Impressive! An entire article about semantic layers, artfully avoids ever defining what a semantic layer is.
Let me take a swipe at it: a semantic layer helps express queries and their results in terms the end-consumers will care about / prefer to reason in, instead of whatever extremely correct and efficient atrocities the database nerds came up with.
Sounds good to me! Semantic layers help expose a more user-friendly view of the data, so it is easier to ask business questions and get accurate results. More technically, it brings modularity and reusability to SQL. Things like joins, aggregate functions, and dimensional expressions are encapsulated as new fields/objects. Typically this logic is rendered at query time rather than pre-computed and materialized. The advantage of that is more flexible iteration and composability. In essence they are highly glorified SQL templating engines.
> A semantic layer, also known as a metrics layer, lies between business users and the database, and lets those users compose queries in the concepts that they understand. It also governs access to the data, manages data transformations, and can tune the database by defining materializations.
> There's a lot of information out there, including from myself about the history and rise [2022], comparing it to an MVC-like approach, or explaining its capabilities. That's why in this article I focus on the why and showcase how to use it in a practical example in the next chapter.
> A semantic layer acts as an intermediary, translating complex data into understandable user business concepts. It bridges the gap between raw data in databases (such as sales data with various attributes) and actionable insights (such as revenue per store or popular brands). This layer helps business users access and interpret data using familiar terms without needing deep technical knowledge. https://www.ssp.sh/brain/semantic-layer#semantic-layer-defin...
I started writing another comment and after looking at your links and was about to quote the definition at the top of "The Rise of the Semantic Layer" as a suggestion, but I realized that it actually isn't that far off, information-wise from the definition youve provided here.
So I took a step back and tried to think about why one "feels" to a reader more like a definition than the other. I think it comes down to phrasing more than informational content. The definition you provide in your comment comes off, for lack of a better term, too much like a sales pitch.
Less is more when it comes to definitions, at least for defining terms in articles/blog posts like these.
Here's my attempt at a better (for this use case) definition:
A semantic layer is an interface to data stores that is designed to be queryable in terms relevant and familiar to those with knowledge of the business domain.
Thanks for your thoughtful comments. I actually refine and update my definition, that's why it is similar.
> A semantic layer is an interface to data stores that is designed to be queryable in terms relevant and familiar to those with knowledge of the business domain.
Sounds good to me, but I think it's too simplified. A semantic layer, IMO, does more. See Julian Hyde's definition, which is also similar to mine, and more involved as well:
> A semantic layer, also known as a metrics layer, lies between business users and the database, and lets those users compose queries in the concepts that they understand. It also governs access to the data, manages data transformations, and can tune the database by defining materializations.
> Like many new ideas, the semantic layer is a distillation and evolution of many old ideas, such as query languages, multidimensional OLAP, and query federation.
I appreciated your feedback. Will think a little more about it.
That's what it does, but that doesn't tell us what it is.
Defining a car as "a vehicular conveyance that helps people get from A to B" is similarly technically correct, but provides little help to the reader in determining if the thing they're looking at is a car or not.
I love a semantic layer as much as the next guy...
Pivoting a decent sized BI shop toward using one instead of splashing the same SQL all over the place is *tough*. It's one of those: "the analyst could have been building important report for director and you want them to create re-usable logic??? we'll do that later, get report done now. Just copy/paste that SQL over here"
This is how you end up with the the 1000 model, "the numbers don't match up", hot mess situations that gain momentum and are hard to slow down.
The flip side is, you often don’t know what needs to be reusable until you’ve had some iterations. Wrong abstractions can be way worse, and also gain their own momentum.
The problem is that often these quick or maybe not reusable are written in such a haste that there's no breadcrumbs left to do the right thing whenever you are done getting that urgent thing out (most likely never because "everything is urgent" :( )
Yeah, minimizing the gap between the semantic layer authoring and adhoc is what you need to do to close that - there has to be a progressive model both for consumption (take this semantic layer, slightly extend/tweak it in an adhoc fashion) and for organically promoting up the adhoc works to the layer.
Right now a lot of semantic tools introduce a big discontinuity in both workflows that keeps the two worlds separate.
That tracks. The semantic layer is like a capital investment that pays off over time. So it can be hard to justify the initial investment if people don't grok the payoff.
We built a transformation library[0] (think a simpler, more performant dbt) for duckdb and I'd really like to create a semantic layer as an extension for it at some point.
Limiting support to only duckdb would make some really useful features trivial to implement. e.g. duckdb has a `json_serialize_sql` function that would handle a lot of the tedious parts of building a semantic layer.
