index-keyspace-evolution

Index Keyspace Evolution

Use this when adding a new Fjall keyspace, extending an existing key encoding, or changing the value shape of an indexed record.

Decide if you need a schema bump

Bumping INDEX_SCHEMA_VER triggers auto-wipe of .basemind/views/<view>/index.fjall/ on the user's next scan. The next scan rebuilds from the msgpack blobs — no data loss, but a one-time scan cost. Bump when:

• The byte layout of an existing key changes.
• A SymbolKind ordinal would shift (you reordered the enum — don't).
• A value's msgpack shape lost a field or renamed one.

Do not bump when:

• Adding a brand-new partition (no existing entries to migrate).
• Adding a #[serde(default)]-defaulted field to an indexed value.
• Adding a new SymbolKind variant at the tail (the u8 mapping is stable).

Steps

1. Encoder + decoder + tests in src/index/keys.rs

   • Add <partition>_key(...) -> Vec<u8> and <partition>_prefix(...) -> Vec<u8>.
   • Add parse_<partition>_key(&[u8]) -> Result<(...), …>.
   • Add unit tests for round-trip + prefix isolation: insert two keys with a shared prefix that differ only in the suffix, scan the prefix, confirm the wrong sibling is excluded.

2. Open the partition in src/index/mod.rs

   • Add the field to IndexDb.
   • Open via db.keyspace("<name>", PartitionCreateOptions::default()) next to the others.
   • If used only for reads, mark #[allow(dead_code)] until a writer / reader exists.

3. Writer in src/index/writer.rs

   • Extend upsert_file to derive secondary keys from the L1/L2 input and stage inserts.
   • Extend the read-before-write deletion to enumerate the partition under the file's primary key and stage deletes for those secondary keys.
   • Add integration tests covering upsert-then-remove leaves the partition empty.

4. Reader in src/mcp/helpers.rs (and/or query.rs)

   • Add a scan_<partition>_by_<key> helper that range-scans the prefix and decodes values.
   • Apply scan_cap = limit * 8.