Sponsored by: Salesforce | Tableau & Databricks: Demystifying the Open Semantic Interchange (OSI)

Summary

The Open Semantic Interchange (OSI) is an open metadata standard co-developed by Salesforce and Databricks that provides a hub-and-spoke model for keeping KPI definitions consistent across BI tools, AI agents, and data platforms without duplicating metric logic in each system.
OSI's core model includes datasets, metrics, AI context, and custom extensions with multi-dialect SQL support, allowing different tools to delegate semantic query logic back to the native authoritative system rather than reimplementing it independently.
Tableau's roadmap includes delegated Unity Catalog metrics to Tableau by end of year, and the OSI community hub on GitHub allows contributors to add converters and extend the standard for additional tools in the growing analytics ecosystem.

Chapters

FAQs

What problem does the Open Semantic Interchange (OSI) solve?

OSI solves the problem of KPI definitions being duplicated and inconsistently implemented across different BI tools, data platforms, and AI agents. When every tool reimplements its own version of a metric like revenue or active users, different systems produce different numbers, eroding organizational trust in data-driven decisions.

How does OSI work as a hub-and-spoke model?

In the hub-and-spoke model, a central semantic definition for each metric is stored once and converters translate it into the native format required by each connected tool. Each tool delegates query logic back to the authoritative semantic definition rather than maintaining its own copy, keeping metrics consistent across the entire analytics ecosystem.

What does delegated semantics in Tableau mean for Databricks users?

Delegated semantics means Tableau respects the native query logic defined in the Databricks Data and AI platform rather than reimplementing metric calculations locally in Tableau. When Tableau connects to Databricks through OSI, it passes metric computations back to the authoritative Databricks definition to execute, ensuring results match what Databricks-native tools would produce.

What is Tableau's OSI roadmap for Databricks integration?

According to this video, Salesforce plans to deliver delegated Unity Catalog metrics to Tableau by end of year as part of its expanded strategic partnership with Databricks. The OSI community hub is available on GitHub for teams that want to contribute converters or extend the standard for additional tools and semantic patterns beyond the initial supported set.

