Context
The DAG viewer should explain why a selected post-ASAP workload is beneficial even when sharing and replacement make the graph look structurally more complex.
The current viewer has one Pre/Post-ASAP workflow: selecting one query displays that query's DAG; selecting multiple queries displays their unioned workload DAG. There are no separate Single, Compare, or Union modes.
This issue adds explicit planner/exported cost and benefit annotations. The viewer must only render metadata present in generated JSON; it must not infer costs, match nodes by guessed signatures, or rely on hand-edited demo JSON.
Cost and benefit semantics
All modeled recurring costs use cost units per second:
maintained_cost_rate =
update_rate * maintenance_cost_per_update
+ evaluation_rate * summary_read_cost
recompute_cost_rate =
evaluation_rate * raw_recompute_cost
evaluation_rate = sum(1 / query_interval_i)
For any selected alternative:
estimated_benefit_rate = baseline_cost_rate - selected_cost_rate
estimated_benefit_ratio = estimated_benefit_rate / baseline_cost_rate
The baseline must be named explicitly, for example pre-ASAP recomputation or the highest-ranked legal non-selected candidate. If baseline_cost_rate <= 0, the ratio is unavailable.
Finite-run or one-shot totals require an explicit horizon H:
total_cost(H) = recurring_cost_rate * H + one_shot_cost
Rate costs and one-shot costs must never be silently added. Shared nodes must be counted once in workload totals.
Annotation schema
Define a structured optional annotation rather than an unqualified number. The exact Rust shape may vary, but it must carry equivalent semantics:
struct CostAnnotation {
value: Option<f64>,
unit: CostUnit,
source: CostSource, // Modeled, Measured, Unavailable
baseline: Option<BaselineRef>,
delta: Option<f64>,
model_version: Option<String>,
benchmark_id: Option<String>,
inputs: Vec<CostInput>,
}
Modeled: generated by a named/versioned cost model.Measured: loaded from a reproducible benchmark artifact.Unavailable: no value; never encode an unknown value as 0 or another synthetic number.
Every numeric value must have units and provenance.
Granularity
Initial supported annotations:
- replacement-region baseline cost, selected cost, and benefit;
- whole selected-workload cost and benefit;
- avoided recomputation for a shared sub-DAG;
- edge cost only when genuinely attributable to the edge, such as transfer or materialization.
Arbitrary path cost is out of scope until path ownership and shared-node double-counting semantics are defined.
Scope
- Extend generated DAG-export JSON with optional, backward-compatible node/edge/replacement-region/workload annotations as justified above.
- Populate modeled annotations from planner/cost-model output, not frontend inference.
- Allow measured annotations only through a versioned benchmark artifact consumed by the exporter/backend.
- Render concise on-graph benefit/cost badges and full inputs/provenance in the resizable sidebar.
- Clearly distinguish Modeled, Measured, and Unavailable states.
- Support the existing one-query and multi-query Pre/Post-ASAP workflow.
Acceptance criteria
- Real queries run through lowering, ASAP-aware mapping, post-ASAP generation, and
dag_export; demo JSON/HTML is regenerated from that pipeline and is not manually patched.
- Viewer reads only explicit JSON mappings and annotations.
- Missing values display
Not estimated, never a fabricated number.
- Modeled and measured values are visually distinct and show units, provenance, and model/benchmark version.
- Workload totals count shared nodes once and tests detect double counting.
- Baseline, selected cost, absolute benefit, and benefit ratio follow the formulas above.
- One-shot totals require an explicit horizon; unit-incompatible aggregation is rejected.
- Selecting one query and multiple queries both display correct Pre/Post-ASAP annotations.
- Existing exports without annotations remain fully supported.
Follow-ups
Context
The DAG viewer should explain why a selected post-ASAP workload is beneficial even when sharing and replacement make the graph look structurally more complex.
The current viewer has one Pre/Post-ASAP workflow: selecting one query displays that query's DAG; selecting multiple queries displays their unioned workload DAG. There are no separate Single, Compare, or Union modes.
This issue adds explicit planner/exported cost and benefit annotations. The viewer must only render metadata present in generated JSON; it must not infer costs, match nodes by guessed signatures, or rely on hand-edited demo JSON.
Cost and benefit semantics
All modeled recurring costs use cost units per second:
For any selected alternative:
The baseline must be named explicitly, for example pre-ASAP recomputation or the highest-ranked legal non-selected candidate. If
baseline_cost_rate <= 0, the ratio is unavailable.Finite-run or one-shot totals require an explicit horizon
H:Rate costs and one-shot costs must never be silently added. Shared nodes must be counted once in workload totals.
Annotation schema
Define a structured optional annotation rather than an unqualified number. The exact Rust shape may vary, but it must carry equivalent semantics:
Modeled: generated by a named/versioned cost model.Measured: loaded from a reproducible benchmark artifact.Unavailable: no value; never encode an unknown value as0or another synthetic number.Every numeric value must have units and provenance.
Granularity
Initial supported annotations:
Arbitrary path cost is out of scope until path ownership and shared-node double-counting semantics are defined.
Scope
Acceptance criteria
dag_export; demo JSON/HTML is regenerated from that pipeline and is not manually patched.Not estimated, never a fabricated number.Follow-ups