Summary
In the azure-functions-durable (V2) package, the parameter that receives the
orchestration context is effectively hardcoded to the name context. As a
consequence:
- A user-supplied
context_name other than "context" on
@app.orchestration_trigger(...) does not work — the host cannot bind the
trigger to the generated handle.
- There is no way for an orchestrator to also receive an
azure.functions.Context (func.Context) parameter, because the Functions
host injects func.Context only into a parameter named exactly context
that is not a trigger binding — and that name is always consumed by the
orchestrationTrigger binding.
This matches V1 behavior (see "Not a regression" below), so it is not a
regression introduced by V2. Filing this to track it as a potential V2
improvement to be considered after the Functions PR merges and the beta
release ships.
Mechanism / relevant code paths
The Azure Functions host binds a trigger to a function parameter by name,
and it inspects the generated handle (not the user's function).
-
The trigger binding is registered with name=context_name:
azure/durable_functions/decorators/durable_app.py —
orchestration_trigger(context_name, ...) calls
OrchestrationTrigger(name=context_name, ...).
-
The generated handle that the host actually indexes hardcodes a parameter
literally named context:
azure/durable_functions/orchestrator.py — Orchestrator.create:
def handle(context: func.OrchestrationContext) -> str:
return Orchestrator(fn).handle(context)
- The parameter must be annotated
azure.functions.OrchestrationContext
for the host's orchestrationTrigger binding converter to accept it.
-
Because the host matches the binding name (context_name) to a parameter
of handle by name, they only line up when context_name == "context".
Any other context_name leaves the trigger unbound.
-
func.Context injection: the Python worker injects func.Context only into
a parameter named exactly context that is not itself a trigger binding.
Since the orchestrationTrigger always occupies the context name, there is
no free context parameter for func.Context to be injected into.
-
Separately, durabletask's executor invokes the user's orchestrator as
fn(ctx, input) — there is no slot in that calling convention for a
func.Context object to be passed through to user code.
Not a regression (V1 parity)
V1 has the same structural constraint. In
azure-functions-durable-python's azure/durable_functions/orchestrator.py,
Orchestrator.create also registers a single-parameter
def handle(context: func.OrchestrationContext), so a V1 orchestrator could
never receive a working func.Context parameter either. V2 preserves this
behavior.
Related parity gap: function_context contents
The V1-compat DurableOrchestrationContext adapter currently exposes an empty
function_context. V1's function_context is not necessarily empty: in
azure/durable_functions/models/DurableOrchestrationContext.py, __init__
consumes a fixed set of named fields (history, instanceId, isReplaying,
parentInstanceId, input, upperSchemaVersion, maximumShortTimerDuration,
longRunningTimerIntervalDuration, upperSchemaVersionNew) and funnels every
remaining field of the orchestration-trigger JSON into
FunctionContext(**kwargs). As of today the WebJobs extension injects at least
one such field — defaultHttpAsyncRequestSleepTimeMillseconds (observed value
30000, the durable-HTTP async polling interval) — so a live V1
function_context carries that value.
V2 receives the orchestration input via a protobuf trigger that does not carry
these arbitrary extra fields, so the compat adapter has no source to populate
function_context from. This is a minor parity gap to consider alongside the
context-name / func.Context work above.
Notes
- Out of scope for the current parity-focused Functions PR; capturing here so
it can be picked up as a V2 improvement after the beta release.
Summary
In the
azure-functions-durable(V2) package, the parameter that receives theorchestration context is effectively hardcoded to the name
context. As aconsequence:
context_nameother than"context"on@app.orchestration_trigger(...)does not work — the host cannot bind thetrigger to the generated handle.
azure.functions.Context(func.Context) parameter, because the Functionshost injects
func.Contextonly into a parameter named exactlycontextthat is not a trigger binding — and that name is always consumed by the
orchestrationTrigger binding.
This matches V1 behavior (see "Not a regression" below), so it is not a
regression introduced by V2. Filing this to track it as a potential V2
improvement to be considered after the Functions PR merges and the beta
release ships.
Mechanism / relevant code paths
The Azure Functions host binds a trigger to a function parameter by name,
and it inspects the generated handle (not the user's function).
The trigger binding is registered with
name=context_name:azure/durable_functions/decorators/durable_app.py—orchestration_trigger(context_name, ...)callsOrchestrationTrigger(name=context_name, ...).The generated handle that the host actually indexes hardcodes a parameter
literally named
context:azure/durable_functions/orchestrator.py—Orchestrator.create:azure.functions.OrchestrationContextfor the host's orchestrationTrigger binding converter to accept it.
Because the host matches the binding
name(context_name) to a parameterof
handleby name, they only line up whencontext_name == "context".Any other
context_nameleaves the trigger unbound.func.Contextinjection: the Python worker injectsfunc.Contextonly intoa parameter named exactly
contextthat is not itself a trigger binding.Since the orchestrationTrigger always occupies the
contextname, there isno free
contextparameter forfunc.Contextto be injected into.Separately, durabletask's executor invokes the user's orchestrator as
fn(ctx, input)— there is no slot in that calling convention for afunc.Contextobject to be passed through to user code.Not a regression (V1 parity)
V1 has the same structural constraint. In
azure-functions-durable-python'sazure/durable_functions/orchestrator.py,Orchestrator.createalso registers a single-parameterdef handle(context: func.OrchestrationContext), so a V1 orchestrator couldnever receive a working
func.Contextparameter either. V2 preserves thisbehavior.
Related parity gap:
function_contextcontentsThe V1-compat
DurableOrchestrationContextadapter currently exposes an emptyfunction_context. V1'sfunction_contextis not necessarily empty: inazure/durable_functions/models/DurableOrchestrationContext.py,__init__consumes a fixed set of named fields (
history,instanceId,isReplaying,parentInstanceId,input,upperSchemaVersion,maximumShortTimerDuration,longRunningTimerIntervalDuration,upperSchemaVersionNew) and funnels everyremaining field of the orchestration-trigger JSON into
FunctionContext(**kwargs). As of today the WebJobs extension injects at leastone such field —
defaultHttpAsyncRequestSleepTimeMillseconds(observed value30000, the durable-HTTP async polling interval) — so a live V1function_contextcarries that value.V2 receives the orchestration input via a protobuf trigger that does not carry
these arbitrary extra fields, so the compat adapter has no source to populate
function_contextfrom. This is a minor parity gap to consider alongside thecontext-name /
func.Contextwork above.Notes
it can be picked up as a V2 improvement after the beta release.