Sponsored By: MinIO | How a Fortune 500 Manufacturer Runs Databricks AI on 250TB of On-Premises Data

Summary

A Fortune 500 semiconductor manufacturer keeps hundreds of terabytes of production data on-premises for regulatory compliance, latency sensitivity, and cost reasons, but needed Databricks AI capabilities for yield improvement and quality assurance modeling.
MinIO AIStor with native Delta Sharing embedded at the storage layer allows Databricks to query live Delta and Iceberg tables on-premises through Unity Catalog with zero data movement—no copies, no ETL pipelines, and no additional infrastructure.
A live demo in this video shows the solution in action, including a performance benchmark where querying on-premises data via WAN completed in 4 seconds compared to 35 seconds for the equivalent data replication approach.

Chapters

FAQs

Why do some enterprises need to keep data on-premises even when using cloud AI?

Enterprises like the semiconductor manufacturer in this video keep data on-premises due to regulatory compliance requirements, latency sensitivity for manufacturing workloads, and the economics of managing very large datasets at scale. Replicating hundreds of terabytes to the cloud creates data drift, pipeline complexity, and significant cost—making zero-copy access a more practical architecture.

What is MinIO AIStor and how does it enable zero-copy access for Databricks?

MinIO AIStor is an on-premises object storage solution with native Delta Sharing embedded at the storage layer. This video explains that it was recognized as Databricks' storage partner of the year for this capability, which allows Databricks to query live Delta and Iceberg tables stored on-premises through Unity Catalog without any data replication or movement.

How does Delta Sharing make on-premises data accessible to Databricks?

Delta Sharing is an open protocol that allows sharing of live Delta and Iceberg tables across platforms and locations without copying data. In this architecture, MinIO embeds Delta Sharing at the storage layer so that Databricks workspaces can query on-premises tables through Unity Catalog as if they were cloud-native, with queries reflecting on-premises changes in real time.

What performance was demonstrated in the live demo?

The demo showed that querying on-premises Delta tables via WAN through the MinIO AIStor and Delta Sharing integration completed in approximately 4 seconds, compared to approximately 35 seconds for a traditional replication approach. This video presents this benchmark as evidence that zero-copy access is not just architecturally simpler but also faster for query workloads.

