Action Outcome Distributed Vendor SQL Replay 011 - Raw Research Artifact

Generated at: 2026-06-21T02:54:17Z Coordinator endpoint: http://127.0.0.1:19241/ SQL seed: docs/research/results/action_outcome_sql_replay_006.sql Node IDs: 1, 8

Load Summary

Seed statements attempted: 203
Seed statements succeeded: 203
Seed statements failed: 0
Seed row-count status: pass
Vendor table row-count status: pass
Distributed ingest status: pass
Full distributed SQL/JOIN/window status: pass
Boundary classification status: pass

Node-Local Stats Delta

Node ID	Stats URL	ticks_ingested	ticks_stored	ticks_dropped	partitions_created
1	`http://127.0.0.1:19241/stats`	161	161	0	4
8	`http://127.0.0.1:19242/stats`	208	208	0	5

Owner Split

Table	Stable table_id	Owner node	Expected/live rows
`action_outcome_episode_metrics`	57172	1	101
`action_outcome_episodes`	3679	8	32
`action_outcome_gate_suppressions`	16692	1	21
`action_outcome_replay_recommendations`	9749	1	18
`action_outcome_retrieval_quality_labels`	37786	8	26
`action_outcome_vendor_queries_010`	58879	8	6
`action_outcome_vendor_recommendations_010`	32465	8	72
`action_outcome_vendor_retrieval_010`	52359	8	72
`action_outcome_vendor_suppressions_010`	58120	1	21

Row Count Verification

Seed Tables

Table	Expected Inserts	Live Rows	Status
`action_outcome_episode_metrics`	101	101	pass
`action_outcome_episodes`	32	32	pass
`action_outcome_gate_suppressions`	21	21	pass
`action_outcome_replay_recommendations`	18	18	pass
`action_outcome_retrieval_quality_labels`	26	26	pass

Vendor Tables

Table	Expected Rows	Live Rows	Status
`action_outcome_vendor_queries_010`	6	6	pass
`action_outcome_vendor_recommendations_010`	72	72	pass
`action_outcome_vendor_retrieval_010`	72	72	pass
`action_outcome_vendor_suppressions_010`	21	21	pass

Distributed JOIN/Window Classification

Check	Left/table owner	Right owner	Rows	Expected	Status
`failed_repeat_join`	`action_outcome_vendor_recommendations_010` -> 8	`action_outcome_vendor_queries_010` -> 8	5	5	`pass`
`context_top_action_join`	`action_outcome_vendor_recommendations_010` -> 8	`action_outcome_vendor_queries_010` -> 8	6	6	`pass`
`suppression_join`	`action_outcome_vendor_suppressions_010` -> 1	`action_outcome_vendor_recommendations_010` -> 8	21	21	`pass`
`misleading_retrieval_join`	`action_outcome_vendor_retrieval_010` -> 8	`action_outcome_vendor_queries_010` -> 8	23	23	`pass`
`row_number_window`	`action_outcome_vendor_recommendations_010` -> 8	n/a	72	72	`pass`
`lag_window`	`action_outcome_vendor_recommendations_010` -> 8	n/a	72	72	`pass`

Interpretation

Experiment 011 separates distributed-safe replay checks from the current distributed planner boundary. Under the default 1/8 two-node ring, the query, recommendation, and retrieval vendor tables are co-located on node 8, so failed-repeat JOIN, context top-action JOIN, and misleading retrieval JOIN all pass through the coordinator.

The suppression table is owned by node 1 while recommendations are owned by node 8. The bounded small-table hash JOIN path now gathers both operational tables under the coordinator row cap, replays them into a temporary typed pipeline, and executes the original hash JOIN locally. This turns the former cross-node suppression JOIN boundary into a passing check without requiring a full distributed SQL optimizer.

The window checks now pass in cluster mode. The coordinator fetch-and-compute path preserves declared operational-table values when it materializes the temporary full-data table, so ROW_NUMBER and LAG match the single-node/vendor baseline.

Next Best Step

Define an explicit shard-key or table-level distribution policy for symbol-less operational tables. The small-table JOIN path proves correctness for bounded control tables; production promotion now needs placement policy, row-cap telemetry, and a larger cost-based JOIN planner only for tables that exceed the small-table boundary.