Hey everyone, welcome to this week's

Azure update. It's the 3rd of April. As

always, we have the chapters so you can

jump to any particular update you care

about the most.

New videos this week. So obviously you

should go to the movies and see Project

Hail Mary. It was phenomenal. The book

was great. The movie was great. So it's

not one of my videos. My videos

I dove into Work IQ.

So when we think about how we work, not

just the knowledge, the artifacts in our

documents, our chats, and our meetings,

but a personalization that learns how

artifacts relate to each other,

how people relate to them, how people

relate to people,

uh

how we talk, how we function, our rhythm

of business,

and then special inferencing engines and

skills around document creation and deep

knowledge retrieval, and creating really

good PowerPoints. That's all part of

Work IQ that obviously

the Microsoft Copilots they're grounded

on, Co-work is grounded on, but you can

ground your own agents.

And then what is Copilot Co-work?

And I don't often say this, but it's a

really, really cool demo. About 8

minutes in if you want to skip the video

but go and look at the demo

to show how the Co-work capability

can just be given some outcomes you

want. I can interrupt it. It's running

as a cloud agent, but it's using these

long-running deep reasoning models to go

and do a bunch of cool stuff including

creating a web app all with a single

prompt. So it's kind of cool.

And then I kind of finished a flow of

videos I was trying to create around AI

and the Microsoft AI ecosystem. So I

just created sample.ai. It's just my

YouTube stuff. It's about 15, 16 videos,

but it's a very curated path if that's

useful trying to learn. Very much the

Microsoft focused

way of leveraging AI.

All right, so onto what's new on the

computer side.

So the Azure Red Hat OpenShift offering.

So that's the jointly created, managed,

supported offering between Red Hat and

Microsoft. Well, this is now available

in Indonesia Central. And so this is

really useful to have in new regions

when I think about I need to run a

workload closer to the customers maybe

from a latency perspective, but it may

also be around meeting different

regulatory requirements.

And this is kind of an interesting one.

So for virtual machines and virtual

machine scale sets, there's now a full

caching capability for ephemeral OS

disks. And this is in preview. So

remember the ephemeral OS disk is where

it's not using a durable managed disk.

It's that it's going to create the the

disk on local host resources. The the

temp space, the cache space.

So what that means is ordinarily

the rights we make to that ephemeral OS

disk go to the local storage in the node

that the VM runs on,

but the actual main OS image it's

reading from is still remote.

So what this new feature does is it

caches the entire OS disk to the local

storage. That removes any remote storage

dependency. So that's going to obviously

increase the resiliency of that running

VM from any remote storage failures, but

also improve the performance with the

super, super low latencies because it's

just local running on that storage. Now

the way it's going to work is the

caching occurs in the background once

the VM boots. So obviously there's a bit

of time for it's going to cache that.

And it's just an enable full caching

flag that I set for the ephemeral OS

disk.

And remember why we use ephemeral OS

disks. It's where there's nothing in the

OS disk state that we care about. It's a

tin soldier. Something's wrong with it,

we just rebuild another one in its

place. It's really common with virtual

machine scale sets where we constantly

create and delete the virtual machines.

And so I don't want to pay for a managed

disk. I want better performance.

Ephemeral OS disks are fantastic for

that.

Networking.

So App Configuration now has an Azure

Front Door integration in preview.

Remember App Configuration enables us to

have all of those different

configurations for the applications to

be centrally managed,

and then it can

be used and delivered to client

applications.

Well, with the integration with Azure

Front Door, which remember Azure Front

Door is that global layer 7 anycast

split TCP

delivery layer with caching,

well, when it integrates with that,

it now enables the scaling of the

delivery of those configurations to be

in terms of millions of clients. And I

don't have to try and develop my own

proxy layer anymore. And obviously this

can work with single page apps, so the

spa, the mobile apps, a whole bunch of

other different scenarios. And what's

going to happen here is

the App Configuration endpoint is

established on Azure Front Door,

and then the Azure App Configuration

store is set as its origin.

So you use managed identity to secure

that communication,

and then Azure Front Door will simply

retrieve the selected key values,

feature flags, um whatever is required.

It can then cache and respond it. So

it's it's avoiding you having to worry

about secrets and other things, but

gives you a really high scalable secure

service.

On the storage side,

so premium SSD V2 is in a new region in

GA. Remember

premium SSD V2 is all about it's a

sub-millisecond latency,

but you have separate dials for the

IOPS, the throughput, and the capacity.

And I can dynamically change the IOPS

and throughput while it's being used. So

now we see these in South India and US

Gov Arizona.

So when I think about scenarios where

I need high IOPS, I need high latency

databases, big data, analytics, gaming,

um

this is really useful and I would use it

in VMs, but I can also use it in

containers where I need some kind of

durable state.

Uh the user delegation shared access

signature for tables, files, and queues

is now GA. So that was already available

for blob,

but now it's been pulled to the other

storage services. And the whole big deal

here is user delegation SAS is more

secure than the regular account or

service SAS because both of those are

tied to the master storage account key,

whereas this is tied to an Entra ID

identity. So whereas the storage account

key has all powerful, it can do

anything, this would be a subset of the

permissions of the identity creating it.

Can never be more.

And it's only valid for up to 7 days

maximum. So there's a lot more control.

