Optimize `executemany()` and `executescript()` to use batch requests

[penberg]

Both APIs are now implemented using interactive transactions. Batches will be more efficient.

[alixTal]

Any updates on this? executemany is very slow since it executes each query individually.

[vmnarayan-byte]

Reproduced this on libsql 0.1.11 (current PyPI release), Python 3.12, against a real Turso Cloud primary (not local-only) via the Embedded Replicas pattern (libsql.connect(path, sync_url=..., auth_token=...)). Adding concrete numbers since the last update here is a year old and it's still very much present.

Bulk-write benchmark, 1000 rows into one table, same connection/DB each time:

Method	Result
execute() in a loop, 1000 calls	187.04ms/row avg (187.99s total)
executemany(), 1000 rows	201.06ms/row avg (201.06s total) — no improvement over the loop
executescript(), 1000 statements in one call	374.95ms/row avg (374.95s total) — worse
Single INSERT ... VALUES (...), (...), ... (1000 rows, one execute() call)	3.879ms/row (3.88s total)

So the only way we found to get real throughput is to hand-construct one multi-row VALUES statement ourselves — executemany()/executescript() give no benefit over a naive per-row loop.

Root cause data, measured independently from inside the same network path (raw socket timing, not through the driver): raw TCP connect to the primary ~58ms, TCP+TLS connect ~200-220ms. That lines up almost exactly with the ~185-230ms-per-call cost we see from execute()/executemany()/executescript() — strongly suggesting each call pays close to a full connection/round-trip-level cost regardless of whether it's inside an already-open Connection/transaction, rather than being multiplexed over one connection.

Worth noting: we also found @tursodatabase/sync (the JS/TS push/pull CDC package, via tursodatabase/turso-sync-benchmark) publishes benchmarks showing similar-magnitude wins over the legacy Embedded Replicas pattern for exactly this kind of workload — so there's prior art within the org for what a fix here might look like, if that's useful context.

Happy to share the harness script if it helps reproduce.