The technical reasons for the boundary are irrelevant; all people care about is creating the charts they need. For them, a gerrymandered visualization stack is a bad experience. (View Highlight)
The same is true for BI. People don’t want to hop from Hubspot to Amplitude when they’re trying to figure out which product features recent webinar attendees are using; they don’t want to recreate a notebook’s output in Looker to make a dashboard out of Prophet forecast. The boundaries we put between these different types of data exploration—vertical-specific analysis, Python analysis, SQL analysis, pivot table analysis—are technical, not experiential. (View Highlight)
I’m skeptical of decentralized consumption apps: Visualizations are really hard to build. Tableau, to take an obvious example, is breathtakingly far from being a simple charting configuration. All of the things it can do—derived calculations, facets, formatting, animations, and much more—are both complex and, to many of Tableau’s customers, necessary. Short of embedding it directly, there’s no reasonable way to port something created in Tableau into another product. (View Highlight)
Note: Unable to port
Consumption, instead, should be split into two categories. The first is generic exploration. This is the type of data consumption that starts with a business question and ends with a quantitative answer. Though the tools used in the middle may differ—Python, R, visual exploration, pivot tables, SQL, Excel—and the form of the output varies, the structure of this type of work is broadly similar.
In other cases, people need to solve specific problems. They want to analyze A/B test results; they want to examine the emotional sentiment in support tickets; they want to create financial forecasts based on standard accounting principles. Specialized tools are hugely valuable here—more valuable, even, than seamless interoperability with other tools. (View Highlight)
Note: Specific problem - specialized tool
Generic exploration - generally accepted structure
This structure is already emerging in other parts of the stack. Snowflake is the storage layer’s planet; specialized warehouses like Clickhouse are its moons, dedicated to particular types of processing. Tools that ingest obscure sources (and potentially reverse ETL services?) resolve around Fivetran; products like Zingg that handle specialized transformations orbit dbt.
The BI planet hasn’t fully formed yet. But that, I think, is what we’re gravitating towards. (View Highlight)
