Pooled live query and flat change tracking oracle review#

Reviewed state and claim#

Base: 84b828c7e, merged with origin/main at 18abceee4. This record reviews the Git tree that contains it; the final commit or pull request identifies that tree.

The pooled live query oracle claims that a live query filtered only by eq(field, literal) on one eager, non-persisted source, served from an equality partition, publishes what its live-query Collection would: the same rows in key order, values, and status. That holds through sync transactions, optimistic writes confirmed or rolled back, peer mounts and unmounts, and source cleanup and restart. The claim excludes on-demand and persisted sources, DbClient hydration, Suspense, and every clause beyond eq conjuncts, which keep the live-query Collection.

The flat change tracking oracle claims that, for a row whose own fields are all primitives or functions with a plain or null prototype and no symbol keys, the flat tracker reports the same change set as the draft proxy and an independent model, with 0 and -0 equal. Nested values, Dates, Maps, Sets, class instances, and symbol keys are outside it, except that they must fall back to the proxy.

Both oracles use mockSyncCollectionOptions or plain rows; neither claims a real sync adapter's behavior.

RED and GREEN evidence#

Mutants ran through each oracle file alone on the reviewed tree. Each file was restored after each run. Every outcome below is an assertion failure.

Every pooled and flat mutant above fails its pinned history and both campaigns. The pooled grammar delivers initial rows in key order or in reverse and weights a pending insert most peers see before cleanup; the flat grammar weights runs that write and delete a hidden field, write an equal object over a hidden object field, assign a getter-only field its own value, and write the opposite-signed zero over each zero field. Before that weighting, arrival order and frozen optimistic rows escaped about one random campaign in three, and the three descriptor mutants escaped both campaigns. The Object.is-alone flat mutant escaped about one random campaign in two until the zero-flip run.

The pooled oracle found that a pooled view followed its source's status after cleanup instead of entering the live query's terminal error; the repair makes the partition terminate. The flat oracle found that the draft proxy dropped a field added as undefined when another field reverted, on main as well; the repair treats a field the original lacks as changed. The grammar now weights that run, so both campaigns also kill the unrepaired proxy.

The loss audit then widened both oracles. The pooled granular observer had been checked by key membership only, so the payload and lifecycle mutants marked "0 of 7 before" survived it. The flat grammar gained frozen rows, non-enumerable fields, accessor and data defineProperty, stored drafts, and throwing callbacks. Four shapes split the trackers, three of them on main's proxy too. The adopted rule is that defining a field acts as assigning it, and that a non-enumerable field is row data only once the callback writes it. The proxy now records accessors and data defines through the assignment path, rejects a getter-only write of its own value, and ignores the delete of a hidden field it wrote. The flat tracker now compares a hidden object field by contents.

Local-only direct writes are compared with a local-only Collection whose handlers resolve, across mixed multi-key batches, failing batches, schema rejection, and every fallback for insert, update, and delete. Mutants that ignore handler types, drop the pending or persisting check, drop the whole transaction check, run before the ambient branch, write only the first mutation, or never complete the transaction fail 4, 4, 4, 8, 12, 1, and 9 tests.

Two shared conformance scenarios, eq-filter-rows and eq-filter-peers, run the pooled path under React, Vue, Solid, Svelte, and Angular. A partition that ignores a row's previous group fails both under every adapter. eq-filter-peers also checks the source's public subscriberCount: an adapter that declares pooledEqFilters must share one subscription for two queries on the same fields. Disabling pooling or using one partition per query fails it under React; before this check the first passed. Each adapter's driver declares pooledEqFilters.

The loss audit found that a partition released its source one second after its last listener, whatever gcTime its views had. A focused release-timing test, pooled-live-query-gc.test.ts, now compares pooled and live-query Collection release times. A fixed delay failed 5 of 10 cases, a missing 50 ms floor for unsubscribed queries 1, a last-gcTime-wins rule 1, and releasing at gcTime 0 2. Release timing is resource lifetime, so the publication oracle does not observe it.

Pooling later admitted residual conjuncts: any where conjunct that reads only the query's own row, beside at least one eq, is evaluated per view with the compiler's evaluator. Peers may add not(eq(g, literal)), which the model evaluates itself, and a pinned history moves rows in and out of a view within one group. A view that ignores the residual, reports unfiltered entries, or sends a row entering or leaving it as an update fails the pinned history and both campaigns. Hiding part of a group makes some earlier histories rarer, so in one full run the arrival-order, normalization, delete-plus-insert, and dropped in-group update mutants escaped the random campaign; the fixed campaign and their pinned histories still kill each. Peers' literals were weighted toward the normalized values to keep the normalization mutant in the fixed campaign.

Pooled live query oracle#

Flat change tracking oracle#

Open work#

• Svelte's suite reads @tanstack/db from its built dist, so a mutant must type-check and be rebuilt before Svelte can observe it.