Semantic Layer is about decomposing views into dimensions and aggregates, then letting downstream apps/users compose their own views on top without having to redefine/re-calculate business level metrics.
This makes data analyis more flexible than sql views which are hardcoded on particular groupings.
It's a lot more. A SQL VIEW is just a saved query, where a semantic layer defines the shared meaning of the data, and helps enforce consistent metrics, joins, and logic across tools. You'd be surprised at how many ways "active customer" can be represented as SQL.
create view active_cx as select * from customer join audit_events using(...) join ... where -- active condition
-- use active_cx wherever
select ... from orders join active_cx using(...) where ts > start_of_month() group by active_cx.id
As one of the consumers of a "semantic layer" for many years now, I am firmly convinced that a "single source of truth" must either be useless or a lie.
Ok, the DBA has produced some joins that I can count up to decide how many "customers" we have. We immediately have the issue that a "customer count" from the semantic layer cannot always be the meaningful or relevant figure. In my experience, outside of the exllicit context it was written it, it cannot be the correct figure. So, I have my single source of truth customer count, but my revenue per customer needs to to use a different count that's slightly off. Another analyst needs to produce customer calls to our call center and that uses a slightly different definition. And so on, until the semantic layer is just a special database for pre-defined executive KPI dashboards and no more.
Nothing to do with linear, meaningful projections on embedding spaces, and everything to do with efficient maintenance of legacy data reporting systems.
I think Common Logic ( https://en.m.wikipedia.org/wiki/Common_Logic - ISO/IEC 24707:2007) would be a good addition to any effort trying to add a semantic layer to any database.
This is a good write up that doesn’t require DuckDB as it isn’t specific to a particular database.
Yeah, I think it's great that there are ARD formats and you can access bytes via low level s3 like protocol. This enables interesting tools like DuckDB which can abstract away some stuff, and be fastish and "serverless". However, clearly there is also some kind of marketing hype train and jargon built around it, and it seems like a concerted movement to displace some other "boring" and "uncool" products and technologies. I actually think it's great to displace proprietary services with open formats and protocols. I hope it takes out "data lakes" and co, but I'd love to keep MVC and not invent completely new terms, APIs and ORMs, for things that have been working fine, for a long time.
hey all, another perspective that I have been thinking about is if semantic layers are like ORM for but BI dashboards. Actually, they I think its more than BI dashboards since a similar idea applies to Features. Features in ML land are nothing but a Measure + Entity metadata + TTL. So, really its about higher-order semantics and as we move up the stack, we need richer expression to describe our world.
I think my key takeaway building this is that we need better expression systems and Ibis is a great foundation to build yours..maybe you want to build a language for some other domain etc.
PS: I am one of the authors of bsl and co-founder of Xorq.
I am one of the authors of bsl and founder of xorq.
I’m familiar with Xorq. One of features of the Xorq library that I find interesting is that it catalogs data processing (compute) expressions as it compiles, along with call lineage. Makes reuse easier for SQL and non-SQL processing.
Impressive! An entire article about semantic layers, artfully avoids ever defining what a semantic layer is.
Let me take a swipe at it: a semantic layer helps express queries and their results in terms the end-consumers will care about / prefer to reason in, instead of whatever extremely correct and efficient atrocities the database nerds came up with.
Did I get that right?
Sounds good to me! Semantic layers help expose a more user-friendly view of the data, so it is easier to ask business questions and get accurate results. More technically, it brings modularity and reusability to SQL. Things like joins, aggregate functions, and dimensional expressions are encapsulated as new fields/objects. Typically this logic is rendered at query time rather than pre-computed and materialized. The advantage of that is more flexible iteration and composability. In essence they are highly glorified SQL templating engines.
Julian Hyde (Apache Calcite, Google) gave a crisp presentation on this and how SQL could express 'measures' to bridge the gap: https://communityovercode.org/wp-content/uploads/2023/10/mon...
> A semantic layer, also known as a metrics layer, lies between business users and the database, and lets those users compose queries in the concepts that they understand. It also governs access to the data, manages data transformations, and can tune the database by defining materializations.
There's also now a paper: https://arxiv.org/pdf/2406.00251
Thanks for your hint, I added the definition of it to the article.
From the article:
> There's a lot of information out there, including from myself about the history and rise [2022], comparing it to an MVC-like approach, or explaining its capabilities. That's why in this article I focus on the why and showcase how to use it in a practical example in the next chapter.
