Every database statement, one function each, as raw SQL.

All functions take the repo explicitly. Step outcomes commit through the batched flush/1 (group commit — see GenDurable.Flusher): one guarded UPDATE … FROM unnest(...) for the whole batch plus the consume / parent-join / await-recheck / children-insert riders, in one transaction. concurrency_key serialization is enforced by the database — the UNIQUE partial index gen_durable_concurrency_active makes a second executing row per key uncommittable, so the pick's claim IS the lock (held exactly for the step window, released by any outcome or the reaper); no advisory locks, no pinned connections.

Every statement has a static SQL text and goes through the connection-level prepared-statement cache (cache_statement:), so Postgres parses and plans it once per connection instead of on every call — bulk inserts pass rows as parallel arrays via unnest to keep the text static (and the parameter count fixed, clear of the 65535-parameter protocol cap). The one exception is upsert_rate_configs (dynamic VALUES, boot-time only). Hosts behind a transaction-pooling proxy set prepare: :unnamed on the repo to bypass the cache.

JSON values (state, result, signal payloads) arrive here as encoded JSON text and are bound as TEXT parameters cast server-side ($n::text::jsonb). A bare $n::jsonb parameter would make the driver JSON-encode the already-encoded string, storing a double-encoded jsonb scalar instead of an object — invisible to ->>/jsonb indexes. Rows written by versions that did exactly that still decode fine: the read paths accept both formats.