Full transcript

[00:09] Welcome everyone and very quick show of hands. How many of you have attending data and AI for the first first time? Very cool. How many of you have on-prem data today and using data bricks? Awesome. Right. Good.
[00:23] So this is the legal slide. Uh I will tell you though that uh it is not a forward-looking session here. This is all GA technology and uh we're happy to share really great
[00:35] content. Make sure to stay till the very end. My co-presenter Dwight here has a awesome demo to show you guys. So um yes and also make sure to complete the
[00:47] survey. It's very important. The NI Summit takes uh a lot of time to go through those surveys. And if you want Dwayne and I to showcase again next year, uh, make sure you put some good positive feedback on this. And again, my
[01:01] name is Dennis Dau. I'm a team member of the partner engineering team at Data Bricks. And I'm joined here, uh, with my co-presenter, Dwight. >> Hi, I'm Dwight Evers. I'm a senior field architect uh, with Mio. I'm out of
[01:13] Minneapolis, Minnesota. So just quickly here um I want to mention that um so we're really happy to recognize Men.io as our uh partner award
[01:26] of the year for a very new category that we launched just about seven months ago. I mean literally this was inspired by an architectural vision. We sat down with Min.io and they turn it around and
[01:38] productize it and now it's the gold standard for all the other storage providers that you also saw in the keynote announcements yesterday. So there's many more storage provider leading that charge, but Minio is
[01:51] basically the gold standard. So just uh really happy that uh we've recognized them as first ever. So now they're first in line and they just got to keep up to stay number one uh next year and win that award again. Right.
[02:05] So let's quickly define the problem and acknowledge that many enterprises today have a large onrem data estate right it's significant and the migrate everything is kind of an era that's gone
[02:19] away and everybody's recognizing that it's a hybrid forever solution instead right that's the reality for all of you that have on premises uh data now there's a couple of things um that
[02:33] are important because most of you are most likely replicating the data today from your on-prem creating a lot of pipelines and um you know probably addressing uh latency that's
[02:46] unacceptable. there's a lot of data drift, you know, it's very complex. And second of all is that um you probably keep this data on prem for good reasons,
[02:58] right? Regulatory reasons, some very strict laws most likely. So that data has got to stay on prem. And thirdly, the economics of moving that data from your on-rem platform, your appliance to
[03:12] the cloud is costly and it's prohibitive and it's challenging and very complex to say the least. Right? So it's not really a data issue. It's more of a compute issue, right? So you get this data on
[03:25] prem and you have a cloud presence as well where you do all your data anal analytics. So when teams are trying to solve this they're good they're going the good you
[03:38] know good oldfashioned ETL copy or they do a dual uh dual write method across environment which again is that complexity we've been talking about. So both approaches they do work um although
[03:54] uh they have compounding tradeoffs. Your cost will grow as your data estate will grow. Obviously, your complexity will also increase over time. I mean,
[04:06] it's not when you start opening this pipeline and you start pushing this data from onrem there, you're going to have more stakeholders demanding more data and it gets really really complicated really fast, right? And the latency is
[04:19] going to get worse over time as well. I mean more data more pipeline there's there's so so much capacity that you can do and with the uh adoption of AI and increasing of adoption AI adoption the
[04:33] penalties that you're incurring today are just growing by the day right it's a simple reality
[04:45] so consequently the hybrid approach is evolving and is no longer a compromise right it's just a fact it's almost a default ult reality in a sense that we know that most of the data today exists
[04:57] on your on an on-remise state. I mean it's slowly you're joining the cloud. You're moving data to the cloud. There's already data that you're consuming and generating on the cloud. But we all know that all this data out there is
[05:11] primarily on on an on-remise state. All right, that's the fact in reality. So you don't want any more pipelines, right? But you want the elasticity of the cloud to do all this AI compute
[05:25] that's necessary. And how do you get this onrem data to the cloud? You don't want more infrastructure. So you want a solution that someone has already embedded built from the ground up where you can leverage in an easy user
[05:40] interface and just point and click and make that data available to the cloud. And you definitely don't want more copies of your data. I mean there's a lot of risk. There's a lot of security and not a you know not not to say the
[05:53] the the actual amount of resource that you invest a lot of these engineers that are writing code just to push you know doing ETL jobs you can probably repurpose that engineering effort to do something a lot more palpable for your
[06:05] industry right a lot more insights so you want to secure a govern access to your data where it already lives right you don't want to move it you want to
[06:17] get just in time just enough access to where the data is actually living and being accumulated.
[06:30] So everybody today will agree that there's many industries out there, many use cases. This is just a sample set and everybody is in in an AI era, right? So every enterprise will benefit from AI but the
[06:44] data a lot of that data is on prem, right? So pick an industry, pick a use case. I mean from that sample list and then this could grow many different ways. But the need for speed and scale
[06:58] as well as governance, right, has never been higher. So folks are looking for a simple way with a native open-source approach to get to that data. You don't
