The sports data stack is quietly being rebuilt: what the last 72 hours actually show

The most useful thing published about sports technology in the last 72 hours has nothing to do with sport. On 30 September, object-storage vendor Tigris published an engineering post-mortem explaining why it moved asynchronous tasks such as garbage collection out of FoundationDB and onto Kafka. The reason matters to anyone running sports analytics: if a worker dies mid-job, the job dies with it. "Once the worker consumes a job, it's no longer in the queue and that job dies with it," the company wrote.

Sport is now a pile of exactly those jobs. Tracking feeds, video pipelines, betting settlement, highlight rendering. None of it tolerates a dropped message.

Tigris is blunt about the trade-off. It still keeps some queues inside FoundationDB, because the database removes the dual-write problem. Scheduling, though, "requires many writes and scans, which puts read load on FoundationDB that directly competes with user requests." The company implemented Apple's QuiCK paper, the queuing system behind CloudKit, and describes it as "a queuing system that uses FoundationDB's Record Layer to implement a message queue based on time." The piece is also unusually funny about Kafka, comparing self-hosting it to waking up as "a monstrous vermin" while everyone politely asks you to move on with life.

The sports angle is not a stretch. Leagues and broadcasters have spent a decade promising data-native coverage, and the pitch usually stops at the dashboard. The undercard is orchestration: which events get processed, in what order, and what happens when a worker falls over at 3am during a match.

Europe's data centres go quiet on numbers

On the same day, Dutch outlet NL Times reported the results of a year-long investigation by Lighthouse Report with Trouw and other European media into data centres' environmental reporting. The finding is stark: in the Netherlands, fewer than a quarter of larger data centres publish figures on electricity and drinking water use.

The numbers behind that gap are specific. According to the Dutch Datacenter Association, there were 186 commercial data centres with a capacity of 500 kilowatts or more in the Netherlands at the end of last year. The Netherlands Enterprise Agency, which collects the data for the EU, holds records on only 104 of them. Public figures exist for the electricity usage of 44 centres and the water usage of 47. The European Energy Efficiency Directive has required reporting at that threshold for three years.

Statistics Netherlands puts Dutch data centre electricity use at 5.1 billion kilowatt-hours in 2024, 4.6 percent of national consumption and close to double the level five years earlier. Grid operator TenneT projects 10 to 15 percent by 2030. The RVO disclosed this summer that Microsoft's largest Dutch data centre alone accounts for 1 percent of national electricity consumption. Google, which runs two large data centres in the country, does not disclose its figures.

For sports this is not abstract. Streaming, real-time tracking and computer vision all sit on the same grid, and the same reporting blind spot.

Querying the lake without moving it

Also on 30 September, AWS announced that Aurora PostgreSQL can now query Apache Iceberg and Parquet data directly, via foreign tables, using DuckDB's query engine embedded in PostgreSQL. The company says the capability is generally available on Aurora PostgreSQL starting with versions 17.11 and 18.6 in all commercial and GovCloud regions, at no additional charge.

The pitch is the removal of ETL. "Accessing it has typically required pipelines that copy data from your data lake into Aurora, driving up costs and engineering work as schemas evolve," the announcement states. For a sports organisation, that is the difference between a data lake that sits in a warehouse and one an analyst can actually interrogate during a season.

Clubs and leagues have been promising data-native sporting ecosystems at trade events all week. Consultancy-me.com covered a Pure Sports leader discussing exactly that at LEAP 2026 on 29 September, and SVG Europe published a piece on the same day about infrastructure for high-volume media. Neither is a product launch. Both describe a market that assumes the data layer works.

It increasingly does. That is the story.

The research layer is moving too

On 29 September, researchers Osayamen Jonathan Aimuyo, Swapnil Gandhi and Christos Kozyrakis submitted a paper to arXiv describing Purlin, a communication framework that separates orchestration from the datapath in GPU collectives. The paper reports latency speedups of up to 5.14x and bandwidth improvements of up to 4.50x across seven collectives on A100, H200 and B200 GPUs. Integrated into the SGLang serving stack, Purlin improved offline LLM serving throughput and interactivity by 1.13x on average and up to 1.37x, and online inference interactivity by 1.26x on average and up to 2.85x, with the largest gain under overload.

Those are not sports benchmarks. But the workloads are the same shape as automated highlight generation, multi-camera tracking and real-time tactical inference, where latency under load is the whole product.

The gap between a 2.85x interactivity gain in a paper and a working touchline system is enormous, and nothing in the dossier suggests anyone has closed it. Still, the direction of travel is consistent across every source published this week: the interesting work is happening below the interface, in orchestration, query engines and network collectives.

What the week did not settle

It did not settle privacy. On 30 September, 404 Media reported that the Trump administration is using the 1980s High Intensity Drug Trafficking Area programme to funnel local automated licence plate reader data from Flock, Axon and other vendors into federal surveillance centres, and in some cases onward to the DEA's National License Plate Reader Program. The records were obtained through public records requests by Cris van Pelt, creator of HaveIBeenFlocked.com. Jeramie Scott of the Electronic Privacy Information Center told 404 Media: "If you're pissed about Flock then you should be pissed about this."

Sports venues are part of that camera estate. So are the parking lots around them.

It did not settle health data either. The Guardian reported on 30 September that more than 44,000 people have filed formal objections under article 21 of the UK GDPR to NHS England's Palantir-powered Federated Data Platform handling their information, with campaigners citing Palantir's work for the Israeli military and for US immigration enforcement. Palantir says the platform is helping cut waiting lists, citing 117,000 additional operations, a 14.3 percent reduction in discharge delays for long-stay patients and a 5.6 percent improvement in cancer diagnosis within 28 days.

And it did not settle the environmental question. The Lighthouse Report investigation found that European data centres largely will not say how much water and power they consume, three years after the reporting obligation began.

Put together, the last 72 hours describe a sports technology sector whose visible layer, the dashboards, the tracking graphics, the AI commentary, is running on an invisible layer that is being rebuilt in public engineering blogs and arXiv preprints while its resource and privacy costs remain largely undisclosed. That is not a scandal. It is a measurement problem, and it is the one worth watching.