In practise this doesn't matter too much, as most jobs are retryable. Just filing a bug for the record

In WMF's job queue, the logic that handles RunnableJob::allowsRetries() lives in MediaWiki\Extension\EventBus\JobExecutor. If the HTTP connection dies before this code is executed (for example, due to apache config or the timeout config in change-propogation), then change-propagation service will retry this job. This is pretty unexpected. If you really don't want your jobs to be retried, you have to set that in the change-propagation config as well as implementing allowsRetries()