resharding::cancelWhenAnyErrorThenQuiesce() uses whenAllSucceed() from future_util.h in combination with whenAll() to wait on all of the data replication components exiting. The whenAllSucceed().onError() pattern is unreliable for this because the onError() lambda won't run when executor (which is the scoped executor here) has already been shut down. The kExecutorShutdownStatus error is propagated back to RecipientStateMachine through _dataReplicationQuiesced and consumed by the onCompletion() which is running on the cleanup executor.

Since the ReshardingOplogFetcher runs on the ReshardingDataReplication::_oplogFetcherExecutor and the whenAll() was skipped, a task from it may still be running at shutdown after the RecipientStateMachine and thus the ReshardingDataReplication has been destroyed. A solution here would be to have RecipientStateMachine::_runMandatoryCleanup() join the ReshardingDataReplication::_oplogFetcherExecutor.

ExecutorFuture<void> cancelWhenAnyErrorThenQuiesce(
    const std::vector<SharedSemiFuture<void>>& futures,
    ExecutorPtr executor,
    CancellationSource cancelSource) {
    return whenAllSucceedOn(futures, executor)
        .onError([futures, executor, cancelSource](Status originalError) mutable {
            cancelSource.cancel();

            return whenAll(thenRunAllOn(futures, executor))
                .ignoreValue()
                .thenRunOn(executor)
                .onCompletion([originalError](auto) { return originalError; });
        });
}