I really like the way Snorkel Flow allows us to write a handful of programmatic labeling functions to auto-label an entire streaming catalog's metadata instantly, instead of manually tagging thousands of hours of video content. I also love how it tracks the data lineage of our weak supervision rules, making it incredibly simple to audit, tweak, and update our recommendation datasets as viewing trends evolve. The platform's ability to write code-based rules to auto-tag vast video transcripts with metadata based on keywords really removes the need for manual video reviews. Plus, the data lineage tracking allows us to instantly update millions of tags across our entire catalog when user search trends change, simply by editing the code rather than starting over. Review collected by and hosted on G2.com.

Snorkel Flow has a steep learning curve because writing effective labeling functions requires data scientists to think mathematically, preventing non-technical content teams from easily using it. Additionally, the platform is strictly an expensive, enterprise-only product with heavy computational demands, lacking a self-serve tier for teams to easily prototype smaller streaming catalogs. One major enhancement would be adding better native out-of-the-box templates for multi-model OTT content, like pre-configured blocks for video thumbnails or audio transcripts. Additionally, the platform desperately needs real-time execution previews for labeling functions, allowing us to see how a code change affects a tiny slice of streaming search logs instantly without having to re-run the entire pipeline. Setting up Snorkel Flow is a heavy enterprise undertaking, requiring substantial alignment between devops and engineers to configure the secure cloud environment. Review collected by and hosted on G2.com.