Grouped metric charts/alerts emit a failing `DESCRIBE TABLE ``.Bucketed` (UNKNOWN_TABLE) on every evaluation

[ZeynelKoca]

Summary

Every render of a metric builder chart that uses a groupBy (and the corresponding alerts) emits an extra ClickHouse statement:

DESCRIBE TABLE ``.Bucketed FORMAT JSON

which always fails with Code: 60. DB::Exception: Table default.Bucketed does not exist. (UNKNOWN_TABLE). The chart/alert itself works fine (the real query runs via the Bucketed CTE), so it's non-fatal — but it logs a ClickHouse <Error> on every evaluation and adds a wasted failing round-trip.

In our deployment a single 5-minute metric alert (a gauge grouped by pod) produced ~250+ ClickHouse error-log lines per 12h, which is noisy and misleading when scanning ClickHouse error logs.

Version: HyperDX 2.27.0 · ClickHouse 25.7.8.71

Reproduce

1. Create a dashboard tile on a gauge metric with a groupBy, e.g. metric k8s.pod.memory_limit_utilization grouped by ResourceAttributes['k8s.pod.name'].
2. (Optionally create an alert on it so it evaluates on a fixed interval.)
3. Watch the ClickHouse server logs — each evaluation logs DESCRIBE TABLE \`.Bucketed ... (UNKNOWN_TABLE)`.

Root cause

For grouped metric rendering, renderChartConfig builds a synthetic Bucketed CTE and then re-points the chart config's from at it:

• Gauge path: packages/common-utils/src/core/renderChartConfig.ts ~L1375 (CTE) / L1407-1410 (from: { databaseName: '', tableName: 'Bucketed' }).
• Sum/increase group-limit path: ~L1502 (CTE) / L1577 (from: { databaseName: '', tableName: 'Bucketed' }).

The Bucketed CTE is intentional and necessary (its comments explain it wraps the aggregation so Rate/Sum/LastValue are exposed as plain columns — avoiding illegal sum(sum(...)) — and enables top-N group ranking). The problem is only that, while rendering against this synthetic from, a column-metadata lookup (Metadata.getColumns, which issues DESCRIBE ... — packages/common-utils/src/core/metadata.ts L275) runs against the CTE name. A CTE can't be DESCRIBEd as a standalone table, so ClickHouse returns UNKNOWN_TABLE.

Notably the codebase already guards for this in the time-filter path — skipColumnLookup at renderChartConfig.ts L738 skips the metadata lookup when hasSubqueryCte(withClauses) && !databaseName (FROM is a CTE alias with no real base table). The same CTE-aware skip just isn't applied to the column-metadata lookup that resolves the select/groupBy columns for the Bucketed-backed config — so that one still DESCRIBEs the CTE.

Suggested direction

Don't remove the Bucketed CTE (it's required). Instead extend the existing CTE-aware skip so a from whose databaseName is empty and whose tableName matches a declared with CTE is treated as "no base table to DESCRIBE" for all metadata column lookups, not just the time-filter one. (Or have getColumns short-circuit when the target is a known CTE.)

Related

• count_distinct aggFn on metric tiles ignores valueExpression, always counts distinct metric Values #2311 — also a metric-tile builder rendering quirk (same general area), though a different symptom.