We saved $500k per year by rolling our own "S3"

Architecture & Serverless Critiques

• Several commenters argue the original design was ill-suited to serverless: pushing 2–6 MB video chunks to S3 for ~2 seconds of life created unnecessary complexity and cost.
• TLS handshakes and disabled keep-alives were viewed as odd bottlenecks; some suggest terminating TLS at the load balancer and using persistent connections internally.
• The new in-memory cache is seen as a good fit for short-lived, loss-tolerant data, but some think this complexity should have been avoided by not overusing serverless in the first place.

Cost Savings vs Engineering Effort

• People question how many engineer-years it takes to build/maintain this versus the $500k/year saved; guesses range from “fraction of one engineer” to “needs three people for safety and bus factor.”
• Others note the team explicitly framed it as worthwhile only because of scale, loss tolerance, and an S3 fallback keeping the system simple.
• Some say negotiations or different AWS tiers (e.g., S3 Express One Zone, PPAs) might have reduced costs without custom infra.

“Rolling Your Own S3” or Just a Cache?

• Multiple commenters point out the system is effectively an in-memory S3-compatible cache in front of S3, not a full S3 replacement; title is seen as overstated.
• Concerns about resiliency: with pure RAM, a node crash could lose pending segments; some suggest NVMe / WAL-style local storage as a cheaper, more durable buffer.

Alternative Designs & Technologies

• Suggestions include: process on upload in a non-serverless service, avoid S3 entirely, or use Kinesis/SQS variants (though size limits are a constraint).
• Other object storage options mentioned: MinIO (now expensive), Garage, Cloudflare R2, on-prem/Ceph, HDFS, or vendor appliances (e.g., Dell object storage).
• Some note this looks like a classic “MVP with S3, optimize later” path, which they consider reasonable.

Cloud vs Self‑Hosted Storage Debate

• Heated discussion over whether “saving a file” is simpler on bare metal vs cloud.
• Pro-cloud side emphasizes built-in redundancy, durability, policies, and managed ops.
• Anti-cloud side stresses simplicity, control, debuggability, and long-term cost; cites ZFS, iSCSI, and custom object stores as viable at scale.

Baby Monitor, Privacy, and Parenting Norms

• Long subthread on Nanit’s model: continuous cloud upload of baby video/audio without E2EE alarms many; some call it “spyware” and worry about exfiltration and training data.
• Others defend it as the only reliably working consumer product they found, especially for remote access (yard, neighbor’s house, hotel bar).
• Alternatives discussed: offline audio monitors, TP-Link/Eufy/Unifi Protect/self-hosted setups, HomeKit Secure Video, VPN/Tailscale, etc.
• Broader philosophical split: some see cloud baby monitoring as unnecessary surveillance and fear-based monetization; others view it as benign convenience, with parenting style and risk tolerance driving choices.