Skip to main content

​
Overview

Firestore uses gRPC with a method-name classifier. Delta tracking works at the document path level — collections serve as “tables” and full document paths serve as “primary keys”. The shadow is a Firestore emulator seeded with production data. Applications connect using the standard Firestore SDK by setting FIRESTORE_EMULATOR_HOST=127.0.0.1:<mori-port>. Connection Parameters:

​
Differences from PostgreSQL

Firestore is a document database — fundamentally different from SQL. The copy-on-write model is adapted:
  • Document-level delta tracking — Uses (collection, fullRelativePath) pairs. Full document paths prevent subcollection key collisions in collection group queries.
  • Shadow seeded with data — Unlike SQL engines (schema-only shadow), the Firestore shadow is seeded with production documents including subcollections up to depth 3.
  • Merged reads by document path — GetDocument checks tombstone/delta per-document. RunQuery queries both backends, merges with shadow-wins semantics, re-evaluates WHERE filters, re-sorts, and applies LIMIT. ListDocuments over-fetches from prod and merges. BatchGetDocuments routes per-document.
  • Aggregation — RunAggregationQuery decomposes to a base RunQuery through the merged pipeline, then re-computes COUNT/SUM/AVG in memory.
  • Transaction snapshot isolation — BeginTransaction captures a ReadTime timestamp. All subsequent prod-side reads within the transaction use this timestamp for consistent point-in-time snapshots.
  • Hydration via GetDocument/UpdateDocument — Fetches the full document from prod and writes to shadow. Falls back to Commit for documents that don’t yet exist in shadow.
  • Document Transforms — serverTimestamp, increment, arrayUnion, arrayRemove transforms are tracked as deltas.
  • Precondition evaluation — Checks exists and update_time preconditions against the correct backend (tombstone, delta, or prod).
  • Listen (real-time) — Forwarded to prod only. Real-time listeners do not see shadow mutations.
  • No schema/DDL — Firestore is schemaless. No DDL tracking, no FK enforcement, no sequence offsets.
  • Collection group queries — Fully supported with full document path keys preventing cross-subcollection key collisions.

​
Known Limitations

  • PartitionQuery returns prod-only cursor boundaries (not a correctness issue — cursors are metadata, not documents).
  • Listen (real-time listeners) only reflects prod state, not shadow mutations.
  • Firestore SDK-specific features (offline persistence, local cache) are not intercepted by the proxy.
  • If GOOGLE_APPLICATION_CREDENTIALS is set, you can omit the credentials file in the connection string.