An ORM costs time: validation, conversion, objects for every row. The question is how much. bench/overhead.ts in sqlite-zod-orm runs every operation twice on the same data, once through the library and once as hand-written bun:sqlite statements. It reports the library's speed as a percentage of the plain version, measured in the same run so that machine load cancels out.
| Operation (10,000 rows, default options) | 4.2.0 | 4.2.2 |
|---|---|---|
get(id) |
12% | 91% |
select().where() range |
9% | 77% |
.with() one level |
10% | 69% |
update(id) |
3% | 58% |
| read and update one row (a game tick) | not measured | 74% |
| change a field by assignment | 2% | 62% |
insert |
7% | 35% |
upsert |
4% | 10% |
| How the gains were made: |
- Zod parsers and per-table plans are built once, not on every call.
- Rows are decoded only for the columns that need it.
- Statements per update went from three to two.
- SQL text and query plans are cached per shape. One further round of changes did not measure faster: insert ×0.82 and upsert ×1.00, against run-to-run noise of about ±15%. It was not released.
Note:
upsertandinsertare still the slowest paths. If your app inserts tens of thousands of rows per second, batch them withinsertMany(61% of plain SQL) or write those statements directly withdb.raw. Next: chapter 9, limits.