### Summary A namespace enrolled in the Knowledge Graph can be permanently skipped by the SDLC dispatcher: its row in the replicated enabled-namespaces table gets `traversal_path = '0/'` baked in at insert time, and every dispatch filter (including the hourly sweep) drops that value forever. ### Expected behavior Enrolling a top-level namespace starts indexing it within a tick or two, regardless of how Siphon happened to order its table writes. ### Actual behavior If Siphon's ClickHouse consumer inserts the enrollment row before the namespace row is visible to the traversal-path dictionary, the row's `traversal_path` DEFAULT evaluates to the `0/` fallback and is stored permanently. The dispatcher's enabled-set queries filter that value out, so the namespace never dispatches and none of its entities are indexed. Nothing self-heals on GDK, Dedicated, or fresh installs; GitLab.com only recovers because later re-snapshot/refresh traffic rewrites the rows. There is a second, related gap: the enrollment CDC fast path reads `traversal_path` from the wire event, but the Postgres table has no such column, so every real enrollment event is skipped with a warning and newly enrolled namespaces wait for the hourly sweep at best. ### Steps to reproduce 1. Bring up GDK with Siphon stopped. 2. Create a root group with a project and some issues, and insert its enrollment row into `knowledge_graph_enabled_namespaces` — all in Postgres, while Siphon is off. 3. Start Siphon and Orbit at the same time. 4. Roughly half the runs, the enrollment row lands in ClickHouse with `traversal_path = '0/'` (the consumer writes tables in parallel workers, so the enrollment insert can beat the namespaces insert by a few milliseconds). The dispatcher's cold-start sweep and every later sweep skip the namespace; no graph rows ever appear for it. ### Environment - [x] **Orbit Remote** — the hosted graph on GitLab.com (`glab orbit remote`, MCP, GitLab Duo, or the REST API) Reproduced on GDK against orbit main (v0.94, schema v88), 2026-08-07. The same insert-time DEFAULT exists in production (`db/click_house/migrate/main/20260423130900_recreate_siphon_knowledge_graph_enabled_namespaces.rb`). ### Logs <details> <summary>Repro evidence (GDK, 2026-08-07)</summary> Two identical runs, sub-100ms apart in consumer insert order, opposite outcomes: ``` run A: enabled row inserted 14:15:47.258, namespaces 14:15:47.316 → stored 1/229/ (indexed) run B: enabled row inserted 18:18:05.913, namespaces 18:18:05.958 → stored 0/ (never indexed) ``` Poisoned state after run B — cold sweep dispatches 4 of 5 enabled namespaces: ``` {"message":"dispatched namespace indexing requests","dispatched":4,"changed_namespaces":4,"sweeping":true} ``` Enrollment CDC fast path skipping every real event (wire event has no traversal_path column): ``` {"level":"WARN","message":"CDC event missing traversal_path; skipping (re-tries next tick via active backfill)","root_namespace_id":231} ``` Production check (gitlab-com-production datalake): 117 historical `0/` rows in `siphon_knowledge_graph_enabled_namespaces`; all currently superseded by later corrective writes (tombstone + reinsert pairs, the refresh/reconciler signature), so no .com namespace is stuck today — but the race demonstrably fires there, and installs without that corrective traffic stay broken. The current sweep query also returns 106 namespaces whose latest state is a tombstone (unenrolled but still dispatched every sweep, until the ReplacingMergeTree merge catches up). </details> <details> <summary><b>Agent context</b> — extended analysis</summary> Root cause chain: - `siphon_knowledge_graph_enabled_namespaces.traversal_path` is `DEFAULT multiIf(..., dictGetOrDefault('namespace_traversal_paths_dict', 'traversal_path', root_namespace_id, '0/'), '0/')`, evaluated at insert time. Postgres has no traversal_path column, so the DEFAULT always runs. - Siphon's ClickHouse consumer processes tables in parallel stream workers; cross-table commit order is not preserved. WAL order does not save you. - `namespace_traversal_paths_dict` is `LAYOUT(CACHE)` with `LIFETIME(60,300)`; a miss is negatively cached, extending the race window beyond the insert ordering itself. - Orbit dispatcher filters: the change-detection CTE requires `traversal_path != ''` and joins on the root path; the sweep reader requires `match(traversal_path, '^[0-9]+/[0-9]+/')`. `0/` fails both. The enabled table is deliberately not a change source, so the row's arrival never triggers anything by itself. - The enrollment CDC route (`EnabledNamespacesRoute`) reads `traversal_path` from event columns that only exist in test fixtures; real Siphon events never carry it, so the "dispatch on enrollment" fast path never fires in practice. Fix directions (implemented in the linked MR): re-resolve the path through the dictionary at query time everywhere the enabled table is read (stored value as fallback, argMax dedup per enrollment row), and have the CDC route resolve missing paths from `namespace_traversal_paths` (the table, not the dictionary, to dodge negative caching). Known remaining readers that still trust the stored path (candidate follow-ups): `crates/indexer/src/modules/namespace_deletion/store.rs` and `crates/indexer/src/orchestrator/scheduled/migration_completion.rs`. Separately, worth asking the Siphon/CDC owners whether the default-value reconciler covers this table on Dedicated. </details> <!-- AI-Sessions dir: ~/.claude/projects/-Users-jean-gabriel-doyon-Work-orbit/ 06a4770e-59b8-44fc-82d0-32c74449666f.jsonl (2026-08-07) -->