads, different conventions. So having just http1.x transport did not suffice their usecases. Hence, many folks preferred a proxy solution that allowed them to support existing system and build a new system that interoperated with such existing system. It would be good to understand specific use-cases.

The protocol was designed with this in mind so we are certainly not opposed to it, it is certainly a subject we have heard a lot about from the community. It’s probably worth pointing out that we have tested this with curl and HTTP1 to HTTP2 converting proxies and it functions just fine. Whether the conversion/compatibility layer is built into gRPC libraries or a separate set of compatibility libraries or proxies is the decision point.

There are three primary things motivating this desire have come up in our conversations with users: browser support, the lack of upstream HTTP2 support in nginx and Cloud layer-7 LBs, and library platform coverage

We are working on a community proposal for a protocol adaptation to enable browser support that will allow as much of GRPC to work as can be enabled within the limitations of the browser platform, specifically the limitations of XHR. We have also been working with the browser vendors to improve their networking APIs so that GRPC can be supported natively at which point we should be able to phase out this adaptation. The GRPC to JSON-REST gateway pattern has also served the browser use-case quite well and is something we will continue to invest in.

We are actively working on getting HTTP2 upstream support into nginx. This has taken longer than we would like. In the meantime there are a large number of other proxy solutions of decent quality that are available to folks. We are working with the large Cloud vendors on this too.

With regard to library coverage GRPC now scores quite well against the large language X platform matrix of potential users so there are very few deployment scenarios which are not covered. The built-in HTTP APIs in many platforms in many cases are quite poor in terms of their ability to express the HTTP protocol and in terms of efficiency. There are many attempts to improve these APIs (Fetch API for browser & node, new HTTP APi proposal for Java 10 etc.) but they are some ways off. The protocol adaptation we intend for browsers can be used in the interim.

> I'm sure we can find CNCF community engineers who would be willing and able to have a go.  How hard can it be?

Based on our initial analysis, the complexity was to ensure support across all implementations. Our experience shows that interop matrix with features/testing and performance takes a lot of effort. We would be happy to support a community effort, but we should have a design conversation whether such support makes more sense in gRPC libraries or a set of separate accompanying libraries or proxies and verify it against different use-cases of users. We believe that the fundamental adoption issues can be addressed without inheriting a significant amount of protocol baggage.

> I'm interested to understand how you measure this.

I have not seen a comprehensive information on how much traffic is http2 vs how much is spdy vs how much is http1.x. Major browsers and most high traffic sites support http2 well. Some charts (https://www.keycdn.com/blog/http2-statistics/) I saw indicate good adoption, but I do not know how comprehensive those are. Internally at Google, we have seen largest traffic being http2 followed by spdy and quic.

> Not having an obvious Joe User adoption path will impede gRPC from being in some sense universal.

> It may also lead to people feeling let down if they hear (a) that gRPC is the new hotness, and that everyone

> should try it, but (b) it has compatibility problems that may not be resolved.

I fully agree with this concern and would love solutions/help. Whenever we considered it, adding more useful features and improving performance seemed to be more important to users than adding compatibility feature unless it gave them complete migration from their existing system. It may be due to set of users we talked to. I would love more input/data here.

On Tue, Oct 25, 2016 at 9:19 AM, Alexis Richardson <alexis@...> wrote:

Jayant

Many thanks.

On Tue, Oct 25, 2016 at 4:50 PM, Jayant Kolhe <jkolhe@...> wrote:

gRPC protocol was designed to build high performance, cross platform and usable libraries for building microservices. It was designed on top of HTTP2 to explicitly make use of

• Full duplex streams to support bi-directional streaming
  
• HPACK compressed headers for efficiently transmitting metadata/sidechannel information. For example, reduces cost for authentication tokens
  
• Connection multiplexing. Reduces the per-RPC connection cost for TLS and high latency connections
  
• Binary Framing layer with good flow control
  

Many of gRPC features work well only on HTTP2 semantics.

Can you please list which ones are notable?  The main one seems to be "streaming" replies.  This shouldn't prevent someone building a http 1 compatibility layer, if that what helps people with adoption.