Full transcript

[00:07] Thank you for coming. I know it's over lunch. Um, my name is Nicholas Breez. I'm Nick. Um, I'm a product manager at Salesforce. I've been with the Tableau part of the business for 11 years. This week was my anniversary, which is
[00:20] incredible to me. Um, so today I'm going to talk about the open semantic interchange, otherwise known as OSI and give you like a shallow overview of what we're doing there and what it is. Has anybody aware of what that is or what
[00:34] we're doing? Showed up because you're curious about it. Is it just marketing nonsense or is it actually happening? Great. All right. So, before we go into that, yes, the famous forward-looking statement. I am going to talk a little bit about some future-looking stuff. So,
[00:47] only make purchasing decisions based on what's available today. And of course, please complete the survey. Um, give the content a thumbs up. Hopefully, the speaker will see. I'd
[00:59] like to get somebody else to do this next year. All right. So, just a reminder of what we're doing with Data Bricks. Salesforce and Data Bricks have a strategic partnership. We have, of course, our zero copy integration with
[01:11] data 360. So that allows you to bring all your Salesforce data natively into data bricks without using ETL or copying it. We literally just turn it on for you. And then you can bring your own model into into data 360 and customer
[01:25] 360. And then Tableau has had an integration with data bricks since 2019 when they released support for DBSQL. You could use Spark before, but that's really a deep integration where you can connect live or extract your data into
[01:38] our in-memory format. We did an announcement yesterday. We really expanded that partnership with data bricks. I got a QR code if you're just curious what the PR looked like. So we're bringing even more capabilities to
[01:51] data 360 and data bricks. And then just a reminder I said on the Tableau side. So we have three main products. We have of course Tableau Server. Many of our customers especially
[02:03] in regulated industries need to self-deploy Tableau and so we have Tableau Server available for them. You can run it on public cloud or private cloud. We have our SAS offering T table Tableau cloud. When I first joined, I
[02:15] was a PM on that team uh back when it launched. Um and then of course we have our latest version of Tableau which is Tableau next which is Tableau native for CRM. So the difference between the two is that is explicitly built for
[02:29] Salesforce and brings a lot of agent capabilities to CRM and Slack. All right, let's talk about the open semantic interchange. So a little bit of history. Um I was pretty heavily involved in this. Um it initially
[02:43] started as the open semantic initiative. Uh I was pretty hesitant to talk about standards or trying to push some kind of format. What I really wanted was vendors in the industry to be able to exchange
[02:56] semantic information. And so we started from there. Um and eventually it evolved into the semantic interchange when we kind of agreed like hey we can propose an open source format and maybe we can align around that to
[03:09] use as the mechanism to exchange that information. So we're going to take a little dive into that. Um you never know what's going to come out of cocktail conversations at a partner summit. Okay. So we formally announced this in
[03:23] September. At the time we launched with I think about a half dozen vendors. Since then, if you go to the OSI main page, um you'll see that there's over 50 vendors, which is awesome. Um and it is
[03:35] an open source effort. So, we are collaborating to enable you to bring your model definitions across all your data and analytics products within your own stack. So, yes, you will see QR codes all through the slides. If you
[03:48] want to get more information about the specific piece that I'm showing, I've updated them one for each section. So the the initial goal is really how do we create from a from your perspective
[04:01] like a unified semantic layer that goes across all the places where you define models and metrics basically avoiding lock in right if you think about what you do today and that's true for Tableau a year ago you recreate
[04:15] KPIs all over the place right they don't exist in a central location we're trying to do that you're familiar with UC metrics but there isn't a good way for us to do that interchange between the different platforms. As my boss's boss
[04:28] says, you should be able to define it once and then use it everywhere. And of course, to do that, they all need to understand the same language. So, going a little layer down, when I first started this work, um it was 2021
[04:42] and the fundamental tenants that apply for agents also apply for just your basic analytics, any kind of data work really. So first and foremost you got to have birectional metadata exchange. If I'm
[04:55] divi defining a semantic model wherever something consuming it has to be able to reflect what that is and not force you to copy it. Problems exist when you start copying things all over the place
[05:07] just like spreadsheets right and then this is super important too. You want the governance to propagate. So if I'm connecting from a dashboard going through the semantic layer, that user
[05:19] context should apply the same access rules all the way down into the native system where that model is defined. And equally important is respecting the
[05:31] native query logic that you have. If you define a calculated field in datab bricks, we should delegate we as in whoever it is that's consuming it should delegate that logic to data bricks so
[05:43] that we can guarantee consistency when you're returning that result. All right, so let's go now OSI works. It's actually quite simple. It's a hub and spoke model. So at the center of it
[05:57] is the OSI specification which is an open source model and each of the vendors or other solutions because you could build your own if you wanted to. We have customers that have built their own semantic layers plug in through this middle model
[06:10] and everybody exchanges their information that way and they work through what's known as a converter. So that's what allows the translation between one vendor going through the hub to the other vendor.
[06:23] Again, there's a I think there's five of them right now. There's a bunch that are being worked on. Again, you can see the full list when you go to that link in GitHub. So, that's the foundational piece for doing the interchange of information.
[06:35] And then there's the main component. So, at the heart of it is the semantic model. Obviously, that's that's what we called it. Um, and then there's some and it's all in YAML, so it's easy to understand. There's some additional key
[06:47] components here. There is what's known as the AI context. So you can embed within your semantic model additional language to help ground the LLM when it's using that semantic model data sets
[06:59] which are the fundamental models that you'll be using and I'll give a little bit more details about that um in the next slide and then custom extensions. So maybe you need to pass on some information because you're using data
[07:11] bricks or you've got mother duck and there's a separate thing that you need to go and pass along to that underlying system. Um, and so there's we also give you the ability for each vendor to specify those things. So within the core
[07:24] building blocks, you've got of course data sets, so your fields, your role level attributes. Um, not only do you have your dimensions, but you can also add additional category attributes like where, when, and who is using this
[07:37] model, the persona behind the model. You've got your classic relationships otherwise known as your joins where you can do single column joins or specify composite keys
[07:49] and then of course very important the metrics. So aggregated business logic where it's only defined once and when a third party access accesses through OSI
[08:01] it will be calculated in that source system. So that's aggregations cross data set metrics and then there's multi- dialect support depending on the dialect of SQL that you're using. Some systems support multiple dialects and of course AI synonyms.
[08:14] >> So here's a sample model. I had to rip out a lot of stuff because it wouldn't have fit on this slide, but you can see just the basic setup for it. And then of course the custom extensions here that we have for data bricks, which is just specifying the default schema and the
