On 19 November 2026, ClickHouse announced a Power BI connector that talks to the database over Arrow Flight SQL, now in public preview. The stack I built used Apache Superset. The post notes that most of a Power BI refresh happens after the data arrives, when Power BI compresses the rows into its in-memory model.
At a US consumer app, reporting used to be forms on MySQL, then an export to a spreadsheet. On a MySQL box with 80 to 90 GB of RAM, those queries took five to seven minutes. Separately, as the case study records, a business question often became an engineering task, and those reports took hours or days.
I moved the events into ClickHouse, a column-oriented database built for analytical queries, which is how ClickHouse describes itself. We write millions of events a day. Those queries now return in seconds. The other change is who asks them. The moderation team, who are non-technical, now ask their own questions through reusable chats and dashboards.
The two clocks
Those two time scales should stay separate. The hours or days on the case study is the time it took a person to prepare an answer: filter the data, export it, and plot the result. The five to seven minutes is the query itself, on that MySQL machine, before the events had a store built for this kind of read. Faster queries matter. They're only one part of the path from a form and a spreadsheet to a team that can ask for itself.
The tradeoffs I accepted
I chose to dual-write into the new store instead of setting up change data capture, because it was simpler. CDC would have followed changes out of the source automatically. For a first version I wanted fewer moving parts. Ingestion in this first version is still manual, and that's the part I want to change next.
Event shapes weren't all the same, so I used a JSON column to let them vary instead of forcing every field into a fixed table up front.
For exploration I put Apache Superset on top, so people could drill down in dashboards instead of waiting for a one-off export. That foundation is the one already on the case study: a near-real-time analytics stack with ClickHouse and Apache Superset, including KPI dashboards and scheduled email reports, plus the internal infrastructure and a Superset deployment on AWS Fargate.
For the language interface I did not write a custom connector. I used Apache Superset's official MCP service, which is part of the Apache Superset codebase. There was no custom build.
When a question is unclear
When Claude isn't sure, it says so. The answers still need a person to review them. I don't want a confident wrong number going back to the team as if it were settled.
The case study already says that metric definitions, query complexity and source freshness still decide whether an answer is useful. In practice the rule is simpler: if the model is unsure, it should say so, and a person should look before the answer is treated as done.
The security setup
Everything in this path runs in the same VPC. The data going through it is PII-free; the only identifiers in it are user IDs.
What I watch once people start asking
The metric I care about is which questions non-technical teams ask, and whether the answers make sense to the business. I log every query. AI helps analyze whether a question is relevant, and team leads review. The risk is a wrong question leading to a wrong decision.
I audit hard for the first couple of months, until the answers are roughly right. I'm not waiting for a perfect score before the team can use the path. Once I trust the questions, the answers and the harness, team leads decide how much auditing to keep.
What I would do next
The first version still depends on manual ingestion. In the coming weeks or months I want to move that over to CDC automation. Dual-write was the right call for simplicity when the stack was new. CDC is the next step if we want ingestion to follow the source without someone moving the data by hand.
The other decisions still hold up: ClickHouse for this event volume, a JSON column for flexible event shapes, Superset for drill-downs, the official Superset MCP service, and a person in the loop when an answer is shaky.
What I took from it
With the data in reach, through this analytics stack and an MCP, query writing, analysis and recommendations can move into non-technical hands. Analysis moves to the people who have the questions. The team stays lean, and it stays fast, because engineers aren't sitting in the middle of every routine question.
Two years out, I think small and medium companies will have departments that just ask questions, instead of IT or analytics teams exporting data. Something like an internal company OS would connect MCPs with context across employee records, payments, sales, marketing and every database, and would answer non-technical people. Cloudflare OS, which Cloudflare open-sourced in August 2026, is an example of that kind of platform. I expect teams to stay lean, processes to get tighter, and models from Anthropic, OpenAI and xAI to help companies decide better. That's a prediction. I haven't deployed a system like this.
The case study for the foundation is From a reporting bottleneck to business answers.
If a reporting bottleneck like this is sitting in your business, I take that work through Developer Labs.