PostgreSQL: what runs differently¶

Where SQLite or ClickHouse compute something their own way, hare leaves their result as it is — see SQLite: what runs differently and ClickHouse: what runs differently. What differs on PostgreSQL is what the server itself does — how it ends a transaction after an error, how long a name may be, what each server version adds. Each row below says what hare does about it.

Transactions¶

PostgreSQL	hare
A failed statement aborts the transaction	Leaving the block without an exception rolls it back and raises `TransactionManagementError`; a nested `atomic()` around the statement keeps the outer transaction usable — see A failed statement in a transaction.
`CREATE INDEX CONCURRENTLY` outside a transaction only	`AddIndex(concurrently=True)` needs `atomic = False` on its migration — see Indexes.
Waiting for a lock without a limit	`migrate(lock_timeout=...)` bounds how long a migration waits for a table's lock — see Lock timeout.

Values and names¶

PostgreSQL	hare
A session time zone of the server's setting	Both drivers run every session in UTC — see The session time zone.
Names cut to 63 bytes	A name hare generates is kept within 63 bytes, a longer one shortened with a digest; an explicit name over 63 bytes raises `ConfigurationError` — see `table`.
`NULL` placement in an ordering	`NULL` sorts as the largest value, last ascending — see NULL ordering.
Text order of the database's collation	Text compares and sorts by the collation of the database; `Collate(...)` asks for another — see Database functions.
Integer and decimal ranges of a column type	The column refuses a value out of its range itself.

Server versions¶

hare reads the server's version when it connects and turns on what the version has; on an older server the feature is refused — see the features table.

Version	Adds
15	`merge()`; `UniqueConstraint(nulls_distinct=False)`
16	`explain(generic_plan=True)`
17	`merge().returning()` and `when_not_matched_by_source()`; `JsonTable`; an `ExclusionConstraint` on a partitioned table; `explain(memory=True)`, `explain(serialize=True)`
18	A `VIRTUAL` `GeneratedField` (`stored=False`); a `UuidV7()` `db_default`; `without_overlaps=True`; `returning(old=...)`

What an older server does itself, hare leaves to it:

PostgreSQL before 18 refuses to cast a JSON null of a JSONField path to another type (Cast(F("data__n"), IntField()) — cannot cast jsonb null to type integer); PostgreSQL 18 gives NULL.
PostgreSQL before 18 doesn't tell a transaction pooler its search path — the schema option and a schema per tenant behind one are refused (see Connection poolers).

Schema and migrations¶

PostgreSQL	hare
Extensions for vectors, geodata, trigrams and some types	`Meta.extensions` declares them, and the migration autodetector writes `CreateExtension` for each — see Extensions.
Partitioned tables, index types	See Indexes and tables.
Schemas	`Meta.schema`, a schema per tenant and row level security work as written — see Schema per tenant.

Drivers¶

rust_pg and asyncpg give the same results through hare's API; raw SQL differs in how the two bind and read values — see Raw SQL on the two drivers.