Cathy Zhang <Cathy.H.Zhang@...>
toggle quoted message
Show quoted text
+1 on a CNCF WG. Will be happy to contribute to it.
I see that different people have different perspectives on serverless or FaaS.
I think it will he helpful to define the semantics of serverless and the scope of serverless (I agree with Dan that we may need a better terminology) in CNCF
so that we can all be on the same page when diving into serverless tech or implementation.
There are already some open source serverless implementations. Most of them concentrates on event driven function hosting and execution, similar to AWS Lambda
function. But in reality there are many use cases that involve multiple functions executing in sequence or in parallel or different function execution upon different events or a function execution triggered by a combination of events, etc.. CNCF can play
a key role in “standardizing” a comprehensive and generic event function execution model, something like a “event+function” graph so that whatever the backend implementation, the user-facing front end will not change.
From: cncf-toc-bounces@... [mailto:cncf-toc-bounces@...]
On Behalf Of Mark Peek via cncf-toc
Sent: Friday, March 31, 2017 2:27 PM
To: Yaron Haviv; Dan Kohn; Alexis Richardson; cncf-toc@...; Ryan S. Brown; Brian Grant
Subject: Re: [cncf-toc] The Cloud-Nativity of Serverless
+1 on a CNCF WG for serverless. Per a side conversation with Alexis this should be discussed at the TOC so it would be nice to have this added as an agenda item for an upcoming
I agree with many of the points, had a bunch of 1:1 chats @ KubeCon on the need for more colaboration in that space
There will be a bunch of implementations by cloud and OpenSource (I presented ours, focused on RT data processing)
And its critical that we have cross platform APIs, the flexibility to use different tools, and allow incremental innovation.
now every “serverless” project 100% overlaps the other, CNI/OCI as good examples for why we should have layering vs monoliths
We can break Serverless tech to few sub-categorize e.g.:
Function APIs/patterns (e.g. definition of context, events, logs, ..)
“Serverless Run-time” (implementation), I guess every cloud will have its own and we will see few open with different focus areas
Edit, build, test, tools (and how they integrate with the run-times & APIs ..)
Event/data sources (API for things that drive function invocations)
Its in CNCF/LF interest to foster collaboration across those areas, allow different build tool to work with multiple run times, and make sure developers don’t have to use
different code for each run-time or cloud vendor. I see nice concepts in the different implementations and would be great if we can sit and agree what are the things we want to adopt in a common definition or make optional.
The best next step is to form a CNCF WG/SIG, happy to contribute to such.
I’m not a fan of the term “serverless” as most people tend to think “no servers” as opposed to “fewer” or “not having to deal with” servers. Of course, FaaS doesn’t quite
roll off the tongue either. Definitely be nice to have better naming around it.
Looking at this from an on-prem or private cloud perspective, you don’t get the same benefits as public cloud FaaS. Instead of having all of the infrastructure handled by
the public cloud, the private clouds will require ease of operation and lifecycle management for the FaaS operation. And will have additional requirements around multi-tenant, RBAC, and security.
While there will be multiple FaaS implementations (stand alone or on top of k8s), as Brian alluded to we should be looking at common eventing models/function execution, interoperability
of the services, and common SDK’s. This would allow for the functions to have portability in mind across a variety of FaaS implementations. Also, I believe this is very pertinent to cloud native as teams are using the a variety of compute (VM/instances, containers,
and FaaS) to design, develop, and deploy their applications. In other words, FaaS is a cloud native design pattern that needs to be supported across all clouds.
And, since I’m replying to Dan’s email I will add this AWS re:invent video link where Coca Cola is using AWS Lambda to run payments from their machines. The beginning talks
about their rationale and goes into pricing break evens on when to switch to dedicated instances. Of course the use of AWS services always need to be kept in mind.
