[disq]

Checkpoints flushed while a resource is being resolved, for example from pre_resource_resolve or post_resource_resolve, now wait for that resource and its relations. Before, the scheduler ran a resource's hooks before putting the resource on the queue, so the CLI could save a cursor past rows it never received.

Also:

• Flushing checkpoints outside their sync now raises SyncNotRunningError, a RuntimeError exported from cloudquery.sdk.stateclient.
• Without checkpoints, setting a key to its current value no longer writes it to the backend again. With checkpoints it still sends a checkpoint, because later values of the key must follow the caller's resources.
• New tests check that checkpoints follow grandchild relations.
• The scheduler's thread pools keep their own max_workers instead of reading ThreadPoolExecutor's private copy, and a comment explains why the queue limit is a semaphore taken before submit().

[erezrokah]

Please address the checkpoint-ordering regression described in the inline comment.

[erezrokah]

[P1] Preserve checkpoint dependencies for unchanged cursor values

After flush() clears self.changes, this early return drops the caller’s scope entirely. For example, resolver A flushes cursor=1; resolver B yields a parent with an unfinished child, then sets and flushes the same cursor; resolver C subsequently flushes cursor=2. B’s flush now schedules no dependency, allowing checkpoint 2 onto the stream before B’s child. An interrupted sync can therefore persist a cursor past rows it never received. Please preserve resource dependencies for unchanged writes in checkpoint mode, even when suppressing redundant state writes.

[disq]

Good catch, fixed in 2346dea. set_key now skips an unchanged value only when it writes to the backend directly. With checkpoints, an unchanged value still records the caller's scope, so B's flush sends a checkpoint that waits for B's child, and checkpoint 2 waits for it in turn. test_unchanged_values_order_later_checkpoints_after_the_callers_resources reproduces your A/B/C scenario and fails on the previous commit.