[07:11] need additional pieces of software and infrastructures and servers and whatnot to push this. What if we'd give you an open mechanism, an open and secure protocol, right, to provide access. You
[07:25] also want this live and easy query uh with a zero copy capability. You don't want to move that data. You want to access it where it lives. Right? Now the next question to me really becomes a
[07:38] little bit more complicated than this is the governance right so you've got this data asset with some rules and policies sitting on your onrem scenario you want to give it access to your cloud and you want to start consuming this data
[07:51] so how do you govern this you've heard this week if some of you have attended some of the keynotes there was a lot of announcement today on this platform most of you here are using data
[08:04] bricks and you've heard all about Genie. You've heard about the AI gateway uh on UC and there's many many innovations that's being done on the platform and really it's not just accessing the data
[08:18] and running query. It's about what's the opportunity of leveraging the data that's on prem with this entire stack. And this is just a sample of that stack. There's been some lot nicer pictures
[08:31] presented at the keynote. Obviously, believe it or not, even uh we're even, you know, not preview to some of these nice nice uh images uh as we build our own presentation. But I mean, it gives you a flavor of what's available
[08:44] assuming that now you can get that data that's on prem and you can start consuming this on the platform and from an AI use case standpoint, right?
[08:56] So it's a lot more than just bringing data from your on-prem to go generate the next dashboard. It's actually taking that data. What if you could actually start using Genie to access that data on that's on
[09:09] prem. So you have that connected tissue and now Genie is now able to operate on your data directly without moving that data. That's a pretty powerful statement, right? But at the end of the
[09:21] day, it's the governance of all this data and it's let's zero in on the unity catalog. So today have the data have the battle access giving you access is one thing.
[09:34] The next thing is how do I govern this data and that data that's available through the method that we're going to show here you today that data becomes just another table asset under UC. So,
[09:47] Unity catalog will maintain and manage that content that sits on prem into your Unity catalog landscape. So, it's not just about the data that
[09:59] you get access to is how you get this data and now blend that data into your modeling exercise, right? What's the next model that you're going to build? How do you going to supplement this content? And you're going to all do this
[10:11] while you're governing that data through UC. So when that data presents itself in UC, it is now an asset that is manageable just like any other assets that you see here on the screen that we
[10:24] do today with UC. Right? So what it really unlocks it unlocks a hybrid access with audit auditable governance model. So
[10:37] this very very important like auditable environment. So when that that data makes it available in UC you can start governing you can audit that content as well just like if it'd be in the cloud right
[10:55] so the real question is can you leverage your data platform today while keeping this data on prem eliminating all these pipelines eliminating all that tech debt that you've built and just access it
[11:07] directly Well, the answer is is rather simple and you've probably heard of the keynote. It's open sharing, right? Open sharing is is a rebrand of delta sharing. You've
[11:20] probably heard about delta sharing if you're a data bricks customer. Uh as of this week, we call it open sharing. We have we support a lot more than just data content. I mean, if you're a data bricks to data bricks open sharing,
[11:32] there's all these different you know components that you can share. What we've done is it's a battle tested protocol that's been used by thousands of customers throughout many industries. So this is not new tech and we were
[11:45] taking the tech but we've repositioned this and minio is actually embedded this directly in their platform. This is a very powerful statement, right? So you're not you're not dealing with
[11:58] something brand new that we're trying to you know you know test and and and scale that protocol today is used today by thousands of customers whether you're
[12:11] you know using within intra sharing capabilities within the cloud the regions and whatnot it's a battle tested protocol
[12:25] all right so what does it unlock I I mean the obvious one is um it unlocks your use cases whether you want to enrich or actually from a zero from an opening standard st standpoint um
[12:39] apologize there. So um so what you want to do is that the zero cloud open sharing this is as well has been talked about uh today at the keynote as well but is just give you a glance of what
[12:51] open sharing is right. So it is a zero copy protocol. So again it's super important to understand that we're not replicating the data you're just getting access directly at the source.
[13:08] All right what does it unlock? So many of you are pushing these pipelines from your on-prem to the cloud to enrich you know existing use cases or create new use cases. It also has the ability to
[13:20] unlock a marketplace data collaboration use case. What if I'd say that today we have many customers that are you know sharing data content and they're collaborating with data sets on the
[13:32] marketplace and now you can tap into your on-prem scenario scenario to basically generate that data content that you can share with other customer. If you're a healthcare customer, if you're financial services customer and
[13:44] you have a lot of data that you want to share collaboratively through the marketplace now you can use this solution to do just that. There's uh also a data clean room content where this is an ability if
