I built p10logs: persistent, multi-cluster Kubernetes pod logs in ~100 MiB RAM. One DaemonSet agent, one hub on a PVC, UI built in. No Loki, no Grafana, no Elasticsearch. Apache-2.0. github.com/p10node/p10logs Demo: demo-p10logs.p10node.org admin/BqJDmpsGRJE85Z0cULFOeoRP Why: Dozzle, kubetail, stern and kubectl logs all stream browser → apiserver → kubelet. One TLS conn per container, a CRI RPC per stream per second, exempt from API Priority & Fairness, killed after 4h, only the current 10 MiB file visible. Pod restarts, logs gone. Structural, not a bug. The agent reads /var/log/pods straight off the node. Metadata from the path, zero apiserver calls. Follows kubelet rotation by inode + first-line hash (inode reuse bit me on k3s after 3h), reassembles 16 KiB CRI partials, checkpoints, spools to disk while the hub is down. Hub: fsynced WAL → dedup by file offset → zstd frames in per-stream chunks per day, trigram bloom per chunk, SQLite catalog that can be rebuilt. Retention by age and % of PVC. Optional S3 offload. Restart anything, lose nothing, duplicate nothing. Laptop bench: 10k lines/s across 50 pods → agent 44 MiB, hub 100 MiB, 7.3× compression. Idle: 18 / 24 MiB. Two static Go binaries, 5 direct deps, no cgo, amd64 + arm64. Multi-cluster with one Helm chart: spokes push to a central hub over HTTPS, egress-only, so home labs behind NAT work. Or hub per cluster + federated queries. v1.1, e2e on kind + my own k3s, not yet in anyone's prod. That's what I need: break it. Demo + Show HN thread in replies.

Your 5-hour AI reset isn’t a bug. It’s a feature of human evolution. In 2026, the only way to distinguish a Soul from a Script is how you handle the "Limit Reached". Here is the Quota Paradox: Why the ceiling is where your best ideas begin. 🧵👇 medium.com/@pierreneter/the-… #AI