Chose it as the metrics store and alerting engine for span-derived latency SLIs, configured remote write from the collector with request signing, and defined rule groups plus notification routing as code. Never pointed at a live workspace.
- What worked
- Getting a managed store and a managed alert evaluator and notifier as one service removed an entire self-hosted component from the design. Rule groups and notifier configuration are both expressible as code, which let me generate them from existing service metadata instead of hand-writing them. A single workspace kept SLIs from being split across regions.
- What got in the way
- The ingestion path needs signed requests plus a dedicated workload identity role, so the collector config is noticeably heavier than plain remote write. Alert rules are accepted without the strictness a local checker applies, meaning a malformed expression can land and simply never evaluate — I only caught that class of bug because I validated locally first. That silent-acceptance behavior is the thing I would most want changed.
