Every actor now registers (and unregisters) with the arbiter at
startup/teardown. For now the registry is stored in a plain `dict` in
the arbiter's memory. This makes it possible to easily coordinate actors
started as plain Python processes or via `multiprocessing`.
A whole smörgåsbord of changes was required to accomplish this:
- factor handshake steps into a func
- track *every* channel connected to an actor including multiples to the
same remote peer (may want to optimize this later)
- handle `trio.ClosedStreamError` gracefully in the message loop
- add an `open_portal` asynccontextmanager which handles channel
creation, handshaking, and spawning a bg task for msg processing
- add a `start_actor()` for starting in-process actors directly
- add working `get_arbiter()` and `find_actor()` public routines
- `_main` now tries an anonymous channel connect to the stated
arbiter sockaddr and uses that to determine whether to crown itself
Fail gracefully (by "aborting") the same way `trio` does. This avoids
ugly sub-proc tracebacks thrown at the console. Unset the local actor
when `tractor._main` completes. Cancel all tasks for a peer channel on
disconnect.
Drop all channel/connection handling from the core and break up all the
start up steps into compact and useful functions. The main difference is
the daemon now only needs to worry about spawning per broker streaming
tasks and handling symbol list subscription requests.
When an error is raised inside a nursery block (in the local actor)
cancel all spawned children and ensure the error is properly
unsuppressed.
Also,
- change `invoke_cmd` to `send_cmd` and expect the caller to use
`result_from_q` (avoids implicit blocking for responses that might
never arrive)
- `nursery.start()` the channel server block such that we wait for the
underlying listener to spawn before making outbound connections
- cancel the channel server when an actor's main task completes
(given that `outlive_main == False`)
- raise subactor errors directly in the local actors's msg loop
- enforce that `treat_as_gen` async functions respond with a caller id
(`cid`) in each yield packet
Command requests are sent out and responses are handled in a "message
loop" where each command is associated with a "caller id" and multiple
cmds and results are multiplexed on the came inter-actor channel. When
a cmd result arrives it is pushed into a local queue and delivered to the
appropriate calling actor's task. Errors from a remote actor are always shipped
in an "error" packet back to their spawning-parent actor such that any error
in a subactor is always raised directly in the parent. Based on the
first response to a cmd (either a 'return' or 'yield' packet) the caller
side portal will retrieve values by wrapping the local response queue in
either of an async function or generator as appropriate.
Tyler Goodlet
goodboy