Thread by @QuinnyPig

Made it to the #reinvent "Monday Night Live with Peter DeSantis."

The band is playing a rousing ode to the @awscloud bill, "You Spin Me (Right Round Baby)."

There's a lot of neat high-level stuff that @awscloud does (and yes, some ridiculous things to), but it all distills down to this: the infrastructure. And I'd say that if AWS misses being "the best in the world at it" it's by a hair's breadth at absolute most.

It is currently very very loud and there are bright lights shining in my eyes constantly.

THIS MUST BE WHAT IT'S LIKE INSIDE OF AN AWS DATA CENTER

"Monday Night Live" at #reinvent goes until 9:30 or so. This of course leads directly to "Tuesday Morning Dead."

Performance, cost, and security are in tension.

Some providers resolve that tension by just not giving the slightest possible shit about any of those three things. This is called "Clown Computing."

We begin by talking about Nitro.

Introducing the Nitro v5 chip. I don't recall seeing specs on previous Nitro chips.

To my understanding, these things manifest like-- oh, yes. There's the slide with the real world impact.

NEW for #reinvent: c7gn instances. These new Nitro chips, Graviton3, enhanced networking. Available in preview today.

(Graviton3 was announced and available at re:Invent last year. They still (strangely) talk mostly about Graviton2 in most conversations.) #reinvent

NEW: HPC7g EC2 instances or possibly HP's latest printer offering because who the hell even knows how these things get named anymore. Certainly not I. #reinvent

Talking now about SRD, their own internal data center networking protocol.

@awscloud is very operationally good and very scaled out. They should write their own networking protocol.

You are probably not and should certainly not, respectively.

Talking about their network topology. "We're able to do this cost effectively--"

Cloud Economist says: CITATION NEEDED

tl;dr: SRD is purpose built for the AWS network, thus it doesn't have to deal with things like "interoperability," "the Internet," "divergent network conditions," and "Comcast."

Now he's talking about EBS.

It's easy to forget that your EC2 instance talks to its EBS volumes *over the network*. #reinvent

4x EBS throughput improvement. So starting in early 2023, all new EBS io2 volumes will be running on SRD.

Now talking about the default EC2 network interface, ENA.

ENA Express is now available in GA today. Brings SRD to EC2 transparently. No software, instance, or config changes needed past *click the box to enable it.*

Yup, sure enough there it is:

Messaging is hard; a more accurate title for what this keynote has pivoted into would have been "Monday Night Computer Science Lecture with Peter DeSantis." #reinvent

NEW: a Trn1n (pronounced "turnin") instance that network optimizes Machine Learning® training instances. #reinvent

Now a Ferrari engineer here to talk about how racing F1 cars (incredibly risky) interacts with cloud computing (incredibly financially risky). #reinvent

Seems like I'll have a new shitposting sporting interest next season!

https://twitter.com/essjaytee/status/1597445438382215168

Now Peter's back to talk about AWS Lambda.

Today, Lambda is in use by over 1 million customers, making over 10 trillion monthly requests.

CLOUD ECONOMICS: That's $2 million in Lambda request charges a month. #reinvent

"I tried and failed to find a picture of a castle with lots of entry points."

I'd have gone with a sandcastle. #reinvent

Describing Lambda as a big cache of compute resources. #reinvent

Ooh. Originally the Lambda compute cache was composed of t2 instances. Each t2 instance was used for one customer's function; that's how they got good performance but also hard isolation between tenants.

#reinvent

Ahhh. Describing it as a compute cache makes sense--a cache miss is a cold start!

The next improvement was to use Firecracker* for isolation; microVMs that launch within fractions of a second. #reinvent

* Firecracker. Not Amazon Firecracker, nor AWS Firecracker. Just Firecracker. Or @awscloud senior engineers get upset.

Ooh. The Cold Start lifecycle.

"So what if we snapshotted post initialization?"
INTRODUCING LAMBDA SNAPSTART
Available now in GA, no additional cost.

#reinvent

I wonder if that means that during the initialization that anything instantiated outside the handler is now static and reused?

...and that's the first thing he's addressing. Of course AWS thought about that...

If the seed is the same, the random numbers become deterministic. Therefore the seed is explicitly changed after it's reinstantiated. Your code has to be snapstart safe. Okay, that's a caveat.

Nothing said yet about things like "making fresh calls to Secrets Manager, establishing a database connection, grabbing and stuffing dynamic data into /tmp," etc. Will have to wait to read the docs to make sense of those cases...

This part's an @awscloud-side problem. Customers don't have to do anything specifically here to my understanding...