[1] https://www.ssp.sh/blog/rise-of-semantic-layer-metrics/ [2] https://cube.dev/blog/exploring-the-semantic-layer-through-t... [3] https://cube.dev/blog/universal-semantic-layer-capabilities-...
My one line definition that I use atm:
> A semantic layer acts as an intermediary, translating complex data into understandable user business concepts. It bridges the gap between raw data in databases (such as sales data with various attributes) and actionable insights (such as revenue per store or popular brands). This layer helps business users access and interpret data using familiar terms without needing deep technical knowledge. https://www.ssp.sh/brain/semantic-layer#semantic-layer-defin...
Edit: I'm the OP.
I started writing another comment and after looking at your links and was about to quote the definition at the top of "The Rise of the Semantic Layer" as a suggestion, but I realized that it actually isn't that far off, information-wise from the definition youve provided here.
So I took a step back and tried to think about why one "feels" to a reader more like a definition than the other. I think it comes down to phrasing more than informational content. The definition you provide in your comment comes off, for lack of a better term, too much like a sales pitch.
Less is more when it comes to definitions, at least for defining terms in articles/blog posts like these.
Here's my attempt at a better (for this use case) definition:
A semantic layer is an interface to data stores that is designed to be queryable in terms relevant and familiar to those with knowledge of the business domain.
Thanks for your thoughtful comments. I actually refine and update my definition, that's why it is similar.
> A semantic layer is an interface to data stores that is designed to be queryable in terms relevant and familiar to those with knowledge of the business domain.
Sounds good to me, but I think it's too simplified. A semantic layer, IMO, does more. See Julian Hyde's definition, which is also similar to mine, and more involved as well:
> A semantic layer, also known as a metrics layer, lies between business users and the database, and lets those users compose queries in the concepts that they understand. It also governs access to the data, manages data transformations, and can tune the database by defining materializations.
> Like many new ideas, the semantic layer is a distillation and evolution of many old ideas, such as query languages, multidimensional OLAP, and query federation.
I appreciated your feedback. Will think a little more about it.
That's what it does, but that doesn't tell us what it is.
Defining a car as "a vehicular conveyance that helps people get from A to B" is similarly technically correct, but provides little help to the reader in determining if the thing they're looking at is a car or not.
I don't think this qualifies as a definition.
Came to make the same comment. I got through like 12 paragraphs, and it still hadn't explained what a Semantic Layer is, so I gave up
I love a semantic layer as much as the next guy...
Pivoting a decent sized BI shop toward using one instead of splashing the same SQL all over the place is *tough*. It's one of those: "the analyst could have been building important report for director and you want them to create re-usable logic??? we'll do that later, get report done now. Just copy/paste that SQL over here"
This is how you end up with the the 1000 model, "the numbers don't match up", hot mess situations that gain momentum and are hard to slow down.
The flip side is, you often don’t know what needs to be reusable until you’ve had some iterations. Wrong abstractions can be way worse, and also gain their own momentum.
The problem is that often these quick or maybe not reusable are written in such a haste that there's no breadcrumbs left to do the right thing whenever you are done getting that urgent thing out (most likely never because "everything is urgent" :( )
Exactly. “Everything is urgent” is the great burden of our time lol
Yeah, minimizing the gap between the semantic layer authoring and adhoc is what you need to do to close that - there has to be a progressive model both for consumption (take this semantic layer, slightly extend/tweak it in an adhoc fashion) and for organically promoting up the adhoc works to the layer.
Right now a lot of semantic tools introduce a big discontinuity in both workflows that keeps the two worlds separate.
That tracks. The semantic layer is like a capital investment that pays off over time. So it can be hard to justify the initial investment if people don't grok the payoff.
this is why I liked Looker. The only way to build reports was from the semantic layer which was easy to use and built into the BI.
we took the same approach when we started https://www.definite.app/.
This is why I liked BusinessObjects
We built a transformation library[0] (think a simpler, more performant dbt) for duckdb and I'd really like to create a semantic layer as an extension for it at some point.
Limiting support to only duckdb would make some really useful features trivial to implement. e.g. duckdb has a `json_serialize_sql` function that would handle a lot of the tedious parts of building a semantic layer.
0 - https://github.com/definite-app/crabwalk
How do you compare your library to SQLMesh?
if you're looking for something like SQLMesh, then I'd stay away from crabwalk.
it's purely meant to run SQL transformations in DuckDB in a reliable way with data lineage.
Is a "semantic layer" nothing more than a fancy name for a SQL VIEW in a NoSQL?
No, it's more than that.