 

While implementing gRPC on top of HTTP1.1 downgrade is feasible, such implementation would lose many of gRPC advantages of efficient usage of network resources and high performance.

How much does that matter to real world users for the practical cases that such an implementation would facilitate?

 

It would add significant complexity across multiple language implementations and would have higher bar and complexity for implementing and testing interoperability across these implementations.

I'm sure we can find CNCF community engineers who would be willing and able to have a go.  How hard can it be?

 

We have also relied on adoption of HTTP2 which has been very rapid and hence the ecosystem is also evolving rapidly to support HTTP2 features.

I'm interested to understand how you measure this.

 

We have also relied on proxies to provide this functionality to allow http1.x only ecosystem to work with gRPC. Such support exists in many proxies (nghttpx, linkerd, envoy) and is coming to others like nginx..Hence we have not implemented gRPC on HTTP 1.x.

Please don't take this the wrong way -- I like gRPC and am excited about it!

But: expecting proxies to solve this stuff, kind of undermines the whole approach.  Not having an obvious Joe User adoption path will impede gRPC from being in some sense universal.  It may also lead to people feeling let down if they hear (a) that gRPC is the new hotness, and that everyone should try it, but (b) it has compatibility problems that may not be resolved.

I vividly recall Brad Fitz telling me back in 2009 (or thereabouts) that, for HTTP, it is prudent to assume the worst when it comes to widespread adoption.  He pointed out that many servers & proxies still spoke 0.9 at the time.  

 

On Fri, Oct 21, 2016 at 11:42 AM, Brian Grant <briangrant@...> wrote:

+Varun and Jayant to answer that

On Fri, Oct 21, 2016 at 10:57 AM, Alexis Richardson via cncf-toc <cncf-toc@...> wrote:

I'd like to understand why gRPC doesn't work with HTTP 1.x

On Fri, 21 Oct 2016, 18:45 Ben Sigelman via cncf-toc, <cncf-toc@...> wrote:

Hi all,

"I am not on the TOC, but" I did want to share a few thoughts about GRPC per the call the other day.

I was recently at one of those moderated VC dinners where everyone gets put on the spot to say something "insightful" (sic) – I'm sure we all know the scenario. Anyway, we had to go around the table and talk about "the one OSS project that's poised to change the way the industry functions". There were lots of mentions of Docker, k8s, etc, and for good reason. I had the bad luck of being last and felt like it wasn't useful to just +1 someone else's comment, and I realized that GRPC was in many ways an excellent answer.

Varun alluded to this in his presentation, but to restate it in different words: the value of an RPC system is mostly not actually about the RPC... it's the service discovery, client-side load balancing, well-factored monitoring, context propagation, and so on.

In that way, a high-quality RPC system is arguably the lynchpin of the "user-level OS" that sits just below the application code but above the actual (kernel) syscalls. An alternative approach moves things like RPC into its own process (a la linkerd(*)) and I think that makes sense in certain situations... but when the RPC system depends on data from its host process beyond the RPC payload and peer identity (which is often the case for more sophisticated microservice deployments), OR when "throughput matters" and extra copies are unacceptable, an in-process RPC subsystem is the right approach.

As for whether GRPC is the right in-process RPC system to incubate: I think that's a no-brainer. It has good momentum, the code is of a much higher quality and works in more languages than the alternatives, and Google's decision to adopt it internally will help to ensure that it works within scaled-out systems (both architecturally and in terms of raw performance). Apache Thrift moves quite slowly in my experience and has glaring problems in many languages; Finagle is mature but is limited to JVM (and perhaps bites off more than it can chew at times); other entrants that I'm aware of don't have a strong community behind them.

So yes, this is just an enthusiastic +1 from me. Hope the above makes sense and isn't blindingly obvious. :)

Comments / disagreements welcome –

Ben

(*): re linkerd specifically: I am a fan, and IMO this is a "both/and" situation, not "either/or"...

ᐧ

_______________________________________________
cncf-toc mailing list
cncf-toc@...
https://lists.cncf.io/mailman/listinfo/cncf-toc

_______________________________________________
cncf-toc mailing list
cncf-toc@...
https://lists.cncf.io/mailman/listinfo/cncf-toc