[13:56] you're not familiar with clean rooms where you can actually provide a secure space and that's the foundation of clean room is using open sharing. However, now you can go and tap into your on-remise estate as well to bring that data
[14:09] content into this private room and obviously the internal sharing uh which is obviously essentially um it's a new emerging use case but also uh if you
[14:22] have a lot of data data mesh strategy folks are talking about mesh strategies sharing content leveraging you know your on-prem data you know fits into that.
[14:34] So in a nutshell I mean if we just summarize and conceptually if you look at this the left hand side is minio that data resides on an appliance in your on-prem data center and through the open
[14:46] sharing protocol now you can connect that to your directly to your data bricks you don't have to do anything right so it's out of the box essentially if you if you have AI store this
[14:59] connectivity exists and then uh Dwight here is going to show you a nice a really nice demo and how the UI experiment experiment is. But it's all out of the box. You don't have to set up some servers. You don't have to
[15:11] establish the protocol yourself. It's all done by Min.io and they've embedded the protocol directly in their stack. Right?
[15:25] So finally I mean essentially what this does is it not only allowing you to get access but it's all done through open sharing and again a simple
[15:40] protocol that's battle tested by thousands of customers numerous industries we have customers like Elsa who's using terabytes terabytes of data through this
[15:53] mechanisms. We have a a lot of financial and uh healthc care customers using this platform either through the marketplace sharing either through a clean room capability and many many many customers
[16:07] using open sharing to share across region across even platforms and open sharing is is agnostic. Um just to let you know I mean even Snowflake just adopted this client solution uh
[16:20] recently. They just decided that this was the protocol that was you know in the uh that was battle tested through the industry. So we're seeing even additional competitors and other platforms adopting and basically uh
[16:34] deploying open sharing as the protocol as well on their platform right but we all want you to use it on data bricks uh along with minio right and obviously what we'll show you today is what has
[16:47] been done thus far and towards the very end we'll also give you a nice picture and and the the northstar of where we're going to bring this to bear in terms of functionality and features.
[17:01] All right, so on that topic, I'm gonna hand it to you, Dwight. Sounds good. Thanks, Dennis.
[17:14] to cover just a couple topics here. So, what we started out with about a year ago was a customer came to both Data Bricks and Minio and said, "Can you two please work together? We have on-prem data. we can't move it but we need to
[17:27] share it. So the whole idea was for the two companies to work together from the engineering perspective as well as from the customer's perspective. So what I'm going to be showing to you
[17:51] and then storing that within AI store. How many of you have heard of Mano before I stood up on stage here today? So, ironically enough, Mino has been around for about as long as the S3 protocol itself from Amazon. We're the
[18:04] premier object storage platform for a lot of our for a lot of companies out there. We have everything from uh Raspberry Pies all the way up to thousand node uh deployments with
[18:16] several actabytes of storage under a single name space. Many of you probably take advantage of the of some of our customers today that are leveraging it. But what they wanted to be able to do was leverage an object
[18:28] storage that would actually scale. So today, when they originally started out as part of our proof of concept, it was about 250 terabytes. They're now several pabytes. So it scales very nicely.
[18:41] From the perspective of how they get it into they're using our REST interface for the iceberg catalog. We have one that's built in. AI store uses a distributed model for both
[18:54] storage as well as the catalog, but also for how we h provide services up for open sharing. It's completely redundant and and hardened underneath covers.
[19:06] From that point, then you go ahead and create a to uh uh a token exchange. A bearer token in this case is what I'll demonstrate today, but it also works with OOTH to be able to share that up to data bricks and then bring it into the
[19:18] Unity catalog itself. So I can write run queries against it directly. The only way this is possible is if you're 100% S3 compli compliant and that is object storage is what they have uh open open
[19:31] sharing is built on top of. So I'll go through what it looks like with both Delta and iceberg tables. we don't discriminate and then also how we create the sh uh the secure credentials
[19:46] how that looks like when you bring it into the data bricks workspace and then run some fun queries against it so I'm going to switch over to my laptop everyone can see that hopefully so this
[19:58] is the basic interface of what it looks like when you go into Menio AI store we're on prem for or wherever you want us for your private cloud for object storage we support both the idea of you know bucket and path when it comes to
[20:10] Delta tables, but also the rest interface for iceberg. One of the first things I'll do here is I'll I'll look at some of my data that I already have out here. I have a warehouse called test ice and in there's
[20:22] some t some wonderful TPCH data. I'll take a look at the customer and I can preview it. Now, from a demonstration perspective, I have full access to everything that's in the table. In a production environment, you would be
[20:34] able to scale this down to just those who who have appropriate identity access. Now the other thing that we have is the ability to work with as I described earlier uh delta tables. In this case my
[20:47] TPCH data for my customer would be along the lines of where the delta logs are stored and and the accompanying parquet files.