Semantic Layer is about decomposing views into dimensions and aggregates, then letting downstream apps/users compose their own views on top without having to redefine/re-calculate business level metrics.
This makes data analyis more flexible than sql views which are hardcoded on particular groupings.
It's a lot more. A SQL VIEW is just a saved query, where a semantic layer defines the shared meaning of the data, and helps enforce consistent metrics, joins, and logic across tools. You'd be surprised at how many ways "active customer" can be represented as SQL.
Doesn't a view do that?
It sounds like "semantic layer" == views/queries created automatically and on the fly.
Kind of annoying the article writes "What is [a semantic layer] anyway?" twice but never defines it directly.
OP here - I wrote extensively about, that's why I linked to existing article rather than explaining once more, and focusing on the why and how to build one. See also comment above: https://news.ycombinator.com/reply?id=44960004&goto=item%3Fi...
I looked for such a link in TFA, and it wasn't obvious.
I like the idea of a semantic Layer but don't think defining it in yaml is the right way to go about it.
Semantic Layer needs proper language and tooling support which Malloy provides.
Yeah, we still have them, MDX before and now DAX :)
I curate some more on here in case of interest: https://www.ssp.sh/brain/data-modeling-languages.
MDX really did it well, at the cost of being impenetrable to the average user - the drilldown flow you could get is still hard to beat.
Shameless plug for the list, though - I work on https://github.com/trilogy-data/pytrilogy - semantic layer directly embedded in otherwise (mostly) SQL syntax.
I'll do an equivalent example on the taxi dataset when I have some time.
oh nice resource! Another one for you: https://docs.lightdash.com/guides/lightdash-semantic-layer#l...
Anything but a markup language / JSON.
are you using Malloy in prod? Do you know of anyone? I remember the Malloy hype during Data Council in 2023. Haven't heard about it since.
Not yet, but it's evolving nicely, the original team moved to meta and they are building interesting open source stuff
i think the team moved on from google ( which was the main benefactor)
>defined once in a single source of truth
As one of the consumers of a "semantic layer" for many years now, I am firmly convinced that a "single source of truth" must either be useless or a lie.
Ok, the DBA has produced some joins that I can count up to decide how many "customers" we have. We immediately have the issue that a "customer count" from the semantic layer cannot always be the meaningful or relevant figure. In my experience, outside of the exllicit context it was written it, it cannot be the correct figure. So, I have my single source of truth customer count, but my revenue per customer needs to to use a different count that's slightly off. Another analyst needs to produce customer calls to our call center and that uses a slightly different definition. And so on, until the semantic layer is just a special database for pre-defined executive KPI dashboards and no more.
Nothing to do with linear, meaningful projections on embedding spaces, and everything to do with efficient maintenance of legacy data reporting systems.
I think Common Logic ( https://en.m.wikipedia.org/wiki/Common_Logic - ISO/IEC 24707:2007) would be a good addition to any effort trying to add a semantic layer to any database.
This is a good write up that doesn’t require DuckDB as it isn’t specific to a particular database.
Yeah, I think it's great that there are ARD formats and you can access bytes via low level s3 like protocol. This enables interesting tools like DuckDB which can abstract away some stuff, and be fastish and "serverless". However, clearly there is also some kind of marketing hype train and jargon built around it, and it seems like a concerted movement to displace some other "boring" and "uncool" products and technologies. I actually think it's great to displace proprietary services with open formats and protocols. I hope it takes out "data lakes" and co, but I'd love to keep MVC and not invent completely new terms, APIs and ORMs, for things that have been working fine, for a long time.
OT, but I really like the design of their site.
hey all, another perspective that I have been thinking about is if semantic layers are like ORM for but BI dashboards. Actually, they I think its more than BI dashboards since a similar idea applies to Features. Features in ML land are nothing but a Measure + Entity metadata + TTL. So, really its about higher-order semantics and as we move up the stack, we need richer expression to describe our world.
Feature stores explored here: https://www.xorq.dev/blog/featurestore-to-featurehouse
I think my key takeaway building this is that we need better expression systems and Ibis is a great foundation to build yours..maybe you want to build a language for some other domain etc.
PS: I am one of the authors of bsl and co-founder of Xorq.
I am one of the authors of bsl and founder of xorq.
I’m familiar with Xorq. One of features of the Xorq library that I find interesting is that it catalogs data processing (compute) expressions as it compiles, along with call lineage. Makes reuse easier for SQL and non-SQL processing.
FYI - https://github.com/xorq-labs/xorq