|
|
|
|
|
|
by jas-park-dev
34 days ago
|
|
|
Auto-instrumentation is great until you hit cardinality. We turned on auto-traces across our LLM gateway and within a week our trace storage costs were dominated by attribute cardinality from user-id and prompt-text labels. Pulled those out into a separate column store with sampling. Query latency dropped 40x. Curious how Superlog handles attribute cardinality at scale, especially for AI workloads where prompt content is high entropy. The 'installs itself' magic is only useful if it doesn't blow up the bill. |
|
|