You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
PR #50 shipped charon_pipeline_block_duration_seconds and charon_executor_profit_usd_cents as Prometheus histograms using default Prometheus buckets (0.005s, 0.01s, ..., 10.0s).
Issues #217 and #218 flagged these buckets as incorrect for their domains:
block_duration_seconds on BSC has a natural value of ~3s. Default bucket upper bounds top out at 10s — almost all observations pile into the +Inf bucket, making p50/p95 quantiles return +Inf or NaN.
profit_usd_cents observations for real Venus liquidations range from ~5 (0.05 USD) to 5,000,000+ (50,000 USD). The default buckets (0.005 to 10.0) are off by a factor of 10,000+ — every observation lands in +Inf.
This PR ships a dashboard that bakes in both broken panels as official tooling. On first import, Panel 2 will display NaN for all quantiles and Panel 8 (heatmap) will render empty.
Impact
An operator importing this dashboard will see broken panels immediately. There is no in-panel indication that the backing histograms have misconfigured buckets. This is a silent correctness failure on the primary observability output.
Refs #54
Location
deploy/grafana/charon.json — Panel 2 (block_duration_seconds p50/p95), Panel 8 (profit_usd_cents heatmap)
Problem
PR #50 shipped charon_pipeline_block_duration_seconds and charon_executor_profit_usd_cents as Prometheus histograms using default Prometheus buckets (0.005s, 0.01s, ..., 10.0s).
Issues #217 and #218 flagged these buckets as incorrect for their domains:
This PR ships a dashboard that bakes in both broken panels as official tooling. On first import, Panel 2 will display NaN for all quantiles and Panel 8 (heatmap) will render empty.
Impact
An operator importing this dashboard will see broken panels immediately. There is no in-panel indication that the backing histograms have misconfigured buckets. This is a silent correctness failure on the primary observability output.
Suggested Fix
This dashboard must not merge before either:
Shipping the dashboard before the exporter is fixed hardcodes incorrect visualization expectations into operator runbooks.