It's a a better secured method.

Um Azure Data Box can now be imported

into Azure Files uh provision V2 storage

accounts. So the whole goal, remember

Azure Data Box is hey, I need to move a

massive amount of data.

Very often it's how I'm migrating from

on-prem to the cloud. I don't want to do

it over the network, but it can work

from the cloud back to on-prem as well.

And so I'm doing a big data migration.

Well, now as part of that migration, I

can use a provision V2 account. The big

thing about provision V2 was like the

dynamic

provision disk V2, the capacity, the

IOPS, and the throughput can all be

separately set.

And so now, hey, I can use those

more fidelity type storage accounts as

part of Data Box.

Uh Azure NetApp Files cool access has

some enhancements in preview. So the

whole goal of the cool access is the

less used data can be moved from the

Azure NetApp Files storage, which is

considered the hot tier, to regular

Azure storage, which is considered the

cool tier, to help drive cost savings.

So with the premium and the ultra

service levels of Azure NetApp Files,

they've enhanced the quality of service

algorithms that drive the allocation of

throughput to really try and minimize

any performance impact you see as a

result of that cooling.

On the database side, so Cosmos DB for

PostgreSQL is being retired

uh end of March 2029, so 3 years' time.

The writing was on the wall for this.

It's been replaced by PostgreSQL Elastic

Cluster. All right. PostgreSQL Elastic

Cluster is

built on the same Citus extensions

that we see for that distributed

sharding up of PostgreSQL databases,

but the Elastic Cluster has the built-in

HA, backups, DR, all the future

engineering investments. So just make

sure you migrate. I mean, honestly, you

want to migrate as soon as possible, but

definitely you need to migrate um

before then. And there is migration

tooling available for you to achieve

this.

Uh Event Grid has a number of updates.

So remember the whole point of Event

Grid is I can build event-driven

solutions at massive scale.

And I as the app that wants to trigger

and do something off of the event don't

have to hammer poll. I I don't have to

constantly ask the source, do you have

something? Do you have something? Do you

have something?

Event Grid takes care of that and then

calls that often serverless

handler that's going to do something

with the event data.

So what I've done is a whole bunch of

MQTT enhancements. So there's things

like in-order message delivery within a

client session.

It's now has the ability for one

connection attempt per second per client

connection limiting. So I'm going to

throttle the number of requests. It's up

to 15 MQTT topic segments and there's

also cross-tenant delivery or when

they're available in GA. In preview,

there's MQTT OAuth 2.0 for the

authentication. There's custom webhook

authentication and there's static client

ID identifiers. Now it can also now use

managed identities for webhooks. There's

a cross-tenant webhook delivery and it

also supports network security

perimeters now. Remember network

security perimeters are about, hey, I

have a bunch of different Azure PaaS

services. If I put them in the same

network security perimeter, they're

allowed to all talk to each other.

And then I can also control

communication

for inbound and outbound to them as a

group of services. So it's a really nice

set of rule controls I can do there.

So Copilot Co-work is now available in

front here. I did a whole separate video

on this. It is crazy, crazy good. And

again, I don't often push to go and

watch one of my videos, but go and watch

the Co-work video. Even if you skip me

waffling on, go to like minute eight and

it's a demo.

It runs in the cloud, so it has no

use of your local machine. It's not

using local resources. It doesn't have

full reign to your local machine.

I can interact with it while it's doing

its work. It's grounded in Work IQ and

Outlook and Teams, all these fantastic

data sources, and I just tell it the

outcomes I want.

And it goes and works out

all of the plan and then just does it.

It It is a game-changer. I've done some

really cool things already just in this

week.

But uh yeah, go and go and watch the

demo. It's crazy good.

Uh Azure Azure Speech Neural HD 2.5. So

this is all about giving more choices

um in the regions you use it, the

quality, the performance, the

expressiveness.

Where I want a really low latency, think

real-time type interaction.

There's a whole number of different

speaking style updates for English

content.

Uh I I know it's all things like

struggling and skeptical styles. I can

do things like sighing and yawning,

which again will be important if I'm

trying to do a voice about my content.

But uh just a lot of work about trying

to have those um synthetic voices.

And then I'm only mentioning this cuz I

thought the core was really named. There

There's constantly new models being

added to foundry. That's one of the

whole points around I think the

Microsoft strategy in general is model

choice. There's no such thing as the

best model. There might be the best

model this week for this type of

requirement,

but

nearly every scenario is using multiple

models and they're going to evolve over

time. But I love the name of this thing.

This is a new Nvidia model, NeMo Triton

3 super 120B

uh A12B.

So it's a mixture of experts. So the

whole point of a mixture of experts is

there's 120 billion parameters,

but

it only activates 10% of them, so 12

billion for any specific inference. So

it's actually fairly compact in its

resource utilization, but based on what

that inference request is, it can choose

which expert it has that is the most

applicable to what it's been asked to

do. It has a 1 million token context and

it's really geared towards token

text generation. Obviously I could then

pass that to a text-to-speech model. I

could have a speech-to-text model in

front of it, but its specific goal is

around that um text generation model. Um

but super, super powerful.

And that is it. As always, I hope this

is useful.

Uh amazing, amazing, amazing for all the

updates this week. And till next video,

take care.