AWS re:Invent 2016: Coca-Cola: Running Serverless Applications with Enterprise Requirements (SVR303)
Yes, I find this story inspiring from Benchling of moving their genome searching to Lambda and both reducing tail latency and dropping costs from thousands of dollars per month to $60:
On Mar 31, 2017, at 14:10, Alexis Richardson via cncf-toc <cncf-toc@...> wrote:
Related to this, many FaaS proponents talk about the economics of only paying for use (function calls). But this economic model is not limited to the serverless app frameworks eg as listed below by Brian. Many container orchestrators
could feasibly provide fine grained billing.
On Fri, 31 Mar 2017, 19:00 Brian Grant via cncf-toc, <cncf-toc@...> wrote:
I don't find the term "serverless" to be useful. It's too broad, and could encompass purely client-based computations and web hosting as well as FaaS.
I see Functions as a Service as an instance of Application Framework as a Service. Data-processing frameworks that support dynamic code loading and managed execution would also
qualify. The services that we have today that support dynamically pushed code are natural evolutions of servlet engines, application frameworks, data-processing frameworks, and plugin-centric systems that have been developed over the
past couple decades.
Even Functions as a Service specially addresses multiple overlapping areas:
1. Event-driven automation. People are using FaaS for simple automation tasks. For these scenarios, the most critical determinant of utility is relevant event sources. In the FaaS space,
Openwhisk is most obviously aimed at this use case. Automation systems such as StackStorm are
fairly similar. The main limitation of a system like Stackstorm is that the actions are pre-packaged plugins rather than dynamically provided functions. IFTTT and Microsoft Flow address points in this spectrum, as well, and configurable actuators capable of
talking to any OpenAPI-compatible API are one reasonable method of linking triggers and actions.
2. Extension implementations. Something to receive extension web hooks from some other service without the need to operate a full-blown application deployment. The main difference with
the event-driven automation use case is that some other system defines the invocation conditions.
3. Data-driven processing and simple ETL workflows. Not unlike Bigtable
4. Rapid application development and deployment, especially for mobile apps, home assistants, and IoT. It's similar to website hosting, but for application
frameworks / servlet engines. As with web app mashups, this model is facilitated by the availability of APIs for other services to do much of the heavy lifting. The line between this scenario and a full-blown PaaS is not about features, agility, the deployment
model, or execution artifacts (most PaaSes support pushing code, and people load and run executable binaries on FaaS), but about who operates the deployed application servers.
Container-based technologies are still improving and I think you'll find that container-centric infrastructure will eventually provide most of the core infrastructure capabilities required by FaaS.
Is FaaS "cloud native"? Yes.
Does FaaS make sense in local development, on prem, hybrid and multi-cloud scenarios? Yes, for all the same reasons that Kubernetes does.
On Fri, Mar 31, 2017 at 9:43 AM, Ryan S. Brown via cncf-toc
On Fri, Mar 31, 2017 at 11:37 AM, Anthony Skipper
> We would like to see a separate group working on serverless as well. At
> Galactic Fog we have had a serverless implementation on DCOS for about 6
> months, and we plan to release our Kubernetes native implementation in the
> next couple weeks in the runup to dockercon.
> From our perspective we would like the following things:
> Agreement on marketing terms. (Call it Serverless or Lambda, everyone
> hates FAAS, but serverless is problematic as well)
Agreement on these terms is probably a bit much to expect. For some
time I was hoping we'd settle on "Jeff". While I'm not a lawyer,
Lambda seems like the kind of thing that will turn into a trademark
issue at some point. I think we're stuck with serverless, and when
offering components that fit in a serverless stack we'll have to stick
with things like "serverless function runtime," FaaS, and similar with
a mind to two different audiences.
Audience A: Technical audience, knows serverless well, and wants to
know exactly what piece your project is providing. So you can say
things like "event router" and function runtime to explain where it
fits exactly. This audience also has some potential contributors in it
if the project is OSS.
Audience B: Thinks of serverless-the-concept as it relates to
developer experience, and would be looking to figure out what they can
do with it generally. The focus for those materials has to be on
distinguishing from plain containers, PaaS, etc more than on the
underlying thing your project is going to provide. Already it's
getting kind of muddy, since Amazon and others are rebranding other
aaS offerings as "serverless," such as DynamoDB.
