Logs vs Structured Events

I got an interesting tweet the other day from @evntdrvn in response to this thread of mine. Paraphrasing,

“So I’ve almost got our group at work up to Step 1 in your observability maturity model, but some of the devs that I work with want to turn OFF our lovely structured logging in prod for informational-level msgs due to their legacy philosophy (‘we only log errors in prod’). The reasons given are mostly philosophical (“I’m a dev and only interested when things error out, I don’t want any other noise in prod logs”, “I don’t want to slow my app down in prod”). Help?!?”

As I was reading this, I was itching to fly out and dive into battle with Eric. I know exactly where his opinionated devs are coming from. I used to say the same things! I even wrote a whole blog post about it.

These developers have internalized a set of rules and best practices for dealing with output data, in the context of “monolith application development in the early 2000s”.

Monolithic systems assumptions

Those systems had many common constraints and assumptions, such as:

  • We have a monolith service, or a very small number of services. We can model the system in our heads.
  • Logging is done to local disk, which can impact performance
  • Disks are expensive
  • Log lines are spat out inline with execution.  A poorly placed printf can take the whole system down.
  • Investigation is rare, and usually means a human reading error logs.
  • Logging is of poor utility for understanding internal states or execution paths; you should just read the code or use a debugger.  (There are few or network hops between functions.)
  • Logging is mostly useful for detecting certain terminal crash states or connection errors.

Monolithic logging best practices


  • We should be very stingy in what we log
  • Debuggers should be used for understanding internal states of the code
  • Logs are a last resort and record of crash dumps.  We do not expect to use log data in the course of our daily work.  We assume log-related manual investigation will be infrequent and of limited utility.

These were exactly the right lessons to learn in the era of expensive hardware and monolithic repos/artifacts. Many people still work in environments like this, and follow logging best practices like these. God bless, more power to em.

Distributed systems assumptions

But more and more of us face systems that are very different.

  • We have many services, possibly many MANY services. A representative request will have “many” hops across “many” services and routers and proxies and meshes and storage systems.
  • We cannot model the system in our heads; it would be a mistake to try. We rely on tooling as the source of truth for those systems.
  • You may or may not have access to those services, or the systems your code runs on. There may or may not be a logging facility, or a centralized log aggregator. Your only view of the system is through the instrumentation of your code.
  • Disks and system resources are cheap, ephemeral, all but disposable.
  • Data services are similarly cheap.  We can almost entirely silo application performance off from the cost of writing perf data out.
  • Investigation is prohibitively slow and expensive for a human to do by hand. Many of the nodes or processes we need to inspect may no longer even exist, but their past states may still be relevant to us in understanding patterns to the present time.
  • Investigation should usually be done distributedly, across all instantiations of your code, however many there might be — and in real time
  • Investigation requires computation — not just string search. We need to ask on the fly involving math and percentiles and breakdowns and group by’s.  And we need access to the raw requests in order to run accurate computations — no pre-aggregates.
  • The hardest part isn’t usually debugging the code, it’s figuring out where is the code you need to debug. Or what the errors or outliers have in common from the perspective of the code.  Fixing the code itself is often comparatively trivial, once found.
  • What even is ‘logging’?
  • What even is ‘local disk’?

This isn’t optional: at some point of complexity or scale or distributedness, it becomes necessary if you want to work with these systems.

Logs can’t help you here.

And you aren’t going to get that kind of explorable data out of loglevel:ERROR, or by chopping up your telemetry into disconnected metrics devoid of context.

You are only going to get this kind of explorable, ad hoc, computation-friendly data if you take a radically new approach to how you output and aggregate telemetry.  You’re going to need to replace your log lines and log levels with a different sort of beast: arbitrarily wide structured events that describe the request and its context, one event per request per service.

If it helps, don’t think of them as log files any more. Think of them as events. Yes, you can stash this stream in a file, but why would you?  on what disk?  will that work for your serverless functions too?  Just stream them over the network to wherever you want to put them.

Log levels are another confusing and unnecessary artifact of yesteryear that you no longer really need. The more you think of structured events as logs, the more tempted you may be to apply the old set of best practices. So just don’t think of them as logs at all.

How to gather and structure your data

Instead of dribbling little pebbles of log effluvia throughout your code, do this.  (If you’re a honeycomb user, our beelines do it all automatically for you *and* pre-propagate the blobs with everything we know of your context.)

  1. Initialize an empty blob at the beginning, when the request first enters the service.
  2. Stuff any and all interesting detail about the request into that blob throughout the lifetime of the request.
    • Any unique id, any high-cardinality variable, any headers passed in, every full query, normalized query, and query execution time; every http call out to a remote service, every http execution time; any shopping cart id, first and last name, execution time — literally anything interesting, append to blob.
  3. Then, when the request is about to exit or error, write the blob off to honeycomb or another service or disk somewhere.

