checkpoint_uri
def checkpoint_uri() -> StringDISTA_CHECKPOINT_URI, without a trailing slash; raises if unset.
dista.checkpoint
Where actor checkpoints live: objects under DISTA_CHECKPOINT_URI.
The URI is a local directory, a file:// one, or anything objectstore
writes to, such as s3://bucket/prefix. A local directory only outlives the
worker that wrote it if the node that restarts the actor sees it too, so a
cluster needs a shared volume or object storage.
Each checkpoint is its own object, named for its actor and the number of
calls it covers, <uri>/<actor ID in hex>/<calls>, so a new one never
overwrites the one the head would restore from. The worker hosting the
actor deletes the ones before that one, those its predecessor left
included. A head that restarts restores each actor from its newest.
A checkpoint's bytes are, for each caller whose calls it includes, that caller's number and the newest call's object ID, then the actor's state: count u64, that many pairs of u64s, the state through the codec. Calls resent after a head restart that it already includes aren't run again.
Functions
checkpoint_location
def checkpoint_location(actor: UInt64, calls: Int) -> StringWhere actor actor's checkpoint after calls calls lives.
newest_checkpoint
def newest_checkpoint(actor: UInt64) -> IntThe calls covered by actor actor's newest checkpoint, or -1.
list_checkpoints
def list_checkpoints(actor: UInt64) -> List[Int]The calls covered by each checkpoint actor actor has, in any order.
write_checkpoint
def write_checkpoint(location: String, data: Span[UInt8])Writes a whole checkpoint, never part of one: after a head restart the newest is restored without anyone having said it was finished. An object store's put is whole already; a local file is written under a temporary name and renamed into place.
read_checkpoint
def read_checkpoint(location: String) -> List[UInt8]delete_checkpoint
def delete_checkpoint(location: String)Deletes a checkpoint nothing will restore from; a failure only leaves it behind.