Page Life Expectancy Doesn’t Mean Jack, and You Should Stop Looking At It.

Page Life Expectancy is a Perfmon counter that’s supposed to track how long pages stay in memory before they get flushed out to make room for other pages that are needed instead.

Paul Randal has blogged this a few times, and rather than rehash ’em, I’d rather point you to a couple of his roundups:

• Knee-Jerk PerfMon Counters: Page Life Expectancy
• Page Life Expectancy isn’t what you think

And layer in a few of my own observations for context:

Page Life Expectancy goes up 1 for each secondthat you’re not under memory pressure. Restart your SQL Server instance and watch PLE: it starts at 1, and goes up by 1 for each second of uptime. 5 minutes of uptime = PLE 300. For the first 5 minutes, it’s not like your server’s under memory pressure – it just woke up, for crying out loud. Give him a minute. Well, 15-20, I guess, because PLE is near useless during that time span.

Page Life Expectancy drops can be triggered by confusing operations.By default, any one running query can get a memory grant the size of 25% of your buffer pool. Run a few of those queries at the same time, and your buffer pool gets drained – but PLE doesn’t necessarily drop. However, the instant an unrelated query runs and needs to get data that isn’t cached in RAM, your PLE will drop catastrophically. Which queries are at fault? The queries getting large grants, or the queries doing reads? “Yes.”

Page Life Expectancy is a lagging indicator.Lagging indicators are something that tell you about an emergency long after the emergency has happened, and lagging indicators don’t recover quickly after the emergency is over. When you combine the above two problems – PLE only rising 1 per second, and PLE dropping at times that aren’t necessarily tied to the dropping buffer pool – then if you’re alerting based on low PLE numbers, you could have already missed the emergency. When you go log in and look for what long-running queries are running, it’s already too late. Instead, you should be using leading indicators: things that tell you a problem is coming up.

With that stuff in mind, I’ve removed the Page Life Expectancy warnings fromsp_BlitzFirst altogether. I think they’re distracting y’all from the real leading indicator, wait stats – and of course, sp_BlitzFirst shows that, so I’d rather focus your attention there.