[20:59] But what's important is not that I can store the data. Of course we can do that but I need to share it. So we can go into uh an option here called table sharing. This lets me define how I want to share something.
[21:14] Now I'm going to generate a share and the first thing I'm going to call it is data bricks AI days summit and this will be uh demo.
[21:32] Now, the first thing I'm going to do is I'm going to grab a table from my iceberg catalog from this the and I can choose iceberg or delta. We'll start out with iceberg and my existing warehouse. I'll grab
[21:44] from my test ice and I'll grab it from the TPCH schema and I'll do customer to start with.
[22:01] Now I can go ahead and change the names. What's nice here is that you don't have to have exactly onetoone mapping. Oftentimes when you're doing kind of like a marketplace or a clean room type of approach, you'll want to anonymize certain things so that people only understand that from uh from the schema
[22:14] or the definitions that you define. So I'll add that right now. I'm going to add another table though. In this case, it's going to be a a delta league table.
[22:26] I'll choose the bucket that I have that in which is test delta and the path will be TPCH
[22:45] Now I again I can choose whatever type of name I want for I'm going to keep it under uh TPCH just to keep it easy to remember. And the name is going to be orders. Now I can go ahead and create the share and I can preview it. Here we provide you some example Python to start working
[22:59] with as well. A lot of a lot of times hey how do I begin to work with it? Dicks has a very uh phenomenal set of of tools and technologies to make this easy from a from a programmatical standpoint.
[23:11] But from an administrative standpoint you want to be able to just hand this to the to the engineer and let them go to town. I can preview at least the definition of the tables. And then when it comes to
[23:23] the token, I'm going to create one here that's easy enough to remember. I can define for how long it needs to be here. So I'll call it Dwight demo. I'll give it seven days. So those of you
[23:35] that are recording this going right ahead, you can steal it for seven days and enjoy. We'll go ahead and create the token. I can copy the bearer token. I can slack it to somebody if I need to or I can
[23:48] just download it as as an encrypted file and then work with it from that standpoint. Of course, if you're using something like OOTH, then you'll have additional authentication factors that are required as part of that. So, just go ahead and download that.
[24:02] It came as a a share file. I'm done with that part for right now. Let's jump over to data bricks. So, everyone's familiar here with at least the Unity catalog interface.
[24:15] my now I already have my uh serverless environment working up and running. I go to open sharing and I can install my share from here. And I'll call it
[24:28] uh
[24:41] And I'll import it. And I can mount it to a catalog. Again, pay close attention to this. Right now, I actually have defined there. That's our public endpoint for working with
[24:54] uh for customers to be able to do evaluations and and testing with. So in this case here, it's talking to port 444. I'm going to mount this to a catalog and just call it let's use the same name as everything else
[25:07] summit. Go ahead and create that. As soon as I get it, I can jump into the catalog itself. It's showing me that I have my TPCH schema out there with the
[25:19] customer and orders. This is being served up from our data center to data bricks which is in being hosted out of the uh I believe the Oregon data
[25:31] center. So when I go to look at the customer data I can see the schema is getting cached. The only thing that's cached or stored in data bricks is that which was
[25:44] which is part of the unity catalog itself. So when it comes to security, if I do a truncate table, it's immediately reflected in terms of the data that's being requested. We don't cache that exclusively up in in
[25:57] uh where it doesn't need to be. Same thing with orders. Now this is the delta table. It looks exactly the same. The idea was that the the open sharing protocol is flexible enough that
[26:09] everything will will adhere to the the the delta table standard when it comes to the uh types and and uh the data types and naming. So what we did under the covers is we have an instant
[26:21] instantaneous translation layer that does this on the fly. So if I make any changes to columns and so forth within my table, that's reflected the next time I do a refresh on the screen or I go to
[26:33] query it the next time from the editor and I can sample the data.
[26:45] Just jumping back here quick to the console. If we go to the monitoring and my tables, I can see that the objects are being pulled up. It's leveraging the predicates that are within the objects
[26:57] themselves. So just as we distribute the query response across all the nodes that are in the cluster, you know, strength through strength through numbers, we do this similarly when it comes to storing
[27:09] of the information to make it resilient. And here you can see that the response time on for example the the list the load table
[27:21] is 2.7 milliseconds. So we're we're less than five millisecond of response time for time to first bite on all these queries. Well, that's wonderful. So let's go ahead and do a query against it.
[27:33] So in this case here, I'm going to go ahead and select my cat uh my catalog and my schema and I'm going to run
[27:46] a join across both of it. So I'm joining a delta table and iceberg table that are sitting on prem and it's loading it and running it directly within data bricks. No data has
[27:59] to be replicated. There's never a time where I have to ship data across. I don't know if you have dogs or kids but it's sharing not giving.
[28:15] Now, something else I'm going to do here real quick as well is I'm going to jump over to one different T. We could actually look at the the performance as well. Let's jump into this. You can see the amount of time that's that's spent on each process within the uh as we
