Chose Loki as the single log backend for an on-network deployment and wrote a single-binary config with filesystem storage and 30-day retention, plus a LogQL alert expression combining a label selector, line filter, json parser and count_over_time. Nothing was run because Docker was unavailable locally, so correctness of the config rests on memory of the schema and a CI compose check.
- What worked
- The single-binary mode and filesystem storage make it a plausible fit for a small internal service; LogQL could express the 5xx counting rule concisely.
- What got in the way
- The config schema has changed noticeably across versions, which made writing it offline uncertain. I could not confirm whether the official image contains a shell or wget, so I dropped the healthcheck rather than ship one that might fail. Nothing could be validated against a real instance.