FOSDEM and the Conference Replay Problem
Every February, thousands of developers descend on Brussels for FOSDEM, one of the largest open-source conferences in the world. Within days, attendees are asking the same question on social media: where are the talk videos? The answer is usually “in a few weeks,” and the reason reveals an interesting infrastructure problem that affects most large conferences.
The Scale Challenge
FOSDEM runs over 800 talks across two days in more than 50 parallel tracks. Each talk needs to be recorded, processed, reviewed, and published with metadata. The conference relies entirely on volunteers, but even well-funded commercial conferences face similar delays. The problem isn’t just manpower—it’s the sheer computational and organizational complexity of the pipeline.
Recording is the easy part. Modern conference setups capture speaker audio, slide feeds, and sometimes audience video simultaneously. These streams need to be synchronized, which isn’t trivial when you’re dealing with multiple hardware clocks that drift independently. A few frames of desync between audio and video becomes noticeable quickly, and slide transitions that don’t align with what the speaker is saying look unprofessional.
Processing Bottlenecks
The real bottleneck is post-processing. Raw recordings need to be transcoded into multiple formats and resolutions for different devices and bandwidth conditions. A single hour of 1080p video can take 20-30 minutes to transcode on a single CPU, even with hardware acceleration. Multiply that by 800 talks and you’re looking at thousands of CPU-hours.
Most conferences can’t simply throw money at cloud compute because video transcoding is expensive at scale. Spinning up hundreds of instances for burst processing costs real money, and volunteer-run events operate on tight budgets. The result is that processing happens on whatever hardware volunteers can access, often spreading the work across personal machines and donated server time.
Quality control adds another layer. Someone needs to verify that each video actually contains the correct talk, that audio levels are reasonable, and that there aren’t technical glitches that render the content unusable. Automated checks can catch some issues—silence detection, black frames, extreme volume spikes—but subtle problems like the wrong slide deck being captured require human review.
The Metadata Problem
Even after videos are processed, they need metadata: titles, speaker names, abstracts, tags, timestamps for chapter markers. This information often comes from multiple sources. The conference schedule database might have one version of a talk title, the speaker might have submitted a different version, and what they actually presented could be a third variation.
Chapter markers are particularly valuable for technical talks where viewers want to jump to specific topics, but generating them requires either manual timestamping or sophisticated speech-to-text processing with topic segmentation. The latter is getting better with modern ML models, but it’s still not reliable enough to ship without review.
Why Not Stream Everything Live?
Live streaming seems like an obvious solution—just publish as you record. Some conferences do this, but it introduces new problems. Live streams need reliable, high-bandwidth uplinks from every room. Conference venues aren’t always equipped for this, and WiFi doesn’t cut it when you’re pushing multiple HD streams simultaneously.
More importantly, live streaming means mistakes go out unedited. If a speaker’s microphone fails for two minutes, or someone walks in front of the camera, or slides don’t advance properly, that’s what the internet sees. Post-processing lets volunteers fix these issues, resulting in a much better viewing experience.
There’s also the speaker consent issue. Many speakers agree to be recorded but want the option to review their talk before it goes public, especially if they misspoke or want to redact something. Live streaming eliminates that option.
Incremental Publishing
The best approach many conferences have found is incremental publishing. Popular talks get prioritized and go out first, while less time-sensitive content follows. Automated processing handles the bulk of transcoding, with human review batched efficiently.
Some conferences publish “fast” versions quickly—minimal processing, basic quality checks—and then update videos with improved versions later. This trades some polish for availability, which most attendees prefer over waiting weeks for perfect output.
The core lesson is that conference video publishing is a systems problem, not just a technical one. It involves coordination between recording teams, processing infrastructure, quality reviewers, and metadata managers. Even with perfect automation, someone needs to orchestrate the pipeline and handle edge cases.
For attendees waiting impatiently for talks to appear online, understanding these constraints helps. For conference organizers, the challenge is balancing quality, speed, and resource constraints in a way that serves the community without burning out volunteers.