You can see immediately how this method has radically different performance implications and risks than the earlier shotgun spray approach. No more “oops i accidentally put a print line INSIDE a for loop”. The write amplification profile is compressed. Most importantly, the incremental cost of capturing more detail about the request per service is nearly zero.

And now you have the kind of structured data that you can feed into something like a columnar store, or honeycomb, and run ad hoc queries to your heart’s delight.

Distributed systems logging events best practices:

Let’s sum up.  (I’m including links to other past rants on this topic):

Just think.

No more doing multi-line regexps trying to look for the same request ID or user ID doing five suspicious things in a row.

No more regexps at all, for fuck’s sake.

No more bullshit percentiles that were computed at write time by averaging over a bunch of other averages

No more having to jump around from dashboards to logs trying to vainly eyeball correlate one spike with another. No more wondering why no two tools can agree if anything even exists or not

Just gather the detail you need to ask the questions when you need them, and store it in a single source of truth.  It’s that simple.

No need to shame people from learning best practices that worked perfectly well for a long time.  You can either let them learn the hard way that this transformation is non optional, or you can help them learn the easy way that it’s simply much better and easier to invest in this telemetry up front.  You seem like a nice enough chap, which is probably why you chose door 2.  (If you wanted to get tougher about it, have a few reformed folks in to tell their horror stories.  Try some ex-twitter engineers.)

The hardest part seems to be getting people to unlearn all the best practices they once learned for dealing with logs.  So just don’t call it logs anymore, if that helps. Call it “structured events”.

– charity.


Logs vs Structured Events

Operational Best Practices #serverless

This post is part two of my recap of last week’s terrific Serverless conference.  If you feel like getting bitchy with me about what serverless means or #NoOps or whatever, please refer back to the prequel post, where I talked about operations engineering in the modern world.

*Then* you can get bitchy with me.  (xoxoxxooxo)

The title of my talk was:

Screen Shot 2016-05-30 at 8.43.39 PM

The theme of my talk was basically: what should software engineers know and care about when it comes to operations in a world where we are outsourcing more and more core functionality?

If you care about running a quality service or product, or providing your customers with a reasonable level of service, you have to care about operational concerns like design, resiliency, instrumentation and debuggability.  No matter how many abstractions there are between you and the bare metal.

If you chose a provider, you do not get to just point your finger at them in the post mortem and say it’s their fault.  You chose them, it’s on you.  It’s tacky to blame the software or the service, and besides your customers don’t give a shit whose “fault” it is.

So given an infinite number of things to care about, where do you start?

What is your mission, and what are your differentiators?

The first question must always be: what is your mission?  Your mission is not writing software.  Your mission is delivering whatever it is your customers are paying you for, and you use software to get there.  (Code is kind of a liability so you should write as little of it as necessary.  hey!! sounds like a good argument for #serverless!)

Second: what are your core differentiators?  What are the things that you are doing that are unique, and difficult to replicate, or the things where you have to actually be world class experts in those things?

Those are the things that you will have the hardest time outsourcing, or that you should think about very carefully before outsourcing.

Screen Shot 2016-05-30 at 7.57.06 PM


You can outsource labor, but you can’t outsource caring.  And nobody but you is in the position to think about your core differentiators and your product in a holistic way.

If you’re a typical early startup, you’re probably using somewhere between 5 and 20 SaaS products to get rid of some of the crap work and offload it to dedicated teams who can do it better than you can, much more cheaply, so you are freed up to work on your core value proposition.


But you still have to think about things like reliability, your security model, your persistent storage models, your query performance, how all these lovely services talk to each other, how you’re going to debug them, how you’re going to repro when things go wrong, etc.  You still own these things, even if you don’t run them.

For example, take AWS Lambda.  It’s a pretty great service on many dimensions.  It’s an early version of the future.  It is also INCREDIBLY irritating and challenging to debug in a practically infinite number of insanity-inducing ways.

** Important side note — I’m talking about actual production systems.  Parse, Heroku, Lambda, etc are GREAT for prototyping and can take you a long, long way.  Early stage startups SHOULD optimize for agility and rapid developer iteration, not reliability.  Thx to @joeemison for reminding me that i left that out of the recap.

Screen Shot 2016-05-30 at 8.03.01 PM

Focus on the critical path

Your users don’t care if your internal jenkins builds are broken.  They don’t care about a whole lot of things that you have to care about … eventually.  They do care a lot if your product isn’t actually functional.  Which means you have to think through the behavioral and failure characteristics of the providers you’re relying on in any user visible fashion.

