This defect was introduced in 9e4d0610ea. I had thought that
this code was exercised in tests but it is bypassed by the use of
grpc_test.beta.test_utilities.not_really_secure_channel. :-(
(1) Move dependency on protobuf from grpcio to grpcio_test. While the
most-commonly-foreseen use case of grpcio makes use of protobuf,
technically protobuf is not strictly needed and there's no actual
in-code relationship between grpcio and protobuf.
(2) Loosen the dependency on protobuf from ==3.0.0a3 to >=3.0.0a3.
(3) Update all references to 0.10.0* to 0.11.0.
(4) Alphabetize the grpcio_test dependencies.
Servers and stubs were context managers in the Alpha API; they may not
need to be in the Beta API but it's easy enough to do, eases migration,
and probably helps some use cases.
For now the stub is given empty __enter__ and __exit__ methods; we can
always come back and implement the actual work of context management in
a later change.
(1) Renamed the "beta" module "implementations" because it hasn't been
monolithic since "interfaces" was factored out of it a few changes
back.
(2) Moved ChannelConnectivity from grpc.beta.beta to
grpc.beta.interfaces since it is constants that don't depend on the
beta implementation.
(3) Moved the Server interface definition from grpc.beta.beta to
grpc.beta.interfaces since it is an interface.
(4) Dropped the "create_" prefix from "create_<...>_channel" functions
to better match the other creation functions throughout the codebase.
It may seem weird that code and details would travel along two paths
now instead of one but it makes sense after considering that sometimes
the code and details are application data from the remote application
and sometimes they are transport data from the transport between the
local and remote applications.
(1) In _ingestion, it's the "details" attribute of a
NoSuchMethodException that we want. The "message" is inherited from the
base Exception class.
(2) In _transmission, use a proper sum type for representing operation
abortion. Trying to overload the existing _completion value for
status-and-details-when-aborting was trying to be too clever.
(3) In _calls... oof. Just look. Oof. Test coverage for this code path
is added.
(4) In _service, the application-provided
face.MultiMethodImplementation isn't directly callable, but rather
exposes a method named "service".
(5) In crust.implementations, the wrapping that we've put around the
application-provided face.MultiMethodImplementation *is* directly
callable, and *does not* expose a method named "service".
(6) Also in crust.implementations, base.NoSuchMethodError's constructor
takes a code value and a details value.
(7) Again in crust.implementations, the application-provided
face.MultiMethodImplementation may be None, and if it is None, we
shouldn't wrap it with an adaptation function that would only raise a
TypeError at a later time.
(1) Plumb the metadata transformer given at the Beta API through to the
InvocationLink where it will be used.
(2) In both InvocationLink and ServiceLink, if there isn't a registered
serializer or deserializer, just pass the payload through rather than
ignoring the entire RPC.
The invoker is an object derived from, and referring to, objects of the
Face implementation under test. If those objects are to be garbage
collected at the appropriate time the invoker that references them must
be made eligible for garbage collection in the test's tearDown method.
Tickets should not be ignored if the end is in a grace period; rather
they should be ignored if they are for an unrecognized (likely new)
operation and the end is in a grace period.
(1) Move metadata and details constants for gRPC-on-the-wire tests into
grpc.test_common.
(2) Drop definitions of setUpModule and tearDownModule from a unit test
module that, because it uses the load_tests protocol, never had those
methods called anyway. :-(
(1) Call "cancel" on each future, not on the list of futures.
(2) If and when futures mature their actions should simply abort all
outstanding operations and cancel any other futures. They should not
shut down the _End's internal thread pool; only the termination action
of the last operation to terminate should shut down the pool (in the
case of their having been active operations at the time at which the
_End's stop(grace) method was called).
- Removing service_accounts credentials. These credentials just have
drawbacks compared to service_account_jwt_access credentials, notably
in terms for security.
- Renaming Google specific credentials with a Google prefix for C and
C++. This should be done as well for wrapped languages.
(1) In grpc._links.service._Kernel.add_ticket, premetadata() the call
if it has not already been premetadataed for any non-None ticket
termination, not just links.Ticket.Termination.COMPLETION.
(2) In grpc.framework.core._reception, add an entry to
_REMOTE_TICKET_TERMINATION_TO_LOCAL_OUTCOME for REMOTE_FAILURE.
REMOTE_FAILURE on a received ticket indicates the remote side of the
operation blaming the local side for operation abortion.
(3) In grpc.framework.core._reception.ReceptionManager._abort, only
abort the operation's other managers if the operation has not already
terminated, as indicated by the "outcome" attribute of the
TerminationManager.
(4) In grpc.framework.core._reception.ReceptionManager._abort, don't
transmit the outcome to the other side of the operation. Either it came
from the other side in the first place and to send it back would be
telling the other side something it already knows, or it arose from a
network failure and there's no confidence that it would reach the other
side.
Jan Tattermusch
Masood Malekghassemi
Nathaniel Manista
Stanley Cheung
Julien Boeuf