Repository navigation
v1.15.0 — transaction functions everywhere; DuckDB correctness fix + connection pool
mentat 1.15.0: transaction functions everywhere, a DuckDB correctness fix, and a DuckDB connection pool
DuckDB users: upgrade
In 1.11.0–1.14.0 the DuckDB extension could silently return no rows. This
happened once a mentat database had been opened by a plain duckdb CLI (or any
client without the extension) after a mentat session. From then on, every
[?e :attr _] query came back empty, and retractions failed with
FATAL ... Failed to delete all rows from index.
The cause is a DuckDB 1.4.0–1.5.6 bug,
duckdb/duckdb#26828: the
first checkpoint after a WAL replay loses rows from secondary indexes. The
upstream fix (#26814) isn't released yet.
1.15.0 stores have no secondary indexes. DuckDB 1.5.6 never used them, and
inserts are 8.5x faster without them. Opening an existing store with 1.15.0
drops the old indexes, which also repairs a store that was already hit. We
verified this on a store that the 1.14.0 extension had corrupted.
New
-
Transaction functions on the embedded SQLite store and the DuckDB
extension (mozilla#333). The:db/fnfunctions that 1.14.0 added to PostgreSQL
now run on every backend that has the scripting interpreter, with the same
semantics:- the embedded store with the
minofeature, the SQLite extension and the
CLI; - the DuckDB extension.
[{:db/ident :person/birthday :db/fn "(fn [db who] (let [a (ffirst (mentat.store/q db '[:find ?a :in $ ?n :where [?e :person/name ?n] [?e :person/age ?a]] who))] [[:db/cas [:person/name who] :person/age a (inc a)]]))"}] [[:person/birthday "Alice"]]The function sees the database as of the start of the transaction and runs
in theedn_evalsandbox. Its tx-data is spliced into the same transaction.
MENTAT_TXFN_MAX_DEPTH(default 16) bounds recursion; on PostgreSQL it is
mentat.txfn_max_depth. Existing stores need no migration. - the embedded store with the
Faster (DuckDB)
-
Concurrent mentat calls now run in parallel. The extension opens a pool
of connections when it is loaded (MENTAT_DUCKDB_POOL, default one per core,
at most 8). Writes to one store still take turns; reads, and writes to
different stores, run concurrently.- Point lookups at 8 client threads: ~320 → ~1500 ops/s.
- Writes from 8 threads, one store each: 65 → 291 tx/s.
- Single-client latency is unchanged.
Benchmarks:
benchmarks/results/duckdb-perf-2026-10-15T010700Z/findings.md.
Other
- The Quack listen backlog (5) is reported upstream as
duckdb/duckdb-quack#318. - sparsemap was evaluated and not adopted
(docs/design/sparsemap-evaluation.md).
Upgrading
- PostgreSQL:
ALTER EXTENSION pg_mentat UPDATE TO '1.15.0';is a no-op
edge, with no schema change. - Embedded / crates.io:
mentatdb = "1.15.0". - DuckDB:
FORCE INSTALL mentat FROM community; LOAD mentat;once the
registry update merges. Until then, use thementat-duckdb-v1.5.6-linux_amd64
asset attached to this release, loaded with-unsigned.
Full details: CHANGELOG.