Ask lots of questions if you can.  (AWS often won’t tell you much, but smaller providers will.)  Find out as much as you can about their cotenancy model (shared hardware or isolation?), their typical performance variance (run your own tests, don’t trust their claims), and the underlying storage systems.

Think about how you can bake in resiliency from the user’s perspective, that doesn’t rely on provider guarantees.  If you’re on mobile, can you provide a reasonable offline experience?  Like Parse did a lot of magic here in the APIs, where it would back off and retry saves if there were any errors.

Can you fail over to another provider if one is down?  Is it even worth it at your company’s stage of maturity and engineering resources to invest in this?

How willing are you to be locked into a vendor or provider, and what is the story if you find yourself forced to switch?  Or if that service goes away, as so many, many, many of them have done and will do.  (RIP, parse.com.)

Screen Shot 2016-05-30 at 8.11.10 PM


Listen, outsourcing is awesome.  I do it as much as I can.  I’m literally helping build a service that provides outsourced metrics, I believe in this version of the future!  It’s basically the latest iteration of capitalism in a nutshell: increased complexity –> increased specialization –> you pay other people to do the job better than you –> everybody wins.

But there are tradeoffs, so let’s be real.

The service, if it is smart, will put strong constraints on how you are able to use it, so they are more likely to deliver on their reliability goals.  When users have flexibility and options it creates chaos and unreliability.  If the platform has to choose between your happiness vs thousands of other customers’ happiness, they will choose the many over the one every time — as they should.

Limits may mysteriously change or be invented as they are discovered, esp with fledgling services.  You may be desperate for a particular feature, but you can’t build it.  (This is why I went for Kafka over Kinesis.)

You need to think way more carefully and more deeply about visibility and introspection up front than you would if you were running your own services, because you have no ability to log in and use strace or gdb or tail a logfile or run any system profiling commands when things go dark.

In the best case, you’re giving up some control and quality in exchange for experts doing the work better than you could for cheaper (e.g. i’m never running a fucking physical data center again, jesus.  EC24lyfe).  In a common worse case, it’s less reliable than what you would build AND it’s also opaque AND you can’t tell if it’s down for you or for everyone because frankly it’s just massively harder to build a service that works for thousands/millions of use cases than for any one of them individually.

Screen Shot 2016-05-30 at 8.21.36 PM.png

Stateful services

Ohhhh and let’s just briefly talk about state.

The serverless utopia mostly ignores the problems of stateful services.  If pressed they will usually say DynamoDB, or Firebase, or RDS or Aurora or something.

This is a big, huge, deep, wide lake of crap to wade in to so all I’m going to say is that there is no such thing as having the luxury of not having to understand how your storage systems work.  Queries will get slow, and you’ll need to be able to figure out why and fix them.  You’ll hit scaling cliffs where suddenly a perfectly-usable app just starts timing everything out because of that extra second of latency coming from …


The hardware underlying your instance will degrade (there’s a server somewhere under all those abstractions, don’t forget).  The provider will have mysterious failures.  They will be better than you, probably, but less inclined to give you satisfactory progress updates because there are hundreds or thousands or millions of you all clamoring.

The more you understand about your storage system (and the more you stay in the lane of how it was intended to be used), the happier you’ll be.

Screen Shot 2016-05-30 at 8.28.29 PM

In conclusion

These trends are both inevitable and, for the most part, very good news for everyone.

Operations engineering is becoming a more fascinating and specialized skill set.  The best engineers are flocking to solve category problems — instead of building the same system at company after company, they are building SaaS solutions to solve it for the internet at large.  Just look at the massive explosion in operational software offerings over the past 5-6 years.

This means that the era of the in-house dedicated ops team, which serves as an absorbent buffer for all the pain of software development, is mostly on its way out the door.  (And good riddance.)

People are waking up to the fact that software quality improves when feedback loops are tighter for software engineers, which means being on call and owning services end to end.  The center of gravity is shifting towards engineering teams owning the services they built.

This is awesome!  You get to rent engineers from Google, AWS, Pagerduty, Pingdom, Heroku, etc for much cheaper than if you hired them in-house — if you could even get them, which you probably can’t because talent is scarce.

Screen Shot 2016-05-30 at 8.39.42 PM

But the flip side of this is that application engineers need to get better at thinking in traditionally operations-oriented ways about reliability, architecture, instrumentation, visibility, security, and storage.  Figure out what your core differentiators are, and own the shit out of those.

Nobody but you can care about your mission as much as you can.  Own it, do it.  Have fun.


Operational Best Practices #serverless