SLO Fo Sho (part two)

Huh uh. We're back. Let's get it Service health monitors set a limit for what's good And objectives set a threshold, somewhere in the neighborhood Careful! Don't get it twisted, gotta differentiate SLOs ain't monitors. There's a difference, don't conflate! Monitors watch your service for alarming signs Expiring certs, resource usage crossin' lines System health checks, traffic up or down low Triggered pagers light up, lettin' operators know But SLOs help us when we're lookin back Review the promise against historic facts The focus is user impact, to see the true cost Respond to lessons learned and error budgets lost We review every week, maybe shift priorities Don't wanna leave users reading through their warranties Define your SLOs separately with thought Renegotiate, don't change 'em on the spot Monitors can use SLO burn rate But know that signal comes too late User pain comes before budgets burn Find proactive signals to discern Right. So the core concern is customer experience Ditch the internal view, be more vicarious Adopt the user's definition of success Let their need for us help drive the process This axiom is foundational, so let it guide: If your objective's met, but they're dissatisfied Then your objectives are invalid, ill-conceived Re-evaluate what you believed Who depends? Which functions matter? What failure modes leave our name in tatters? We go hard for the service level, naming our objectives Don't be scared to dive deep into the customer perspective Imagine you're a shoppin' website selling stuff to users We could look up all the items that they see as they maneuver Then search our competition stores to check for better prices So think outside the box, that's where the real advice is But in most situations you'll find latency and error rates As categories of measured concerns always dominate First, error rates focus on the pain we inflict Server-side errors are failures that we must admit The 500s family, yes, you know we must include But also timeouts that fail to connect, when they conclude - The request badly. It's a failure we can fix Is it bugs or bad deploys, or maybe infra playing tricks Exclude the client errors, that's their domain: Bad creds, bad input, and junk sent in vain A 400 response? If it's correct It fulfills the contract. It gets respect Now, latency defines the experienced delay What users are living is what we must convey Count the milliseconds 'till an API call is through Or minutes 'til a cluster that is being built for you Sits fully provisioned, and ready for your task Whatever the time these operations ask We measure start to end, beyond our application Right up to the border of our user conversation From the gateway proxy, watching the transaction Or synthetic clients, simulating user action Pass a timestamp in a pipeline with a winding flow Measure when the job is finished, find out if it's slow Yo, so we zoomed in on some things that make our users frown When they're faced with errors or the speed they really need is down Latency and error rates -- we laid out the definition But just countin' up events may not give a clear rendition Gotta break into partitions, see the pain minute by minute Part Three is in the queue to show you how to really pin it Grab yourself some milk and cookies, we'll be back to get into it Learn what separates the rookies, you'll be dunkin' when you do it Word