Arvados

I suspect the "write log, then save log collection" is doing the opposite of what we want, because

• "write log entry" just writes the message to the throttled-logging buffer, not yet to the log collection
• "save log collection" saves the log collection and resets the auto-flush timer, minimizing the chance auto-flush will happen before the preempted instance shuts down

It's possible something else is going on too, but either way, we should rearrange the logging pipeline so the log collection gets updated immediately instead of after the "group logs into chunks" step. If nothing else, that will reduce latency for showing logs in workbench.