Date: 2026-04-23
Subsystem: data-pipeline
Severity: Medium
Impact: Integrity
Hypothesis by: gpt-5.4, high
Expected Behavior
Two aliased operation selections on the same BaseStateChange should both resolve against the real operation row while respecting their own requested fields. Asking for id under one alias must not cause a sibling alias requesting operationXdr or resultCode to receive zero values.
Mechanism
resolveStateChangeOperation puts the alias-local DB column set into the loader key, but operationByStateChangeIDLoader fetches all batched keys with keys[0].Columns. newOneToOneLoader then maps the loaded Operation only by the synthetic stateChangeID, so both aliases for the same state change get the same sparse struct shaped by the first alias rather than their own selection.
Trigger
Query a state change's parent operation twice in one request with different aliases and different sub-selections, for example: stateChanges(first: 1) { edges { node { opLite: operation { id } opFull: operation { operationXdr resultCode } } } }. When opLite is processed first, opFull should expose empty or incorrect operation data.
Target Code
internal/serve/graphql/resolvers/resolver.go:124-139 — builds the StateChangeID/Columns loader key from the alias-local selection setinternal/serve/graphql/dataloaders/operation_loaders.go:77-99 — state-change batch path uses only keys[0].Columnsinternal/serve/graphql/dataloaders/loaders.go:156-166 — one-to-one loader reuses one fetched row per synthetic stateChangeIDinternal/data/operations.go:295-323 — sparse projection query for operations by state-change tuple
Evidence
The dataloader key intentionally carries Columns, proving the resolver expects projection-sensitive batching. The batch function discards that distinction, and the generic loader returns a single fetched struct to every matching key regardless of which alias asked for which operation fields.
Anti-Evidence
Single-alias requests are safe, and matching alias projections would hide the problem. The issue also requires the duplicated aliased field loads to coalesce into the same dataloader batch.
Date: 2026-04-23
Subsystem: data-pipeline
Severity: Medium
Impact: Integrity
Hypothesis by: gpt-5.4, high
Expected Behavior
Two aliased
operationselections on the sameBaseStateChangeshould both resolve against the real operation row while respecting their own requested fields. Asking foridunder one alias must not cause a sibling alias requestingoperationXdrorresultCodeto receive zero values.Mechanism
resolveStateChangeOperationputs the alias-local DB column set into the loader key, butoperationByStateChangeIDLoaderfetches all batched keys withkeys[0].Columns.newOneToOneLoaderthen maps the loadedOperationonly by the syntheticstateChangeID, so both aliases for the same state change get the same sparse struct shaped by the first alias rather than their own selection.Trigger
Query a state change's parent operation twice in one request with different aliases and different sub-selections, for example:
stateChanges(first: 1) { edges { node { opLite: operation { id } opFull: operation { operationXdr resultCode } } } }. WhenopLiteis processed first,opFullshould expose empty or incorrect operation data.Target Code
internal/serve/graphql/resolvers/resolver.go:124-139— builds theStateChangeID/Columnsloader key from the alias-local selection setinternal/serve/graphql/dataloaders/operation_loaders.go:77-99— state-change batch path uses onlykeys[0].Columnsinternal/serve/graphql/dataloaders/loaders.go:156-166— one-to-one loader reuses one fetched row per syntheticstateChangeIDinternal/data/operations.go:295-323— sparse projection query for operations by state-change tupleEvidence
The dataloader key intentionally carries
Columns, proving the resolver expects projection-sensitive batching. The batch function discards that distinction, and the generic loader returns a single fetched struct to every matching key regardless of which alias asked for which operation fields.Anti-Evidence
Single-alias requests are safe, and matching alias projections would hide the problem. The issue also requires the duplicated aliased field loads to coalesce into the same dataloader batch.