Tag
Three signals and the different questions they answer, the trace-ID join that delivers 80% of the value, the collector pattern, tail-based sampling that keeps the interesting 100%, and the incident-driven rollout.
The four flag species and why conflating them causes misery, sticky percentage rollouts with one decision point, the staged rollout playbook with observability hooks, and the hygiene that prevents flag archaeology.
The real leak vectors ranked, the non-negotiables at any size, a three-rung maturity ladder that mostly stops at rung two, and rotation as a rehearsed muscle with two-key overlap.
The prime directive (code and schema never jump together), batched backfills as jobs, each engine's locking personality — CONCURRENTLY, NOT VALID, gh-ost — and the destructive-change quarantine.
One protocol, three servers, one reversible decision: what managed Pusher buys, what Reverb and Soketi cost to operate, and the connection-count crossover math that settles it.
Shard sizing between the two failure modes, the 50%/30GB heap rules and why the other half matters, master quorum and watermarks, ILM tiering for log workloads, and snapshots you have actually restored.
Organized by the incident it prevents: maxmemory and eviction policies, AOF vs RDB persistence blends, Sentinel vs Cluster reality, the fork-memory trap and the five-number dashboard.
The four objects that carry a production service, annotated line by line: probes that don't lie, honest resource requests, zero-downtime rollouts, HPA caveats and the boring starter kit that ages well.
From the while-true-Task.Delay era to clustered schedulers: persistent job stores, misfire policies, timezone traps, idempotent job design and the observability layer everyone skips.
Why I keep coming back to Graylog for self-hosted log management: the GELF setup, structured logging from .NET and Laravel, streams that route and alert, and the retention math nobody does.