> Agreement on core capabilities, from our perspective they are:
> Runtime Support
> API Gateway Support
> Config / Secret Capabilities
> Security Implementation
> Logging Support
> Monitoring Support
> Performance/Scalability Capabilities (eg. Gestalt and Fission are a couple
> order of magnitude faster than Amazon, and that changes the art of the
I agree with these, but I'd put performance as non-core because there
are plenty of workloads where it doesn't matter all that much. Think
about the class of back-office examples that are common: transforming
streams, resizing images, propagating changes to other systems. As
long as they get done, the difference between 100ms and 1000ms can
pass unnoticed since each event is eventually spawning a new function,
and the queue/event system handles backpressure transparently.
Then there's the category of user-facing synchronous workloads that
you'd see an API Gateway used for, where perf matters and users just
abandon anything that's perceivably slow.
> None Core Capabilities
> Ability to inter-operate between serverless implementations (eg, migration
> between them, include up to ad back from public cloud)
> Lambda Chaining
> Data management capabilities (exposing filesystems or other services in)
> Making the implementation of the serveless solution portable across
> Data Layer Integration approaches.
I'd probably bump chaining up to core, since all but the very simplest
projects end up with a series of functions that either call each
other, or create events that invoke others.
> I wouldn't worry to much about the other big vendor stuff right now.
> Serverless is at such an early stage any R&D done by anyone is really
> helpful and not really competitive or problematic. (eg Openwhisk has
> really cool ideas, and Amazon's attempts to standardize lambda portability
> show an approach that is helpful for discussion)
> On Fri, Mar 31, 2017 at 11:17 AM, Ryan S. Brown via cncf-toc
> <cncf-toc@...> wrote:
>> Hello all,
>> If haven't heard Amazon&others raising a general ruckus about serverless
>> lately, I sincerely hope your vacation to the backwoods was relaxing.
>> I'm Ryan, and I've been interested in FaaS/serverless for a while now.
>> Also CC'd on this message are Ben Kehoe (iRobot) and Peter Sbarski
>> (ServerlessConf/A Cloud Guru). Lately, it seems the open-source interest has
>> been picking up significantly in addition to all the use in the public
>> cloud. Just to name a few FaaS/serverless provider projects: Fission &
>> Funktion on Kubernetes, FaaS on Swarm, and standalone OpenWhisk
>> (primarily IBM-driven). Even Microsoft's Azure Functions is OSS.
>> A cynical observer might say that the MS/IBM efforts are open to help
>> compensate for them starting so late relative to Lambda, but either way the
>> result is a lot of open or nominally open projects in the FaaS/serverless
>> area. And with cloud providers looking to embed their various FaaS deeper
>> into their clouds by integrating their FaaS with cloud-specific events,
>> making their FaaS the way into customizing how their infra reacts to events.
>> So why am I writing this email? Well I've been thinking about serverless
>> as the next step in "cloud native" developer tooling. Look back to the state
>> of the art in the 00's and you'll see the beginnings of
>> autoscaling/immutable infrastructure, then move ahead a bit to containerized
>> applications, then container schedulers, and you can see a trend towards
>> shorter and shorter lifespans of persistent machines/processes.
>> Function-as-a-Service is another step in that direction where containers
>> live for seconds rather than persistently listening. This trajectory seems
>> pretty intuitive as a developer: as lower layers of the stack become more
>> standard I should be able to automate/outsource management of them.
>> I'd like to help the TOC think about where (or whether) serverless/FaaS
>> should fit into the CNCF's plans for the future. Do you want to talk about
>> what serverless actually is? Figure out how various OSS fits into a
>> serverless ecosystem? Compare how FaaS provided in the public cloud differs
>> from what users need in a hybrid/on-prem environment? Ask away - Ben, Pete,
>> and I are all here to help out.
>> Ryan Brown / Senior Software Engineer / Red Hat, Inc.
>> cncf-toc mailing list
Ryan Brown / Senior Software Engineer, Ansible / Red Hat, Inc.
cncf-toc mailing list