O’Reilly Software Architecture Conference 2019 – Berlin, DE

About this course

so in a recent study by IDC they saved that by the year twenty five or twenty twenty five which is only six years away less than six years away we're going to

expect that the storage industry will ship an additional forty to zettabytes of capacity and if you're not familiar with how big is zettabyte is that's a trillion gigabytes so we're going to be

adding 42 trillion gigabytes in capacity we're gonna see IOT devices generating things at a rate of about ninety zettabytes we're expecting about half of the data that's going to be produced

will be stored on the cloud for the reasons that we all know and like and that roughly thirty percent of the data is gonna have to be handled in real time right because we're not going to be

storing absolutely everything we're going to capture things that happen interact with them reason over them make choices from it but that's going to translate into a shift from about thirty

three zettabytes in existence now to about a hundred and seventy-five zettabytes by 2025 that's an accommodative about sixty one percent which is insane and it's going to come

in these three basic tiers right there's the backend the cloud-based systems where we push a lot of the data up spend some time training in the cloud and chunking on

the data that's there there's also the front tier the mobile tier the internet-of-things tier where we have the data being generated we've got sensors that are producing data we've

got cameras and microphones on our on our phones that are being used as sources to treat people for medical analysis and and whatnot to determine we know if there's been evidence of some

kind of neurological damage in somebody's you know facial expressions and whatnot and there's sensitivities around that like we can't just take that data off of a phone and shove it up into

the cloud without thinking about regulation and those sorts of things so we're actually imagining this multi tier heterogeneous highly paralyzed architecture that we're going to have to

deploy into and we get back to the kinds of problems that we were facing at Parab on with respect to like the compute and data ratios it doesn't make sense to be pushing

zettabytes of data around from the front to the back but we also don't necessarily want to push our big models all the way into small small profile devices like phones and whatnot so we

have to start to think about the models how big they are how can we reduce the size and that's where things like quantization of the models and tooling around optimizations to target the

actual hardware that's going to run on are going to become increasingly interesting and useful and so this world of hey let's do some training let's build a model and let's roll it into

production it's kind of an older view of what it means to take machine learning into production because we have to deal with all of these situations and all these scenarios