[28:28] profile the query. We see where the amount of time is actually spent. Now, we're not going to lie, it takes more. There's more latency going over across a WAN than there is a LAN. Everyone knows that. We're not
[28:40] hiding that. I have some performance numbers I'll show more in more depth about that. But the fact that we're able to do handle the predicate push down directly into open table formats like iceberg and delta. You're not having to
[28:52] retrieve the whole file back. Just those blocks that are required to service the query itself. And it's compressed all the way across because data bricks does a fantastic job of handling the the
[29:05] consumer side within the Unity catalog. So, I have another table that's out here to take a look at it
[29:22] and it's my bake house. This came from onrem. I'm going to quickly add this to a share that I already have out there. So, go to tables. I'll go to Dwight01 my bake house and I have
[29:37] customers and transactions. This is where it was.
[29:52] I'll just add the transactions to to my other one. That's probably easier. Go to my tables sharing.
[30:05] I have one table there. Dwight. I'm going to add my for my bake house my transactions table.
[30:19] Transactions. And I'm going to go ahead and add that. So I'm adding a table on the fly. So with this,
[30:35] all I have to do is refresh it and it should pick up the transactions here momentarily. And there's the transactions table. Let's jump over to the SQL editor.
[30:49] Let's do a select count from uh transactions. So, it's got 2200 transactions in that table. Let me just finish out what I had
[31:02] started here earlier, which is I'm going to just add some more data to it. So, the data is live. We're not caching anything.
[31:15] And I just added some additional tables or some additional rows.
[31:27] So now when I go and look at it, if I run that same query again from 2200, it jumps up to 3,300. So any changes you make on prem are immediately reflected with what the share is coming from data
[31:41] bricks to look at. We're not there's no there's nothing that we're doing in between. And we're making sure that it's available at both ends of the spectrum.
[32:03] So you got to see the data stays on prem. You can see the how long it takes to get to the information itself. We're leveraging directly within the Unity catalog because it's by far the easiest way to work with it. All the policies and requirements that you have there as part of governance or security still are
[32:17] still in play and there's virtually no setup required. I hand the share file to my data engineer. They get they they're able to authenticate and they can start working with the data.
[32:32] So how fast does this actually work? Well, there you saw a small data set as part of a proof of concept that we had to do with an with a com with our customer. They wanted they wanted to test it out. They said the first thing is let's look at what it takes for local sharing for example within the the
[32:45] public cloud itself versus working with data that might be on prem. So with a with a 2 Gbit link over uh the WAN about 15 millisecond latency is what they measured it's about from 1.7 seconds to
[33:00] to about 4 seconds. So about double the time most of that of course is WAN latency. Compare this to what they were doing just prior to this as part of their testing procedures is they're having to
[33:13] do replication. Well, the time it took them to replicate just that one data set, it took 35 seconds just to move the data and then another 10 seconds to ETL it. And this is all in data bricks. It's it's not that data bicks is doing
[33:26] anything wrong. It's just that they're using it le uh less efficient than what they would like to. Of course, the other option that they had in place was something like Golden Gate, which required, you know, a CDC
[33:38] pipeline that was very complex and very expensive. We're data bricks and min are trying to help our customers save on those extraneous costs that aren't providing value. Moving data is not valuable. Quering data is what they're
[33:51] trying to get at. So, in the time it took to do either one of those methods, they could run 10 times as many queries, get 10 times as much response, but they're not paying any API costs as part of it.
[34:09] So it really is about about how we want to help our customers and the our organizations build this modern data lakehouse that stretches from the cloud to onrem. Data bricks is without a doubt the premier data intelligence platform.
[34:22] When working with men, we are without we we've demonstrated time and again that we are a high performance object storage that has native object uh native open sharing built in. It's not a bolt-on.
[34:34] It's not a separate process. It's integrated in with everything that we do from identity management security as well as the resiliency of the platform itself. And of course fully leveraging what the capabilities of iceberg and delta tables
[34:48] because open tables afford things such as time travel and asset transactions and they understand the protocol themselves. But having that understanding of how to
[35:00] how to deploy a private hybrid type of approach is what our customers are asking from us. So what does it take to do it today?
[35:12] If you want to be able to work with uh uh unstructured data, we have uh data bricks in the room and we'd like to hear more from you on what you'd like to see. So today it's tabular information that we're focused on.
[35:24] Unstructured is one of the things we're very much keeping an eye on as well and listening to our customers on this whether it's CSV or JSON or image and video files using male. I'm here find me. We're at
[35:38] booth 527 so you can have an opportunity to to discuss what it takes to actually deploy. It can be bare metal, it can be containers, Kubernetes, private cloud. There's even hosting providers that we work with on this as well. What's nice
[35:51] is there's no API costs.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.