Description
Problem
A PromQL aggregate with no grouping keys, over the output of another aggregate, answers one row when its input is empty. Prometheus answers no sample.
Over a gauge m whose values are all below 1,000, exact, planned with plan_installable and run with the executor:
| Query | Prometheus | Today |
|---|---|---|
count(m >= 1000) |
empty | empty |
count(sum by (job) (m >= 1000)) |
empty | {} 0 |
count by (instance) (count by (job, zone, region) (m{job="a"}) > 10) (instance is absent after the inner by, so the outer grouping has no keys) |
empty | {} 0 |
Probably the same cause (not traced) makes some nested aggregates fail to install instead, with native output type differs from Planner output. For example sum(min without (resource) (kube_resourcequota{type="hard"})) and quantile(0.9, sum without (cpu) (rate(node_cpu_seconds_total{mode="system"}[5m]))). The executor's aggregate nullability reads the same identity-column test (below).
The first two rows were measured on #1242's head before its round-5 fixes, so this is not caused by that PR; its fuzz harness (a PromQL reference evaluator) finds it through without queries that now plan, such as count without (zone) (count without (instance, job) (max by (instance, zone, job) (m{job="a"})) > 4).
Why
The executor decides between SQL and PromQL semantics by whether its input schema has a series identity column:
crates/executor/src/operators/summary/mod.rs, thePerGroupSummaryBuildrun:emit_empty_global: !input_schema.has_promql_series_identity(). With no groups and that flag, an empty input still gets the global group, as a SQLSELECT count(*)does.crates/executor/src/operators/aggregate/mod.rs(nullability ofsum,min,max,avg) andaggregate/columnar.rsuse the same test.
An aggregate's output ([job, sum] for sum by (job)) has no identity column, so an aggregate over it is treated as a SQL aggregate over a relation. The IR already knows which it is: OperatorNode::result_kind is InstantVector there and Relation only for SQL tables. The summary merge path already uses it (built_from_relation in crates/executor/src/physical_planner/mod.rs).
Proposal
Decide the SQL-versus-PromQL rule in the lowering, from the IR's result kind, and carry it in the operator:
Kind::PerGroupSummaryBuild(and the keyed and all-groups builds) gets the same opt-in the merge has (with_emit_empty_global), set only when the build's input is aRelation.Operator::aggregate's nullability takes the same flag instead of reading the identity column.
Scope and acceptance
- In scope: global aggregates and their output types over PromQL aggregate outputs.
- Done when
count(sum by (job) (m >= 1000))answers no sample,sum(min without (resource) (kube_resourcequota))installs and matches its reference, and SQLSELECT count(*) FROM t WHERE falsestill answers one row of 0.
Activity