[08:27] default catalog for that particular object. Okay. Uh little bit of vendor speak here. So because I work for Tableau, we've been doing a bunch of work to support the notion of delegated
[08:39] semantics. So you can think of like in any enterprise, you know, ignoring the logos, you're going to see something that kind of looks like this, right? You're going to have your data platform, there's probably somebody doing data modeling in some set of tools. You will
[08:52] have some semantic layer, whether it's explicit or implicit, your analytics layer, and then where you consume those analytics in apps like Teams or Slack. So today we're focused on that on that
[09:05] middle part of course, but the goal is to bring not just the ability to consume those things, but ultimately the trust and the governance that you need so that when somebody's looking at total sales, they know that that's been defined in a single place and it's not something
[09:17] that's been replicated somewhere else. So you can think of it as a semantic stack. So within Tableau, there's of course you can go connect to relational tables. So your basic data model,
[09:30] sometimes it's a raw a raw table. Other times it's a business view or some other semantic view of the data. Um cubes, yes, many customers still use cubes. They will be around for a while,
[09:42] especially if it's older companies. If you're not familiar with what a cube is, it's essentially a collection of pre-aggrated data. So it sits a level above the raw information. Um and of
[09:54] course we support connecting to cubes but it is a level up and that will matter in a little bit. You can then define additional models on top of that. So let's say for example you wanted to do a prediction column within a data
[10:06] source. You could say okay pull this raw data run it through the data science and now we're building up our semantic stack within Tableau itself or any BI tool really you can do more aggregation at a higher level um which ultimately gets
[10:20] visualized in some kind of encoding a bar chart spider chart pick your favorite chart the missing component here of course is how do we bring in metrics because that is really typically the key component that matters when it
[10:33] comes to building trust in anything that a user is using whether it's a visualization or an conversation or an
[10:49] capabilities, you're going to have a mix of things coming from all over the place. You may not have a semantic model that's been designed defined for everything. So in this example here, you can see I've got models coming out of data bricks. I've got some marketing
[11:01] data. Maybe it's Google Analytics 4 pulling it out of Google and then some additional information that's been modeled in DBT within Tableau. These all work like your basic definitions that
[11:14] have been created within Tableau. Same thing for metrics. So you can take advantage of all the blending and the joining. Um you can refine those metrics further down the stack, but all the time you're starting from a place of trusted
[11:26] lineage and definition for that model. And then of course I mentioned Tableau next earlier which is the native to CRM Tableau. Um equally there they are supporting all the same functionality.
[11:38] If you want to see what that looks like in a little bit more detail I'm not doing any demos today. Salesforce booth we're showcasing some of the the Tableau next semantics. I'll also be there 3 to 6. So if you got more questions please come by. I'm happy to chat. Um and
[11:50] that's integrated into data 360 which is that customer data platform that we offer where I talked about the zero copy earlier. All right, let's take a look at the road map. So, first today I mentioned earlier
[12:02] we have an integration with data data bricks via our DBS SQL connector. Uh I don't know if you're aware but we have the ability to initiate a Tableau data exploration experience from within Unity
[12:14] catalog itself. So you pick a model in the upper right hand corner there's a little dropown you'll see PowerBI Tableau and you can open up in desktop or you can open up in Tableau cloud or in Tableau Server and in Tableau Server
[12:27] Tableau cloud it's an ephemeral experience so you can work with that data you're not persisting it anywhere yet until you decide that you want to do that and then lastly um if you're familiar
[12:39] with private link in AWS so it's this ability to connect your data without going over the public cloud so that's coming soon for data bricks uh for Tableau cloud. So on our road map I mentioned what we
[12:51] have today. Tableau next has a translator converter available within the OSI on the GitHub page. Today what Tableau connects to are just the semantic models. We will not respect
[13:04] the metric definition. So if you bring in a calculated field from a third party system, we will reagregate it. And that is just because of the way that we were designed originally when we just assumed everything was role level data. So in
[13:18] the fall we are bringing in the notion of that delegated metric being defined somewhere else. So you'll be able to bring it in kind of like a calculated field within your data model and then blend it with other things. Initial
[13:31] release is for Looker and DBT. Um and that will also be integrated with the relationship model that I showed you earlier. But early next year and we're actively working on this right now. I manage connectivity at Tableau. So I'm very
[13:43] familiar with the road map for that. Um both Snowflake and data bricks will also be supported. So things that you define in UC metrics you will now be able to bring in not have to recopy it into the native Tableau language that we use to
[13:56] do our calculated fields and we will support that delegated query back out to data bricks for that metric. We already support the semantic models in a way already. It's really that key component that we're missing which is the notion
[14:08] of that defined metric and then next year because we started our work in 2021 we don't support OSI today that is going to come later because we have work to do because we had already started down that path
[14:20] before OSI of course and so we will put in a formal support for it create translators for Tableau as it is today if you're familiar with extracts still some ideation around that what does it
[14:33] mean if you extract something that's been centrally defined somewhere. Should we even allow that? We'll see what customers have to say. Yeah, I agree with you nodding your head over there. You shouldn't do that. Maybe it'll be a flag. Um and then of course aggregate
[14:45] awareness. All right, so here's the good news. This is live as of April. OSI is officially live. Um there is a community hub that you can go partake in if you
[14:57] want to participate or if there is a vendor that you use that has not is not showing up on the list of vendors supporting feel free to push them to go commit to supporting OSI everything is
[15:09] open source thank you to our friends at DBT who drove this this came out of their metrics flow technology and it is we donated this to the Apache Foundation so there's a number of companies that are you can see it's all listed on the
[15:21] website all the contributors there's different parts s to it. Um, and then of course we extended the original idea with that AI support with the additional AI context I mentioned earlier. Again, use a QR code. You don't need to remember this. Just go take a look at
[15:33] the website and then please participate. You know, I I I'm going to make a joke about O data. I don't want this to be O data. I really want this to take hold in the industry. I think as customers and
[15:46] vendors, we really need this and we want to stop that whole problem that we have with breaking trust in data when people look at numbers and they don't trust that number. Um, but we can only do that if we get all the major vendors on
[15:58] board. We've had a good start. We've got the the important ones right now, but we have more work to do. Okay, as I mentioned, I will be at the booth. If you want to come talk to me, got more questions about this, I'm happy to
[16:10] discuss roadmap, whatever. Um, again, you can go check out all the specifications on the website and thank you very much for coming and joining me today.

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.