You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Define and Implement Input Freshness and Overload Policies #764
Define and implement explicit policies for stale or missing inputs, bounded buffers, and processing overruns in the Awkernel localization pipeline. Ensure that abnormal conditions produce observable and predictable behavior rather than unbounded backlog or silently invalid estimates.
Scope:
Define freshness limits and missing-input timeouts for each input stream
Specify buffer capacities and overflow behavior, such as dropping oldest or newest messages
Define component-specific behavior when required data is unavailable, including bounded waiting, skipping updates, or explicitly flagged fallback
Define handling of deadline overruns, overlapping activations, and late results
Specify recovery conditions and estimator-reset behavior where necessary
Report policy decisions through structured status events and counters
Acceptance criteria:
A policy matrix documents each condition, detection rule, threshold, action, output-validity status, and recovery condition.
Waiting periods and buffer capacities are bounded and configurable.
Component-specific requirements are respected; motion history is not discarded while still required for an admitted scan without an explicit failure outcome.
Stale or missing data is not silently reused as a fresh measurement, and prediction-only outputs are distinguishable from measurement-updated outputs.
Deadline overruns do not create uncontrolled overlapping executions or unlimited catch-up work.
Late or cancelled computations cannot overwrite newer state or access released buffers; cancellation is used only where supported safely.
Fault-injection tests verify the expected behavior for stale inputs, missing messages, buffer overflow, processing overruns, and recovery.
Logs distinguish input rejection, queue overflow, skipped updates, deadline misses, and discarded late results.
Any deviations from the selected Autoware reference behavior are documented.
Out of scope:
Fault-injection infrastructure, new scheduling algorithms, vehicle-level emergency-stop behavior, and formal safety certification.
Activity