ID

R34

Status

Backlog

Bucket

cleanup

Priority

13

Theme

legacy-migration

sis-graphql-spec migration to graphitron-rewrite

Track the consumer-side schema work needed to bring sis-graphql-spec cleanly onto graphitron-rewrite. This plan exists because sis is the canonical large-scale consumer; closing it out validates the rewrite’s classification contracts end-to-end and lets us close courtesy windows on shims (notably retire-synthesis-shims, which gates on this work).

The earlier blocker, ~200 sis event types colliding on __NODE_TYPE_ID = "195" after silent promotion to NodeType, has been resolved upstream by retiring the type-level shim in TypeBuilder.buildTableType: a @table SDL type without implements Node @node is now a TableType regardless of metadata. The latest sis build against rewrite confirms no typeId-collision errors remain. The residual work splits into three independent phases plus one upstream bucket.

Phase 1: declare @nodeId on every Relay-style ID field

The two surviving field-level synthesis shims still fire today on bare scalar ID fields whose context implies they should be Relay node IDs:

  • FieldBuilder Path-2: output ID on a NodeType parent (parent type implements Node @node).

  • BuildContext.classifyInputField: input ID on a @table input whose backing jOOQ class carries NODE_TYPE_ID / NODE_KEY_COLUMNS.

Each fires a per-occurrence WARN. The roadmap item retire-synthesis-shims flips the WARN to a terminal classifier error once consumer schemas declare @nodeId explicitly.

For every WARN line in the sis build log:

  1. Open the SDL field it names.

  2. Add @nodeId to the field declaration.

  3. If the field’s parent type isn’t already implements Node @node, decide whether it should be (the WARN trigger condition implies it usually should), and add it.

Done when the sis build log contains zero @nodeId synthesis WARNs. At that point the upstream shim retirement can land safely.

Estimate: ~250+ fields per the latest build; mechanical, and ideally split across PRs by logical type group.

Phase 2: annotate filter inputs with @table

~25 filter input types currently fail with "table … could not be resolved". Filter inputs need @table(name: "…") so the generator knows which jOOQ table’s columns the filter can target.

For each error in the build log:

  1. Identify the filter input type (typically *Filter or *Where).

  2. Add @table(name: "<jooq-table-name>") matching the table the filter targets.

  3. Verify the filter’s fields all reference columns that exist on that table.

Done when no "table … could not be resolved" errors remain for filter input types.

Estimate: ~25 types; straightforward once the table mapping is known.

Phase 3: clean up author-error directives

A handful of types declare @node or @nodeId but lack the corresponding metadata (no id column, missing keyColumns, etc.). These are pre-existing schema bugs surfaced more loudly now that the type-level shim no longer papers over them.

For each remaining error:

  1. Determine intent: was this meant to be a NodeType, a TableType, or neither?

  2. Either add the missing metadata or remove the directive that doesn’t apply.

Done when the build log contains only deferred-feature errors (Phase 4) and no author errors.

Estimate: small, case-by-case.

Phase 4 (out of scope, tracked upstream): deferred features

Remaining errors flag features the rewrite hasn’t implemented yet. These are graphitron-rewrite roadmap items, not sis migration work. Track them so the migration owner knows what’s blocking final green; do not attempt schema workarounds.

Sequencing

Recommended order:

  1. Phase 2 first: smallest, unblocks the most types from UnclassifiedType.

  2. Phase 3 next: small, removes noise.

  3. Phase 1 last: largest, mechanical, splittable across PRs.

Phase 4 is parallel work upstream, not on the sis side.

Verification per phase

After each phase, rerun the sis build against graphitron-rewrite and diff WARN/ERROR counts by category against the previous run. Each phase should drive its target category to zero without introducing new errors in other categories.

History

The original blocker was the TypeBuilder.buildTableType shim that silently promoted any @table SDL type whose backing jOOQ class carried NODE_TYPE_ID / NODE_KEY_COLUMNS to NodeType, even without implements Node @node. In sis, ~200 event types shared __NODE_TYPE_ID = "195" and were promoted in lockstep, then symmetrically demoted to UnclassifiedType by the registry-uniqueness check. That shim has been retired; @table without implements Node @node is now a TableType regardless of metadata, and field-level synthesis only applies inside types that opt in via implements Node @node. The remainder of this plan tracks the consumer-side schema cleanups that the type-level shim used to obscure.